WO2023138190A1 - 目标检测模型的训练方法及对应的检测方法 - Google Patents

目标检测模型的训练方法及对应的检测方法 Download PDF

Info

Publication number
WO2023138190A1
WO2023138190A1 PCT/CN2022/131716 CN2022131716W WO2023138190A1 WO 2023138190 A1 WO2023138190 A1 WO 2023138190A1 CN 2022131716 W CN2022131716 W CN 2022131716W WO 2023138190 A1 WO2023138190 A1 WO 2023138190A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
area
feature
feature map
detection model
Prior art date
Application number
PCT/CN2022/131716
Other languages
English (en)
French (fr)
Inventor
王娜
刘星龙
黄宁
陈翼男
Original Assignee
上海商汤智能科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海商汤智能科技有限公司 filed Critical 上海商汤智能科技有限公司
Publication of WO2023138190A1 publication Critical patent/WO2023138190A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0012—Biomedical image inspection
    • G06T7/0014—Biomedical image inspection using an image reference approach
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/22—Matching criteria, e.g. proximity measures
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10072—Tomographic images
    • G06T2207/10081—Computed x-ray tomography [CT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10116—X-ray image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10132—Ultrasound image
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30004—Biomedical image processing
    • G06T2207/30061—Lung
    • G06T2207/30064—Lung nodule

Definitions

  • the present application relates to the technical field of image processing, in particular to a method for training a target detection model and a corresponding detection method.
  • the target detection model can be used to detect organs in the future, which can greatly reduce the workload of medical staff and improve the efficiency of diagnosis and treatment.
  • the present application at least provides a method for training a target detection model and a corresponding detection method, device, equipment, storage medium, and computer program product.
  • the first aspect of the present application provides a method for training a target detection model, the training method comprising: acquiring a sample medical image containing a preset organ, wherein the sample medical image is marked with a labeling result of at least one target located on the preset organ, and the labeling result includes an actual area where the target is located; using the target detection model to match at least one first candidate area for at least one (for example, each) target in a matching order, and obtaining a final prediction result about the target based on the first candidate area, wherein the matching order is determined based on the size of the actual area where at least one (for example, each) target is located; using the final prediction result and the labeling result, Tune the parameters of the object detection model.
  • At least one (for example, each) target is matched with at least one first candidate area according to the matching order, and the matching order is determined based on the size of the actual area where the target is located on the preset organ, so that the matching order can be adjusted according to the size of the actual area where the target is located, so that the matching order can be more adapted to the situation of different sizes of the target, which helps to improve the recall rate during the training of the target detection model and improve the effect of target detection.
  • the above matching sequence is: the smaller the size of the actual area, the earlier the target is matched.
  • the target with the smaller size of the actual area can be matched first, so that the target with the smaller size of the actual area can be matched to a more suitable first candidate area, so as to improve the recall rate during the training of the target detection model, especially the recall rate of small targets, which helps to improve the effect of target detection.
  • the aforementioned target detection model is used to match at least one (for example, each) target with at least one first candidate area according to the matching order, including: based on the size of the actual area where the at least one (for example, each) target is located, at least one (for example, each) target is divided into different target groups, wherein the size ranges corresponding to the target groups are different; based on the size ranges of different target groups, the matching order corresponding to different target groups is determined; according to the matching order, at least one (for example, each) target group is matched with at least one first candidate area.
  • the targets can be divided into groups based on the size of the actual area where the target is located. Further, by determining the matching order corresponding to different target groups based on the size range of different target groups, the matching order can be determined based on the size of the actual area where the target is located.
  • the above-mentioned matching of at least one first candidate region for the targets in at least one (for example, each) target group according to the matching order includes: performing the following matching steps on at least one (for example, each) target group according to the matching order: obtaining the matching degree between at least one (for example, each) target in the target group and different anchor regions of the sample medical image; based on the matching degree, selecting at least one anchor region for at least one (for example, each) target of the target group as the first candidate region of the target.
  • At least one anchor region can be selected for at least one (for example, each) target in the target group as the first candidate region of the target, and the corresponding first candidate region can be determined for at least one (for example, each) target in the target group.
  • the number of the first candidate regions of the target in the above-mentioned different target groups is different, and the smaller the size range of the target group, the more the number of the first candidate regions of the target; and/or, the degree of matching between the target and the anchor region is the degree of coincidence between the target and the anchor region.
  • the matching degree is determined according to the degree of coincidence.
  • the number of first candidate areas of targets in the target group with smaller size ranges is greater, it is possible to use more first candidate areas that match targets with small size ranges during training to train the target detection model, which helps to improve the sensitivity of the target detection model to small target detection and the accuracy of small target detection in practical applications.
  • the above-mentioned acquisition of the matching degree between at least one (for example, each) target in the target group and different anchor regions of the sample medical image includes: selecting an anchor region that is not used as the first candidate region as the anchor region to be matched from a plurality of anchor regions generated for the sample medical image, and obtaining the matching degree between at least one (for example, each) target in the target group and at least one anchor region to be matched; It also includes: generating a preset number of anchor point regions of different sizes for at least one (for example, each) position point of the sample medical image, wherein the sizes of the preset number of anchor point regions are respectively determined based on a preset number of first feature maps of different scales of the sample medical image.
  • anchor region that is not the first candidate region as the anchor region to be matched, and obtaining the degree of matching between at least one (for example, each) target in the target group and at least one (for example, each) anchor region to be matched, more anchor regions can be selected as the first candidate region.
  • multiple anchor point regions may be generated.
  • the method before using the above-mentioned target detection model to match at least one (for example, each) target with at least one first candidate region according to the matching order, and obtaining a prediction result about the target based on the first candidate region, the method also includes: using the target detection model to obtain a preset number of first feature maps of different scales of the sample medical image, wherein the preset number is greater than or equal to 1; obtaining a final prediction result about the target based on the first candidate region, including: predicting and obtaining the final prediction result based on the first candidate region and the first feature map.
  • the first feature maps of different sizes can be used to train the target detection model, thereby improving the detection effect of the target detection model on targets of different sizes.
  • the aforementioned acquisition of a preset number of first feature maps of different scales of the sample medical image includes: performing feature extraction on the sample medical image to obtain a preset number of second feature maps of different scales; performing preset attention processing on the second feature maps respectively to obtain the first feature map corresponding to the second feature map, wherein the preset attention processing includes one or more of dimension attention processing and feature channel attention processing.
  • the target detection model by performing preset attention processing on the second feature map, it is helpful for the target detection model to extract more accurate feature information about the target, thereby improving the accuracy and recall rate of target detection.
  • the above-mentioned preset attention processing is performed on the second feature map to obtain the first feature map corresponding to the second feature map, including: obtaining the dimension weight corresponding to at least one (for example, each) dimension of the second feature map, and using the dimension weight corresponding to at least one (for example, each) dimension to perform weight processing on at least one (for example, each) dimension feature in the second feature map to obtain a spatially focused feature map; dividing the features of different channels in the spatially focused feature map into several channel feature groups, and obtaining channel weights corresponding to the channel feature groups respectively, using the channel weights to perform weighting processing on several channel feature groups to obtain a classic The first feature map obtained by feature channel attention processing.
  • the target detection model by using the second feature map to obtain the spatially focused feature map, and further using the spatially focused feature map to obtain the first feature map obtained through feature channel attention processing, it is helpful for the target detection model to extract more accurate feature information about the target in the spatial and channel dimensions.
  • the above-mentioned acquisition of dimension weights corresponding to at least one (for example, each) dimension of the second feature map includes: taking at least one (for example, each) dimension of the second feature map as a target dimension, and performing average pooling on the second feature map for the remaining dimensions except the target dimension to obtain a third feature map on the target dimension; using the third feature map on different target dimensions to determine the dimension weight corresponding to at least one (for example, each) dimension of the second feature map; Chord transformation to obtain channel weights corresponding to at least one (for example, each) channel feature group.
  • the third feature map on the target dimension is obtained, and using the third feature maps on different target dimensions, so as to determine the dimension weight corresponding to at least one (for example, each) dimension of the second feature map, which helps the target detection model to extract more accurate feature information about the target.
  • the above-mentioned at least the first candidate region is selected from a plurality of anchor regions of different sizes in the sample medical image, and the first candidate regions of different sizes correspond to first feature maps of different scales; based on the first candidate region and the first feature map, predicting and obtaining a final prediction result includes: obtaining feature information of the first candidate region based on the first feature map corresponding to the size of the first candidate region; using the feature information of the first candidate region to predict and obtain a final prediction result about the target.
  • the feature information of the first candidate area can be used for target detection to obtain the final prediction result.
  • the above-mentioned use of the feature information of the first candidate area to predict and obtain the final prediction result about the target includes: using the feature information of the first candidate area to adjust the first candidate area to obtain an initial prediction result corresponding to at least one (for example, each) first candidate area, wherein, the initial prediction result corresponding to the first candidate area includes an initial prediction area of the target adjusted based on the first candidate area; using the initial prediction result corresponding to at least one (for example each) first candidate area to perform optimized prediction to obtain a final prediction result about the target.
  • the feature information of the first candidate area to adjust the first candidate area to obtain an initial prediction result corresponding to at least one (for example, each) first candidate area, and further using at least one (for example each) initial prediction result corresponding to the first candidate area to perform optimized prediction, and obtain a final prediction result about the target, so that the sensitivity of the target detection model to the target can be improved, and the accuracy of target detection can be improved.
  • the training of the target detection model can be realized based on the classification loss of the target detection model and the regression loss of the prediction result.
  • the above-mentioned final prediction result is obtained by predicting the target detection model using the first candidate area to obtain the initial prediction result and then optimizing the initial prediction result.
  • the initial prediction result includes the initial prediction area of the target and the initial confidence of the category to which the initial prediction area belongs;
  • the above-mentioned use of the final prediction result and labeling results to adjust the parameters of the target detection model also includes: obtaining the second category loss based on the initial confidence;
  • the loss, adjusting the parameters of the target detection model includes: adjusting the parameters of the target detection model by using the first category loss, the first regression loss, the second category loss and the second regression loss.
  • the parameters of the target detection model can be adjusted, so as to realize the training of the target detection model.
  • the above-mentioned first regression loss is obtained based on the offset between the final prediction area and the actual area, and the intersection ratio between the final prediction area and the actual area, or the second regression loss is obtained based on the offset between the initial prediction area and the actual area, and the intersection ratio between the initial prediction area and the actual area, including: using the offset corresponding to the corresponding prediction area to obtain the first offset loss of the corresponding prediction area;
  • the second offset loss of the corresponding prediction area is obtained, wherein the larger the intersection and union ratio, the smaller the loss weight; based on the intersection and union ratio corresponding to the corresponding prediction area, the GIOU loss of the corresponding prediction area is obtained; the corresponding regression loss is obtained by using the second offset loss and GIOU loss.
  • the loss weight is smaller. In this way, it can be realized that the final predicted area corresponding to the target is less coincident with the actual area where the target is located, and the result of the lower coincidence degree will be larger, so that the parameter update of the target detection model is more vigorous when optimizing the positioning, which helps to improve the accuracy of target detection.
  • the second offset loss and GIOU loss to obtain the corresponding regression loss, the predicted region location of the trained target detection model can be more accurate.
  • the above-mentioned sample medical image is a three-dimensional medical image; and/or, the preset organ is lung, and the target is nodule.
  • the trained target detection model can perform targeted detection on the nodule in the lung.
  • the second aspect of the present application provides a target detection method, the detection method comprising: acquiring a target medical image containing a preset organ; using a target detection model to obtain a first feature map of the target medical image, and determining at least one first candidate region of the target, and obtaining a final prediction result about the target based on the first candidate region and the first feature map; wherein, the target detection model is obtained by using the training method of the target detection model in the first aspect above, and/or the first feature map is obtained by performing preset attention processing on the second feature map obtained by extracting the features of the target medical image, and the preset attention processing includes dimension attention One or more of processing and feature channel attention processing.
  • the accuracy and recall rate of target detection can be improved.
  • the target detection model trained by the above target detection model training method to perform target detection, the accuracy and recall rate of target detection can be improved.
  • the third aspect of the present application provides a training device for a target detection model.
  • the device includes an acquisition module, a detection module, and an adjustment module, wherein the acquisition module is used to acquire a sample medical image containing a preset organ, wherein the sample medical image is marked with a labeling result of at least one target located on the preset organ, and the labeling result includes the actual area where the target is located; the detection module is used to use the target detection model to match at least one first candidate area for at least one (for example, each) target according to the matching order, and obtain a final prediction result about the target based on the first candidate area, wherein the matching order is based on at least one (for example, each) target The size of the actual area is determined; the adjustment module is used to adjust the parameters of the target detection model by using the final prediction result and labeling result.
  • the fourth aspect of the present application provides a target detection device.
  • the target detection device includes: an acquisition module and a detection module, wherein the acquisition module is used to acquire a target medical image containing a preset organ; the detection module is used to use a target detection model to obtain a first feature map of the target medical image, determine at least one first candidate region of the target, and obtain a final prediction result about the target based on the first candidate region and the first feature map; wherein the target detection model is obtained by using the above-mentioned training method of the target detection model in the first aspect, and/or the first feature map is the second feature map obtained by extracting the features of the target medical image.
  • the preset attention processing includes one or more of dimension attention processing and feature channel attention processing.
  • the fifth aspect of the present application provides an electronic device, the electronic device includes a memory and a processor coupled to each other, the processor is used to execute the program instructions stored in the memory, so as to implement the method for training the target detection model described in the first aspect above, or realize the target detection method described in the second aspect above.
  • the sixth aspect of the present application provides a computer-readable storage medium on which program instructions are stored.
  • the program instructions are executed by a processor, the method for training the object detection model described in the first aspect above is implemented, or the method for object detection described in the second aspect above is implemented.
  • the seventh aspect of the present application provides a computer program product, including computer-readable codes, or a computer-readable storage medium carrying computer-readable codes.
  • the processor in the electronic device executes the method for training the target detection model described in the first aspect above, or realizes the target detection method described in the second aspect above.
  • At least one (for example, each) target is matched with at least one first candidate area according to the matching order, and the matching order is determined based on the size of the actual area where the target is located on the preset organ, so that the matching order can be adjusted according to the size of the actual area where the target is located, so that the matching order can be more adapted to the situation of different sizes of targets, which helps to improve the recall rate during target detection model training and improve the effect of target detection.
  • Fig. 1 is the first schematic flow chart of an embodiment of the training method of the target detection model of the present application
  • Fig. 2 is the second schematic flow chart of an embodiment of the training method of the target detection model of the present application
  • Fig. 3 is a first flowchart of another embodiment of the training method of the target detection model of the present application.
  • Fig. 4 is a second schematic flowchart of another embodiment of the training method of the target detection model of the present application.
  • Fig. 5 is a third schematic flowchart of another embodiment of the training method of the target detection model of the present application.
  • Fig. 6 is a schematic flowchart of another embodiment of the training method of the target detection model of the present application.
  • FIG. 7 is a schematic structural diagram of a target detection model in the training method of the target detection model in the present application.
  • Fig. 8 is a schematic flow chart of an embodiment of the target detection method of the present application.
  • FIG. 9 is a schematic structural diagram of a training device for a target detection model of the present application.
  • Fig. 10 is a schematic structural diagram of the target detection device of the present application.
  • Fig. 11 is a schematic frame diagram of an embodiment of the electronic device of the present application.
  • Fig. 12 is a schematic diagram of an embodiment of a computer-readable storage medium of the present application.
  • FIG. 1 is a schematic flow chart of an embodiment of a method for training a target detection model in the present application. Specifically, the following steps may be included:
  • Step S11 Obtain a sample medical image including preset organs.
  • the predetermined organ may be an organ of an animal or a human body.
  • Animal organs are, for example, dog kidneys, hearts, and the like.
  • Organs of the human body are, for example, kidneys, lungs, heart, and the like.
  • the predetermined organ is the lungs of a human body.
  • the sample medical image may be a two-dimensional image or a three-dimensional image.
  • the three-dimensional image may be a three-dimensional image obtained by scanning an organ.
  • three-dimensional imaging may be performed by using computerized tomography (Computed Tomography, CT) imaging technology to obtain a sample medical image.
  • CT computerized tomography
  • the two-dimensional image is, for example, a sample medical image obtained by ultrasonic imaging technology or X-ray imaging technology. It can be understood that the imaging method of the sample medical image is not limited.
  • the sample medical image is marked with a mark result of at least one target located on a preset organ, and the mark result includes the actual area where the target is located.
  • the target on the predetermined organ may be a specific substance present on the organ. For example, nodules in the lungs, cysts in the kidneys, etc.
  • the actual region where the target is located is the region where the target exists on the sample medical image. For example, the area where the nodules of the lungs are located, the area where the cysts of the kidneys are located, and so on.
  • the predetermined organ is the lung
  • the target on the predetermined organ is a nodule.
  • the sample medical image may be obtained by resampling the initial sample medical image.
  • the resolution of the sample medical image can meet the requirements, which helps to improve the accuracy of target detection.
  • the normalization operation can also be performed on the pixel values in the sample medical images, which is convenient for the training of the subsequent target detection model.
  • operations such as rotation, translation, mirroring, and scaling can be performed on the sample medical image to achieve data enhancement, and can balance the positive and negative samples in the sample medical image, amplify the amount of data, and help improve the generalization of the target detection model and reduce the possibility of overfitting.
  • Step S12 using the target detection model to match at least one first candidate region for at least one (eg, each) target according to the matching order, and obtain a final prediction result about the target based on the first candidate region.
  • the first candidate area is, for example, an anchor area (anchor) matched with a target in a one-stage detection algorithm or a two-stage detection algorithm.
  • the matching order is determined based on the size of the actual area where at least one (for example, each) target is located.
  • Determine the matching order based on the size of the actual area where at least one (for example, each) target is located for example, determine the priority matching target according to the size of the actual area where the target is located, or determine the number of first candidate areas that match targets of different sizes according to the size of the actual area where the target is located.
  • the matching sequence is: the smaller the size of the actual area, the earlier the target is matched. Therefore, by setting the target with the smaller size of the actual area to be matched earlier, the target with the smaller size of the actual area can be matched first, so that the target with the smaller size of the actual area can be matched to a more suitable first candidate area, so as to improve the recall rate during the training of the target detection model, especially the recall rate of small targets, which helps to improve the effect of target detection.
  • a final prediction result about the target can be obtained based on the first candidate region.
  • the final prediction result includes a final prediction area of the predicted target.
  • the specific process of obtaining the final prediction result of the target may be a specific process of a one-stage detection algorithm commonly used in the art, or a specific process of a two-stage detection algorithm, which will not be repeated here.
  • Step S13 Using the final prediction results and labeling results, adjust the parameters of the target detection model.
  • the loss function can be used to determine the corresponding loss value according to the difference between the final prediction result and the labeling result, and the parameters of the target detection model can be adjusted according to the loss value.
  • each target is matched with at least one first candidate area according to the matching order, and the matching order is determined based on the size of the actual area where the target is located on the preset organ, so that the matching order can be adjusted according to the size of the actual area where the target is located, so that the matching order can be more adapted to the situation of different sizes of the target, which helps to improve the recall rate during the training of the target detection model and improve the effect of target detection.
  • FIG. 2 is a second schematic flowchart of an embodiment of a method for training a target detection model in the present application.
  • the "using the target detection model to match at least one first candidate region for at least one (for example, each) target according to the matching order" mentioned in the above steps specifically includes steps S121 to S123.
  • Step S121 Divide at least one (for example, each) target into different target groups based on the size of the actual area where the at least one (for example, each) target is located.
  • the target groups correspond to different size ranges. That is to say, the size ranges to which the sizes of the actual regions where the objects belonging to different object groups are located are different. In this way, the objects can be divided into groups based on the size of the actual area where the objects are located.
  • each nodule may be divided into different target groups according to the size of the actual area of the nodule.
  • the nodules can be divided into different groups according to different sizes, for example, nodules smaller than 6 mm are divided into small nodules, nodules from 6 mm to 12 mm are medium nodules, and nodules larger than 12 mm are large nodules.
  • nodules smaller than 6 mm are divided into small nodules
  • nodules from 6 mm to 12 mm are medium nodules
  • nodules larger than 12 mm are large nodules.
  • the small nodules were divided into one group
  • the middle nodules were divided into one group
  • the large nodules were divided into one group.
  • the targets may be divided into different target groups directly according to the size of the actual area where the target is located.
  • the actual area where the target is located may be an area
  • the sample medical image is a three-dimensional image
  • the actual area where the target is located may be a volume. Then, those in the actual area where the target is located are divided into one group, and those in the actual area where the target is located are in the second preset range, and so on.
  • Step S122 Based on the size ranges of the different target groups, determine the matching order corresponding to the different target groups.
  • the targets are divided into groups based on the size of the actual area where the target is located.
  • the matching order can be determined based on the size of the actual area where at least one (for example, each) target is located.
  • the size range of the first target group is less than 6 mm.
  • the size range of the second target group is 6 mm to 10 mm
  • the size range of the third target group is greater than 10 mm to 15 mm
  • the size range of the fourth target group is greater than 15 mm.
  • the matching order can be determined as follows: the first matches the first target group, the second matches the second target group, the third matches the third target group, and the fourth matches the fourth target group.
  • Step S123 Match at least one first candidate region for the objects in at least one (for example, each) object group according to the matching sequence.
  • At least one first candidate region can be matched to the targets in at least one (eg, each) target group according to the matching order.
  • the at least one first candidate area may be sequentially matched according to the size order of the actual area where the targets in the target group are located. For example, matching can be performed from small to large, or from small to large.
  • the targets can be divided into groups based on the size of the actual area where the target is located. Further, by determining the matching order corresponding to different target groups based on the size range of different target groups, the matching order can be determined based on the size of the actual area where the target is located.
  • the following matching steps S1231 and S1232 may be performed on at least one (for example, each) target group according to the matching order.
  • Step S1231 Obtain the degree of matching between at least one (for example, each) target in the target group and different anchor regions of the sample medical image.
  • the anchor region may be a default region generated in the sample medical image.
  • four anchor regions can be generated on each pixel of the sample medical image, which are anchor regions with sizes of 4*4, 8*8, 16*16 and 32*32 respectively.
  • four anchor point regions can be generated on each voxel of the sample medical image, which are respectively 4*4*4, 8*8*8, 16*16*16 and 32*32*32 anchor point regions. It can be understood that the size of the anchor area can be set as required, and there is no limitation here.
  • the anchor regions can be used to obtain the matching degree of each anchor region and the target, and then obtain the matching degree between at least one (for example, each) target in the target group and different anchor regions of the sample medical image.
  • the matching degree between the target and the anchor region is the coincidence degree between the target and the anchor region.
  • the Intersection of Union (IoU) between the actual area where the target is located and the anchor area can be used as the coincidence degree between the target and the anchor area.
  • the ratio of the overlap between the actual area where the target is located and the anchor area to the anchor area may be directly taken as the coincidence degree. Therefore, by taking the coincidence degree between the target and the anchor region as the matching degree between the target and the anchor region, the matching degree is determined according to the degree of coincidence.
  • Step S1232 Based on the matching degree, select at least one anchor region as the first candidate region of the target for at least one (for example, each) target of the target group.
  • At least one anchor region can be selected for each object in the object group as the first candidate region of the object. For example, for each target in the target group, several anchor regions with the highest matching degree may be selected as the first candidate regions. For example, 1000 anchor regions are generated on a sample medical image, and by calculating the degree of matching between the 1000 anchor regions and a certain target, the degree of matching between the 1000 anchor regions and a certain target can be determined, so that the 6 anchor regions with the highest matching degree can be selected as the first candidate regions for a certain target.
  • an anchor region not used as the first candidate region may be selected from a plurality of anchor regions generated for the sample medical image as an anchor region to be matched, and the degree of matching between at least one (for example, each) target in the target group and at least one (for example, each) anchor region to be matched is obtained. For example, among the 1000 anchor regions generated on the sample medical image, there are already 50 anchor regions as the first candidate regions for some objects, then when selecting the first candidate regions for other objects, several anchor regions can be selected from the remaining 950 anchor regions as the first candidate regions for other objects.
  • anchor region that is not the first candidate region as the anchor region to be matched, and obtaining the degree of matching between at least one (for example, each) target in the target group and at least one (for example, each) anchor region to be matched, more anchor regions can be selected as the first candidate region.
  • At least one anchor region can be selected for at least one (for example, each) target in the target group as the first candidate region of the target, and the first candidate region can be determined for at least one (for example, each) target in the target group.
  • the number of first candidate regions of objects in different object groups is different, and the number of first candidate regions of objects in an object group with a smaller size range is greater.
  • there are 4 target groups in total and the size range of the first target group is less than 6mm.
  • the size range of the second target group is 6 mm to 10 mm
  • the size range of the third target group is greater than 10 mm to 15 mm
  • the size range of the fourth target group is greater than 15 mm.
  • the number of the first candidate areas of the targets of the first target group is the largest, which is 6; the number of the first candidate areas of the targets of the second target group is 4; the number of the first candidate areas of the targets of the third target group is 3; Therefore, by determining the number of first candidate areas of targets in the target group with a smaller size range is greater, so that more first candidate areas that match targets with a smaller size range can be used during training to train the target detection model, which helps to improve the sensitivity of the target detection model to small target detection and the accuracy of small target detection in practical applications.
  • a part of the anchor region in addition to matching at least one first candidate region for each target, a part of the anchor region can also be selected as the second candidate region, and the second part of the candidate region can be used as a region where no target exists to train the target detection model.
  • an anchor region whose matching degree with a certain target is within a certain interval may be selected as the second candidate region.
  • the matching degree between the target and the anchor area is the intersection ratio between the actual area where the target is located and the anchor area
  • the anchor area with an intersection ratio of 0.02-0.2 can be used as the second candidate area to train the target detection model, so as to balance the number of positive and negative samples and improve the training effect of the target detection model.
  • the step of "matching at least one first candidate region for at least one (for example, each) target group according to the matching order" it may also be performed first: generating a preset number of anchor point regions of different sizes for at least one (for example, each) position point of the sample medical image, wherein the sizes of the preset number of anchor point regions are respectively determined based on a preset number of first feature maps of different scales in the sample medical image.
  • the preset number of first feature maps of different scales based on the sample medical image may be obtained by using the feature extraction network of the target detection model to extract the feature of the sample medical image.
  • the feature extraction network in the feature map pyramid network (Feature Pyramid Networks, FPN) and single-step multi-frame target detection SSD (Single Shot MultiBox Detector) model is used to obtain a preset number of first feature maps of different scales. It can be understood that the method for obtaining a preset number of first feature maps of different scales is not limited.
  • an anchor region of one size may be generated on the sample medical image corresponding to a first feature map, so that a preset number of anchor regions of different sizes may be generated on the sample medical image.
  • the specific size of the anchor region with a larger size is determined based on the first feature map with a smaller size.
  • a 4*4*4 anchor region can be generated on the sample medical image based on the first feature map of 48*48*48 size
  • an 8*8*8 anchor point region can be generated on the sample medical image based on the first feature map of 24*24*24
  • an 8*8*8 anchor point region can be generated on the sample medical image based on the first feature map of 12*12*12 size.
  • a 16*16*16 anchor region is generated on the image
  • a 32*32*32 anchor region is generated on the sample medical image based on the first feature map of 6*6*6 size.
  • the feature information on the first feature map corresponding to the first candidate area may be used as the feature information corresponding to the first candidate area. Since the sizes of the preset number of anchor regions are respectively determined based on the preset number of different-scale first feature maps of the sample medical image, the anchor region determined based on a certain first feature map is the first feature map corresponding to the anchor region, so that the first feature map corresponding to the size of the first candidate region can be correspondingly determined. For example.
  • the size of the anchor point area is 16*16*16, then the size of the first feature map corresponding to the size of the anchor point area is 12*12*12 in size, and the feature information corresponding to the anchor point area is the feature information of the area on the first feature map of 12*12*12 size corresponding to the anchor point area.
  • the anchor region generated in the sample medical image may be directly used as the first candidate region for target detection.
  • FIG. 3 is a schematic flow chart of another embodiment of a method for training a target detection model of the present application. In this embodiment, it specifically includes step S21 to step S24.
  • Step S21 Obtain a sample medical image including a preset organ.
  • step S11 For the specific description of this step, please refer to the above step S11, which will not be repeated here.
  • Step S22 Obtain a preset number of first feature maps of different scales of the sample medical image by using the target detection model.
  • the preset number is greater than or equal to 1.
  • Using the target detection model to obtain a preset number of first feature maps of different scales of the sample medical image may be obtained by using the feature extraction network of the target detection model to extract features from the sample medical image.
  • the feature extraction network in the feature map pyramid network (Feature Pyramid Networks, FPN) and single-step multi-frame target detection SSD (Single Shot MultiBox Detector) model is used to obtain a preset number of first feature maps of different scales.
  • the method for obtaining a preset number of first feature maps of different scales is not limited.
  • the bottom-up part of the feature map pyramid network is a residual network (Residual Network, ResNet).
  • the residual network is, for example, ResNet18.
  • Step S23 using the target detection model to match at least one first candidate region for at least one (for example, each) target according to the matching sequence, and obtain a final prediction result about the target based on the first candidate region.
  • step S12 For the specific description of this step, please refer to the above step S12, which will not be repeated here.
  • step S23 specifically includes: based on the first candidate region and the first feature map, predicting to obtain a final prediction result. For each first candidate region, its corresponding region on the first feature map can be determined, and then the feature information of the first candidate region on the first feature map can be determined, and then the final prediction result can be obtained based on the first candidate region and the feature information of the first candidate region on the first feature map.
  • the first candidate region may be selected from several anchor point regions of different sizes in the sample medical image.
  • the first candidate regions of different sizes correspond to the first feature maps of different scales respectively.
  • the first feature maps of different sizes can be used to train the target detection model, thereby improving the detection effect of the target detection model on targets of different sizes.
  • the above step of "predicting and obtaining the final prediction result based on the first candidate region and the first feature map” specifically includes step S231 and step S232 (not shown in the figure).
  • Step S231 Obtain feature information of the first candidate region based on the first feature maps respectively corresponding to the sizes of the first candidate region.
  • Step S232 Using feature information of the first candidate area, predict to obtain a final prediction result about the target.
  • a target detection algorithm may be used to detect the target, so as to predict and obtain a final prediction result about the target.
  • the target detection algorithm is, for example, a one-stage detection algorithm or a two-stage detection algorithm, which will not be repeated here.
  • the feature information of the first candidate area can be used for target detection to obtain the final prediction result.
  • the step "Using the feature information of the first candidate area to predict and obtain the final prediction result about the target” specifically includes step S2321 and step S2322 (not shown in the figure).
  • Step S2321 Using the feature information of the first candidate area to adjust the first candidate area to obtain an initial prediction result corresponding to at least one (for example, each) first candidate area.
  • the initial prediction result corresponding to the first candidate area includes the initial prediction area of the target adjusted based on the first candidate area.
  • the initial prediction area is, for example, the Proposal obtained by the two-stage detection algorithm. That is, the target detection model can perform regression (adjustment) on the first candidate region based on the feature information of the first candidate region, so as to obtain an initial prediction result corresponding to the first candidate region.
  • the target detection model can also perform regression on the second candidate area based on the feature information of the second candidate area, so as to obtain an initial prediction result corresponding to the second candidate area.
  • Step S2322 Use the initial prediction results corresponding to at least one (for example, each) first candidate area to perform optimal prediction, and obtain a final prediction result about the target.
  • the initial prediction region of the initial prediction result can be used as a region of interest (RoI) to perform another prediction.
  • the specific process can refer to the relevant process in the two-stage detection algorithm, and will not be repeated here.
  • the feature information of the first candidate area to adjust the first candidate area to obtain an initial prediction result corresponding to at least one (for example, each) first candidate area, and further using at least one (for example each) initial prediction result corresponding to the first candidate area to perform optimized prediction, and obtain a final prediction result about the target, so that the sensitivity of the target detection model to the target can be improved, and the accuracy of target detection can be improved.
  • Step S24 Using the final prediction result and labeling result, adjust the parameters of the target detection model.
  • step S13 For the specific description of this step, please refer to the above step S13, which will not be repeated here.
  • FIG. 4 is a second schematic flowchart of another embodiment of a method for training a target detection model of the present application.
  • the above-mentioned step "obtaining a preset number of first feature maps of different scales of the sample medical image” specifically includes step S221 and step S222.
  • Step S221 Perform feature extraction on the sample medical image to obtain a preset number of second feature maps of different scales.
  • Feature extraction is performed on sample medical images to obtain a preset number of second feature maps of different scales. Specifically, it can be obtained by using Feature Pyramid Networks (Feature Pyramid Networks, FPN) and the feature extraction network in the single-step multi-frame target detection SSD (Single Shot MultiBox Detector) model, which will not be described here.
  • FPN Feature Pyramid Networks
  • SSD Single Shot MultiBox Detector
  • Step S222 Perform preset attention processing on the second feature map respectively to obtain the first feature map corresponding to the second feature map.
  • the preset attention processing includes one or more of dimension attention processing and feature channel attention processing.
  • Dimensional attention processing is, for example, the coordinate attention processing in target detection algorithms.
  • the feature channel attention processing is, for example, the channel attention (Channel Attention) processing in the target detection algorithm. I won't repeat them here.
  • the target detection model by performing preset attention processing on the second feature map, it is helpful for the target detection model to extract more accurate feature information about the target, thereby improving the accuracy and recall rate of target detection.
  • step S222 specifically includes step S2221 and step S2222 (not shown in the figure).
  • Step S2221 Acquire dimension weights corresponding to at least one (for example, each) dimension of the second feature map, and use the dimension weights corresponding to at least one (for example, each) dimension to perform weighting processing on at least one (for example, each) dimensional feature in the second feature map to obtain a spatially focused feature map.
  • the sample medical image is a two-dimensional image
  • it may be the dimension weights corresponding to the X and Y dimensions of the second feature map, and then the dimension weights corresponding to the X and Y dimensions perform weighting processing on the X and Y dimension features in the second feature map to obtain a spatially focused feature map.
  • the sample medical image is a three-dimensional image
  • it may be the dimension weights corresponding to the X, Y, and Z dimensions of the second feature map, and then the dimension weights corresponding to the X, Y, and Z dimensions perform weighting processing on the X, Y, and Z dimensions of the second feature map to obtain a spatially focused feature map.
  • Obtaining the dimension weights corresponding to each dimension may be obtained after processing the feature information of each dimension of the second feature map by using a coordinate attention mechanism.
  • the dimension weight corresponding to at least one (for example, each) dimension of the second feature map may be obtained through the following steps 1 and 2 (not shown in the figure).
  • Step 1 Take at least one (for example, each) dimension of the second feature map as the target dimension, perform average pooling on the remaining dimensions except the target dimension on the second feature map, and obtain the third feature map on the target dimension.
  • the second feature map is also a three-dimensional feature map.
  • the X, Y, and Z dimensions of the second feature map may be used as target dimensions respectively. Taking the X dimension of the second feature map as the target dimension, and then performing average pooling on the Y and Z dimensions to obtain the third feature map on the X dimension.
  • the third feature map in the Y and Z dimensions can also be obtained in the same way.
  • Step 2 Using the third feature maps on different target dimensions, determine the dimension weights corresponding to at least one (for example, each) dimension of the second feature map.
  • the third feature map on at least one (for example, each) dimension After obtaining the third feature map on at least one (for example, each) dimension, use the third feature map on at least one (for example, each) dimension to determine the dimension weight corresponding to at least one (for example, each) dimension of the second feature map.
  • the third feature map in each dimension can be concatenated, and then processed by a convolutional layer with batch normalization and nonlinear activation, and then the output of the convolutional layer is re-divided into feature maps in each dimension, and then after a layer of convolution and activation, the dimension weights corresponding to each dimension of the second feature map are obtained.
  • the third feature map in the X, Y, and Z dimensions can be spliced, and then processed by a convolutional layer with batch normalization and nonlinear activation, and then the output of the convolution layer is re-divided into feature maps in the X, Y, and Z dimensions, and then after a layer of convolution and activation, the dimension weights corresponding to the X, Y, and Z dimensions of the second feature map are obtained.
  • the second feature map can be first divided into several parts in the channel dimension, and then step 1 and step 2 are respectively performed on each part to obtain the dimension weight corresponding to each dimension of the second feature map of each part, and then the results of each part are combined to obtain the complete dimension weight corresponding to each dimension of the second feature map.
  • the third feature map on the target dimension is obtained, and using the third feature maps on different target dimensions, so as to determine the dimension weight corresponding to at least one (for example, each) dimension of the second feature map, which helps the target detection model to extract more accurate feature information about the target.
  • Step S2222 Divide the features of different channels in the spatially focused feature map into several channel feature groups, and respectively obtain the channel weights corresponding to the channel feature groups, and use the channel weights to weight the several channel feature groups to obtain the first feature map obtained through feature channel attention processing.
  • the features of different channels in the spatial focusing feature map are divided into several channel feature groups, for example, the 256-dimensional channel features in the spatial focusing feature map are divided into four channel feature groups, and each channel feature group is a 64-dimensional channel feature.
  • obtaining the channel weights corresponding to the channel feature groups may specifically be to perform cosine transformation on at least one (for example, each) channel feature group, that is, perform frequency-domain channel attention (Frequency Channel Attention) processing on at least one (for example, each) channel feature group, so as to obtain the channel weight corresponding to at least one (for example, each) channel feature group.
  • the features of the spatially focused feature map can be divided into several equal parts in the channel dimension, and each feature can be multiplied by the cosine series to obtain the cosine-transformed frequency. These frequencies are then merged in the channel dimension, and pass through a fully connected layer with activation function sigmoid activation to obtain channel weights. Therefore, by performing cosine transformation on at least one (for example, each) channel feature group, the utilization rate of feature information can be improved, which helps to improve the accuracy of target detection.
  • the target detection model by using the second feature map to obtain the spatially focused feature map, and further using the spatially focused feature map to obtain the first feature map obtained through feature channel attention processing, it is helpful for the target detection model to extract more accurate feature information about the target in the spatial and channel dimensions.
  • FIG. 5 is a schematic flowchart of a third embodiment of another embodiment of a method for training a target detection model in the present application.
  • the above-mentioned final prediction result further includes the final confidence level of the category to which the final prediction region belongs.
  • the above-mentioned step of "using the final prediction result and labeling result to adjust the parameters of the target detection model" specifically includes steps S241 to S243.
  • Step S241 Obtain the first category loss based on the final confidence.
  • the optimal final prediction result can be selected from several final prediction results, so as to realize the detection of the target. For example, non-maximum suppression (Non-Maximum Suppression, NMS) processing can be performed on the final confidence of each final prediction result, so as to obtain the optimal final prediction result, and at the same time, the classification score of the optimal final prediction result can also be determined accordingly.
  • NMS Non-Maximum Suppression
  • the first category loss may be calculated based on the classification scores of the optimal final prediction results corresponding to the several targets and the classification scores of the final prediction results corresponding to the second candidate region.
  • the Focal loss function can be used to obtain the first category loss.
  • the first category loss can be calculated using the following formula (1):
  • y represents the actual classification information of the optimal final prediction result obtained
  • y′ represents the classification score of the optimal final prediction result
  • ⁇ is the loss weight
  • ⁇ is the adjustment weight
  • the label [0, 1] of the real classification information may be softened to [0.1, 0.9], so as to enhance the generalization performance of the object detection model.
  • Step S242 Obtain the first regression loss based on the offset between the final predicted region and the actual region, and the intersection ratio between the final predicted region and the actual region.
  • the final prediction area of this step may be the final prediction area of the optimal final prediction result of an object determined in step S231.
  • Determining the offset between the final prediction area and the actual area may be a general method in the art, for example, using a smooth-L1 loss function to determine the offset between the final prediction area and the actual area.
  • the method of determining the intersection-over-union ratio of the final predicted area and the actual area can be a general calculation method, and will not be repeated here.
  • the offset between the final predicted area and the actual area can be weighted by using the intersection ratio between the final predicted area and the actual area as an adjustment weight, or the loss value can be calculated by using the intersection ratio between the final predicted area and the actual area, and then weighted and summed with the offset between the final predicted area and the actual area to obtain the first regression loss.
  • Step S243 Using the first category loss and the first regression loss, adjust the parameters of the target detection model.
  • the final loss can be determined based on these two losses, for example, by means of weighted summation to obtain the final loss. And adjust the parameters of the object detection model based on the final loss.
  • the training of the target detection model can be realized based on the classification loss of the target detection model and the regression loss of the prediction result.
  • the final prediction result is obtained by predicting the target detection model using the first candidate region to obtain an initial prediction result and optimizing the initial prediction result.
  • the initial prediction result includes the initial prediction area of the target and the initial confidence of the category to which the initial prediction area belongs.
  • the initial prediction result further includes an initial prediction result corresponding to the second candidate area.
  • the above step of "Using the final prediction result and labeling result to adjust the parameters of the target detection model" also includes Step S234 and Step S235 (not shown in the figure).
  • Step S244 Obtain the second category loss based on the initial confidence.
  • Step S245 Obtain the second regression loss based on the offset between the initial prediction area and the actual area, and the intersection ratio between the initial prediction area and the actual area.
  • step S244 and step S245 please refer to the above step S241 and step S242, which will not be repeated here.
  • the above-mentioned step of "using the first category loss and the first regression loss to adjust the parameters of the target detection model” specifically includes: using the first category loss, the first regression loss, the second category loss and the second regression loss to adjust the parameters of the target detection model.
  • the first loss can be obtained by using the first category loss and the first regression loss
  • the second loss can be obtained by using the second category loss and the second regression loss
  • the final loss value can be obtained based on the first loss and the second loss
  • the parameters of the target detection model can be adjusted according to the final loss value.
  • the parameters of the target detection model can be adjusted, so as to realize the training of the target detection model.
  • FIG. 6 is a schematic flowchart of another embodiment of a method for training a target detection model of the present application.
  • the above-mentioned first regression loss is obtained based on the offset between the final prediction area and the actual area, and the intersection ratio between the final prediction area and the actual area
  • the second regression loss is obtained, or the second regression loss is obtained based on the offset between the initial prediction area and the actual area, and the intersection ratio between the initial prediction area and the actual area, including steps S31 to S33.
  • Step S31 Obtain the first offset loss of the corresponding prediction area by using the offset corresponding to the corresponding prediction area; obtain the loss weight of the corresponding prediction area based on the intersection and union ratio corresponding to the corresponding prediction area, and multiply the first offset loss of the corresponding prediction area by the loss weight of the corresponding prediction area to obtain the second offset loss of the corresponding prediction area.
  • the greater the intersection-over-union ratio the smaller the loss weight. In this way, it can be realized that the final predicted area corresponding to the target is less coincident with the actual area where the target is located, and the result of the lower coincidence degree will be larger, so that the parameter update of the target detection model is more vigorous when optimizing the positioning, which helps to improve the accuracy of target detection.
  • the corresponding prediction area may be the initial prediction area corresponding to the object in the initial prediction result, or the final prediction area corresponding to the object in the final prediction result.
  • the intersection ratio corresponding to the corresponding prediction area may be the intersection ratio of the initial prediction area and the actual area, and the intersection ratio of the final prediction area and the actual area.
  • the second offset loss can be determined by the following formula (2) and formula (3):
  • iou is the intersection and union ratio corresponding to the corresponding prediction area.
  • W iou is the loss weight of the corresponding prediction area, and formula (3) is the smooth-L1 loss function after weighting based on the loss weight W iou .
  • Step S32 Obtain the GIOU loss of the corresponding prediction area based on the intersection-over-union ratio corresponding to the corresponding prediction area.
  • the GIOU (Generalized Intersection over Union) loss can be obtained based on the following formula (4):
  • A is the corresponding prediction area
  • B is the actual area
  • C is the minimum closed area between the corresponding prediction area and the actual area.
  • Step S33 Using the second offset loss and the GIOU loss to obtain the corresponding regression loss.
  • the weighted addition of the second offset loss and the GIOU loss can be used to obtain the corresponding regression loss.
  • the first regression loss or the second regression loss can be obtained respectively.
  • the predicted region location of the trained target detection model can be more accurate.
  • FIG. 7 is a schematic structural diagram of the target detection model in the training method of the target detection model in the present application.
  • the target detection model 10 includes a feature extraction module 11 , an attention module 12 and a detection module 13 .
  • the feature extraction module 11 is, for example, a feature map pyramid network, a feature extraction network in an SSD model, and the like.
  • the feature extraction module 11 can perform feature extraction on the input sample medical images to obtain a preset number of second feature maps of different scales.
  • the sample medical image is marked with a marking result of at least one target located on a preset organ, and the marking result includes an actual area where the target is located.
  • the attention module 12 may perform preset attention processing on at least one (for example, each) second feature map.
  • the preset attention processing includes one or more of dimension attention processing and feature channel attention processing, so that the first feature map can be obtained.
  • the attention module 12 includes a spatial attention submodule 121 and a feature attention submodule 122.
  • the spatial attention submodule 121 can perform dimension attention processing on the second feature map to obtain a spatially focused feature map
  • the feature attention submodule 122 can perform feature channel attention on the spatially focused feature map to obtain a first feature map.
  • the detection module 13 can respectively match at least one first candidate area for at least one (for example, each) target according to the matching order, and obtain a final prediction result about the target based on the first candidate area.
  • the detection module 13 includes an initial prediction submodule 131 and a final prediction submodule 132 .
  • the initial prediction sub-module 131 can use the feature information of the first candidate area to adjust the first candidate area to obtain an initial prediction result corresponding to at least one (for example, each) first candidate area.
  • the final prediction sub-module 132 can use the initial prediction results corresponding to at least one (for example, each) first candidate area to perform optimized prediction to obtain a final prediction result about the target.
  • the final prediction sub-module 132 optimizes the prediction of the initial prediction result, which may be to use the region of interest pooling (ROI pooling) layer of the final prediction sub-module 132 to perform region of interest pooling (ROI pooling), then use two layers of fully connected layers with non-linear activation to process, and then detect, so as to obtain the final prediction result about the target.
  • ROI pooling region of interest pooling
  • ROI pooling region of interest pooling
  • FIG. 8 is a schematic flowchart of an embodiment of a target detection method of the present application.
  • the target detection method specifically includes:
  • Step S41 Obtain a target medical image including a preset organ.
  • step S11 Regarding the acquisition of the target medical image including the preset organ, please refer to the above step S11, which will not be repeated here.
  • Step S42 Using the target detection model to obtain a first feature map of the target medical image, and determine at least one first candidate region of the target, and obtain a final prediction result about the target based on the first candidate region and the first feature map.
  • the target detection model is trained by using the above-mentioned target detection model training method.
  • the first candidate region when the target detection model is used for target detection, the first candidate region may be an anchor region directly generated on the target medical image.
  • the final prediction result about the target is obtained based on the first candidate region and the first feature map, specifically, detection may be performed according to the feature information of the first candidate region on the first feature map, so as to obtain the final prediction result about the target.
  • detection may be performed according to the feature information of the first candidate region on the first feature map, so as to obtain the final prediction result about the target.
  • Step S43 Using the initial prediction results corresponding to at least one (for example, each) first candidate area to perform optimal prediction, and obtain a final prediction result about the target.
  • step S2322 For a detailed description of this step, please refer to the above step S2322 and other related descriptions, and details will not be repeated here.
  • the accuracy and recall rate of target detection can be improved.
  • the first feature map is obtained by performing preset attention processing on the second feature map obtained by feature extraction of the target medical image
  • the preset attention processing includes one or more of dimension attention processing and feature channel attention processing.
  • the target detection model it is helpful for the target detection model to extract more accurate feature information about the target, thereby improving the accuracy and recall rate of target detection.
  • FIG. 9 is a schematic structural diagram of a training device for a target detection model of the present application.
  • the training device 90 includes an acquisition module 91 , a detection module 92 and an adjustment module 96 .
  • the acquisition module 91 is used to acquire a sample medical image containing a preset organ, wherein the sample medical image is marked with an annotation result of at least one target located on the preset organ, and the annotation result includes the actual area where the target is located;
  • the detection module 92 is used to use the target detection model to match at least one first candidate area for at least one (for example, each) target according to the matching order, and obtain a final prediction result about the target based on the first candidate area, wherein the matching order is determined based on the size of the actual area where at least one (for example, each) target is located;
  • the parameters of the model is used to acquire a sample medical image containing a preset organ, wherein the sample medical image is marked with an annotation result of at least one target located on the preset organ, and the annotation
  • the above matching sequence is: the smaller the size of the actual area, the earlier the target is matched.
  • the above-mentioned detection module 92 is used to use the target detection model to respectively match at least one first candidate area for at least one (for example, each) target according to the matching order, including: based on the size of the actual area where the at least one (for example, each) target is located, at least one (for example, each) target is divided into different target groups, wherein the size ranges corresponding to the target groups are different;
  • the above-mentioned detection module 92 is used to match at least one first candidate region for the targets in at least one (for example, each) target group according to the matching order, including: performing the following matching steps on at least one (for example, each) target group according to the matching order: obtaining the matching degree between at least one (for example, each) target in the target group and different anchor regions of the sample medical image; based on the matching degree, selecting at least one anchor region for at least one (for example, each) target in the target group as the first candidate region of the target.
  • the number of the first candidate regions of the target in the above-mentioned different target groups is different, and the smaller the size range of the target group, the more the number of the first candidate regions of the target; and/or, the degree of matching between the target and the anchor region is the degree of coincidence between the target and the anchor region.
  • the above-mentioned detection module 92 is used to obtain the matching degree between at least one (for example, each) target in the target group and different anchor point regions of the sample medical image, including: from a plurality of anchor point regions generated for the sample medical image, select the anchor point region that is not used as the first candidate region as the anchor point region to be matched, and obtain the matching degree between at least one (for example, each) target in the target group and at least one (for example, each) anchor point region to be matched; Before the targets in each) target group match at least one first candidate region, the detection module 92 is also used to generate a preset number of anchor point regions of different sizes for at least one (for example, each) position point of the sample medical image, wherein the sizes of the preset number of anchor point regions are determined based on a preset number of first feature maps of different scales in the sample medical image.
  • the above-mentioned detection module 92 is used to use the target detection model to match at least one first candidate area for at least one (for example, each) target according to the matching order, and obtain the prediction result about the target based on the first candidate area.
  • the above detection module 92 is used to obtain a preset number of first feature maps of different scales of the sample medical image, including: performing feature extraction on the sample medical image to obtain a preset number of second feature maps of different scales; performing preset attention processing on the second feature maps respectively to obtain the first feature map corresponding to the second feature map, wherein the preset attention processing includes one or more of dimension attention processing and feature channel attention processing.
  • the above-mentioned detection module 92 is used to perform preset attention processing on the second feature map to obtain the first feature map corresponding to the second feature map, including: obtaining the dimension weight corresponding to at least one (for example, each) dimension of the second feature map, and using the dimension weight corresponding to at least one (for example, each) dimension to perform weighting processing on at least one (for example, each) dimension feature in the second feature map to obtain a spatially focused feature map; dividing the features of different channels in the spatially focused feature map into several channel feature groups, and obtaining channel weights corresponding to the channel feature groups, and using the channel weights to add the several channel feature groups Weight processing to obtain the first feature map obtained by feature channel attention processing.
  • the above-mentioned detection module 92 is used to obtain the dimension weights corresponding to at least one (for example, each) dimension of the second feature map, including: taking at least one (for example, each) dimension of the second feature map as the target dimension, performing average pooling on the second feature map for the remaining dimensions except the target dimension, to obtain the third feature map on the target dimension; using the third feature maps on different target dimensions to determine the dimension weight corresponding to at least one (for example, each) dimension of the second feature map; Cosine transform is performed on the channel feature group to obtain channel weights corresponding to at least one (for example, each) channel feature group.
  • the above-mentioned at least the first candidate region is selected from several anchor regions of different sizes in the sample medical image, and the first candidate regions of different sizes respectively correspond to first feature maps of different scales; the above-mentioned detection module 92 is used to predict and obtain the final prediction result based on the first candidate region and the first feature map, including: obtaining feature information of the first candidate region based on the first feature map corresponding to the size of the first candidate region; using the feature information of the first candidate region to predict and obtain the final prediction result about the target.
  • the above-mentioned detection module 92 is configured to use the characteristic information of the first candidate region to predict and obtain the final prediction result about the target, including: using the characteristic information of the first candidate region to adjust the first candidate region to obtain an initial prediction result corresponding to at least one (for example, each) first candidate region, wherein, the initial prediction result corresponding to the first candidate region includes an initial prediction region of the target adjusted based on the first candidate region; performing optimized prediction by using the initial prediction result corresponding to at least one (for example each) first candidate region to obtain a final prediction result for the target.
  • the above-mentioned final prediction result also includes the final confidence degree of the category to which the final prediction region belongs; the above-mentioned adjustment module 96 is used to adjust the parameters of the target detection model by using the final prediction result and the labeling result, including: obtaining the first category loss based on the final confidence degree; obtaining the first regression loss based on the offset between the final prediction region and the actual region, and the cross-merge ratio between the final prediction region and the actual region; and adjusting the parameters of the target detection model by using the first category loss and the first regression loss.
  • the above-mentioned final prediction result is obtained by predicting the target detection model using the first candidate area to obtain the initial prediction result and optimizing the prediction of the initial prediction result.
  • the initial prediction result includes the initial prediction area of the target and the initial confidence of the category to which the initial prediction area belongs;
  • the above-mentioned adjustment module 96 is used to adjust the parameters of the target detection model by using the final prediction result and labeling results, and also includes: obtaining the second category loss based on the initial confidence;
  • the adjustment module 96 is used to adjust the parameters of the target detection model by using the first category loss and the first regression loss, including: adjusting the parameters of the target detection model by using the first category loss, the first regression loss, the second category loss and the second regression loss.
  • the above-mentioned adjustment module 96 is used to obtain the first regression loss based on the offset between the final prediction area and the actual area and the intersection and combination ratio between the final prediction area and the actual area, or obtain the second regression loss based on the offset between the initial prediction area and the actual area and the intersection and integration ratio between the initial prediction area and the actual area, including: using the offset corresponding to the corresponding prediction area to obtain the first offset loss of the corresponding prediction area; The losses are multiplied to obtain the second offset loss of the corresponding prediction area.
  • intersection and union ratio the smaller the loss weight; based on the intersection and union ratio corresponding to the corresponding prediction area, the GIOU loss of the corresponding prediction area is obtained; the second offset loss and GIOU loss are used to obtain the corresponding regression loss.
  • the above-mentioned sample medical image is a three-dimensional medical image; and/or, the preset organ is lung, and the target is nodule.
  • FIG. 10 is a schematic structural diagram of an object detection device of the present application.
  • the object detection device 100 includes an acquisition module 101 and a detection module 102 .
  • the acquisition module 101 is used to acquire a target medical image containing a preset organ;
  • the detection module 102 is used to use a target detection model to obtain a first feature map of the target medical image, and determine at least one first candidate region of the target, and obtain a final prediction result about the target based on the first candidate region and the first feature map;
  • the target detection model 100 is obtained by using the above-mentioned target detection model training method, and/or, the first feature map is obtained by performing preset attention processing on the second feature map obtained by extracting features from the target medical image, and the preset attention processing includes Dimensional attention processing and feature channel attention One or more of processing.
  • the subject of execution of the method steps of the present application may be executed by hardware, or by means of a processor running computer executable codes.
  • FIG. 11 is a schematic frame diagram of an embodiment of an electronic device of the present application.
  • the electronic device 110 includes a memory 111 and a processor 112 coupled to each other, and the processor 112 is configured to execute the program instructions stored in the memory 111, so as to realize the steps in any of the above embodiments of the object detection model training method, or realize the steps in any of the above embodiments of the object detection method.
  • the electronic device 110 may include, but is not limited to: a microcomputer and a server.
  • the electronic device 110 may also include mobile devices such as notebook computers and tablet computers, which are not limited here.
  • the processor 112 is configured to control itself and the memory 111 to implement the steps in any of the above embodiments of the method for training an image segmentation model, or to implement the steps in any of the above embodiments of the image segmentation method.
  • the processor 112 may also be called a CPU (Central Processing Unit, central processing unit).
  • the processor 112 may be an integrated circuit chip with signal processing capability.
  • the processor 112 may also be a general-purpose processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field-programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.
  • a general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.
  • the processor 112 may be jointly realized by an integrated circuit chip.
  • FIG. 12 is a schematic frame diagram of an embodiment of a computer-readable storage medium of the present application.
  • the computer-readable storage medium 120 stores program instructions 121 that can be executed by the processor, and the program instructions 121 are used to implement the steps in any of the above embodiments of the method for training an object detection model, or to implement the steps in any of the above embodiments of the object detection method.
  • the computer-readable storage medium may be a volatile storage medium or a non-volatile storage medium.
  • the functions or modules included in the device provided by the embodiments of the present disclosure can be used to execute the methods described in the above method embodiments, and its specific implementation can refer to the description of the above method embodiments, and for the sake of brevity, details are not repeated here.
  • the disclosed methods and devices may be implemented in other ways.
  • the device implementations described above are only illustrative.
  • the division of modules or units is only a logical function division. In actual implementation, there may be other division methods.
  • units or components may be combined or integrated into another system, or some features may be ignored or not implemented.
  • the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.
  • each functional unit in each embodiment of the present application may be integrated into one processing unit, each unit may exist separately physically, or two or more units may be integrated into one unit.
  • the above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.
  • the integrated unit is realized in the form of a software function unit and sold or used as an independent product, it can be stored in a computer-readable storage medium.
  • the technical solution of the present application essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product
  • the computer software product is stored in a storage medium, and includes several instructions to make a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) execute all or part of the steps of the methods of each embodiment of the application.
  • the aforementioned storage medium includes: various media that can store program codes such as U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk.
  • Embodiments of the present disclosure further provide a computer program product, the computer program product carries program code, including computer readable code, or a computer readable storage medium carrying computer readable code, when the computer readable code is run in a processor of an electronic device, the instructions included in the program code can be used to execute the steps of the method described in the above method embodiment, for details, refer to the above method embodiment, and details will not be repeated here.
  • program code including computer readable code
  • the above-mentioned computer program product may be specifically implemented by means of hardware, software or a combination thereof.
  • the computer program product is embodied as a computer storage medium, and in another optional embodiment, the computer program product is embodied as a software product, such as a software development kit (Software Development Kit, SDK) and the like.
  • the product applying the technical solution of this application has clearly notified the personal information processing rules and obtained the individual's independent consent before processing personal information.
  • the technical solution of this application involves sensitive personal information
  • the products applying the technical solution of this application have obtained individual consent before processing sensitive personal information, and at the same time meet the requirements of "express consent". For example, at a personal information collection device such as a camera, set up a clear and prominent sign to inform that it has entered the scope of personal information collection, and the personal information will be collected.
  • the individual voluntarily enters the collection scope, it is deemed to agree to the collection of his personal information; or on the personal information processing device, when the personal information processing rules are notified with obvious signs/information, personal authorization is obtained through pop-up messages or by asking individuals to upload their personal information.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Radiology & Medical Imaging (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

本申请公开了一种目标检测模型的训练方法及对应的检测方法、装置、设备、存储介质以及计算机程序产品，其中，训练方法包括：获取包含预设器官的样本医学图像，其中，样本医学图像标注有位于预设器官上的至少一个目标的标注结果，标注结果包括目标所在的实际区域；利用目标检测模型按照匹配顺序分别为至少一个目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的最终预测结果，其中，匹配顺序是基于至少一个目标所在的实际区域的尺寸确定的；利用最终预测结果与标注结果，调整目标检测模型的参数。通过该方法，能够提高目标检测模型训练时的召回率。

Description

目标检测模型的训练方法及对应的检测方法
本申请要求在2022年01月24日提交中国专利局，申请号为202210080240.1，申请名称为“目标检测模型的训练方法及对应的检测方法”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及图像处理技术领域，特别是涉及一种目标检测模型的训练方法及对应的检测方法。
背景技术
通过对器官进行检测，可以发现器官可能存在的病灶，有助于提高诊疗效率。目前，通过对目标检测模型进行训练，后续便可以利用目标检测模型来对器官进行检测，以此能够大幅降低医务人员的工作量，提高诊疗效率。
然而，由于器官中的目标大小不一，现有的目标检测模型的训练方法在对不同大小的目标进行检测时，召回率不高，导致目标检测的效果不好，这限制了该技术的进一步发展。
因此，如何改进目标检测模型的训练方法，以提高检测结果的召回率，以提高目标检测的效果，有非常重要的意义。
发明内容
本申请至少提供一种目标检测模型的训练方法及对应的检测方法、装置、设备、存储介质以及计算机程序产品。
本申请第一方面提供了一种目标检测模型的训练方法，该训练方法包括：获取包含预设器官的样本医学图像，其中，样本医学图像标注有位于预设器官上的至少一个目标的标注结果，标注结果包括目标所在的实际区域；利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的最终预测结果，其中，匹配顺序是基于至少一个(例如各个)目标所在的实际区域的尺寸确定的；利用最终预测结果与标注结果，调整目标检测模型的参数。
因此，在训练的过程中，通过按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并且，通过基于预设器官上的目标所在的实际区域的尺寸来确定匹配顺序，以此使得匹配顺序能够针对目标所在的实际区域的尺寸大小进行调整，使得匹配顺序能够更加适应目标大小不一的情况，有助于提高目标检测模型训练时的召回率，提高目标检测的效果。
其中，上述的匹配顺序为：实际区域的尺寸越小，越早对目标进行匹配。
因此，通过设置实际区域的尺寸越小的目标越早进行匹配，可以是实际区域的尺寸较小的目标能够优先被匹配，使得实际区域的尺寸较小的目标能够被匹配到更加合适的第一候选区域，以此能够提高目标检测模型训练时的召回率，尤其是小目标的召回率，有助于提高目标检测的效果。
其中，上述的利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，包括：基于至少一个(例如各个)目标所在的实际区域的尺寸，将至少一个(例如各个)目标划分至不同目标组，其中，目标组分别对应的尺寸范围不同；基于不同目标组的尺寸范围，确定不同目标组对应的匹配顺序；按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域。
因此，通过基于不同目标组的尺寸范围，确定不同目标组对应的匹配顺序，可以实现基于目标所在的实际区域的尺寸大小，分组对目标进行划分，进一步地，通过基于不同目标组的尺寸范围，确定不同目标组对应的匹配顺序，实现了基于目标所在的实际区域的尺寸确定匹配顺序。
其中，上述的按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域，包括：按照匹配顺序对至少一个(例如各个)目标组进行如下匹配步骤：获取目标组中的至少一个(例如各个)目标分别与样本医学图像的不同锚点区域之间的匹配程度；基于匹配程度，为目标组的至少一个(例如各个)目标选出至少一个锚点区域作为目标的第一候选区域。
因此，通过获取目标组中的至少一个(例如各个)目标分别与样本医学图像的不同锚点区域之间的匹配程度，以此便能够基于匹配程度，为目标组的至少一个(例如各个)目标选出至少一个锚点区域作为目标的第一候选区域，实现了对目标组的至少一个(例如每个)目标确定对应的第一候选区域。
其中，上述的不同目标组中的目标的第一候选区域的数量不同，且尺寸范围越小的目标组中的目标的第一候选区域的数量越多；和/或，目标与锚点区域之间的匹配程度为目标与锚点区域之间的重合度。
因此，通过将目标与锚点区域之间的重合度作为目标与锚点区域之间的匹配程度，以此根据重合度的大小确定匹配程度。另外，通过确定尺寸范围越小的目标组中的目标的第一候选区域的数量越多，使得在训练时能够利用更多与尺寸范围小的目标匹配的第一候选区域来对目标检测模型进行训练，以此有助于提高目标检测模型对小目标检测的敏感性，以及实际应用时对小目标进行检测的准确度。
其中，上述的获取目标组中的至少一个(例如各个)目标分别与样本医学图像的不同锚点区域之间的匹配程度，包括：从为样本医学图像生成的多个锚点区域中，选择未作为第一候选区域的锚点区域作为待匹配的锚点区域，并获取目标组中的至少一个(例如各个)目标分别与至少一个待匹配的锚点区域之间的匹配程度；和/或，在按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域之前，方法还包括：为样本医学图像的至少一个(例如各个)位置点生成不同尺寸的预设数量个锚点区域，其中，预设数量个锚点区域的尺寸是分别基于样本医学图像的预设数量个不同尺度第一特征图确定的。
因此，通过选择未作为第一候选区域的锚点区域作为待匹配的锚点区域，并获取目标组中的至少一个(例如各个)目标分别与至少一个(例如各个)待匹配的锚点区域之间的匹配程度，可以选择更多的锚点区域作为第一候选区域。另外，通过基于样本医学图像的预设数量个不同尺度的第一特征图来确定设数量个锚点区域的尺寸，可以生成多个锚点区域。
其中，上述的利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的预测结果之前，方法还包括：利用目标检测模型获取样本医学图像的预设数量个不同尺度的第一特征图，其中，预设数量大于或等于1；基于第一候选区域得到关于目标的最终预测结果，包括：基于第一候选区域和第一特征图，预测得到最终预测结果。
因此，通过生成不同尺度的预设数量个第一特征图，可以利用不同尺寸的第一特征图来对目标检测模型进行训练，以此能够提高目标检测模型对不同大小的目标的检测效果。
其中，上述的获取样本医学图像的预设数量个不同尺度的第一特征图，包括：对样本医学图像进行特征提取，得到不同尺度的预设数量个第二特征图；分别对第二特征图进行预设注意力处理，得到第二特征图对应的第一特征图，其中，预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。
因此，通过对第二特征图进行预设注意力处理，有助于目标检测模型提取到更准确的关于目标的特征信息，以此能够目标检测的准确度和召回率。
其中，上述的对第二特征图进行预设注意力处理，得到第二特征图对应的第一特征图，包括：获取第二特征图的至少一个(例如各个)维度对应的维度权重，利用至少一个(例如各个)维度对应的维度权重对第二特征图中的至少一个(例如各个)维度特征进行加权处理，得到空间聚焦特征图；将空间聚焦特征图中的不同通道的特征分为若干份通道特征组，并分别获取通道特征组对应的通道权重，利用通道权重对若干份通道特征组进行加权处理，得到经特征通道注意力处理得到的第一特征图。
因此，通过利用第二特征图得到空间聚焦特征图，并进一步地利用空间聚焦特征图得到经特征通道注意力处理得到的第一特征图，有助于目标检测模型在空间维度和通道维度上提取到更准确的关于目标的特征信息。
其中，上述的获取第二特征图的至少一个(例如各个)维度对应的维度权重，包括：分别将第二特征图的至少一个(例如各个)维度作为目标维度，对第二特征图进行除目标维度以外的剩余维度的平均池化，得到目标维度上的第三特征图；利用不同目标维度上的第三特征图，确定第二特征图的至少一个(例如各个)维度对应的维度权重；和/或，分别获取通道特征组对应的通道权重，包括：对至少一份(例如每份)通道特征组进行余弦变换，以得到与至少一份(例如每份)通道特征组对应的通道权重。
因此，通过分别将第二特征图的至少一个(例如各个)维度作为目标维度，对第二特征图进行除目标维度以外的剩余维度的平均池化，得到目标维度上的第三特征图，并利用不同目标维度上的第三特征图，以此可以确定第二特征图的至少一个(例如各个)维度对应的维度权重，有助于目标检测模型提取到更准确的关于目标的特征信息。
其中，上述的至少第一候选区域是从样本医学图像的不同尺寸的若干锚点区域中选择得到的，不同尺寸的第一候选区域分别对应不同尺度的第一特征图；基于第一候选区域和第一特征图，预测得到最终预测结果，包括：基于与第一候选区域的尺寸分别对应的第一特征图，获得第一候选区域的特征信息；利用第一候选区域的特征信息，预测得到关于目标的最终预测结果。
因此，通过确定第一候选区域的尺寸对应的第一特征图，并获得第一候选区域的特征信息，以此便能够利用第一候选区域的特征信息进行目标检测，以得到最终预测结果。
其中，上述的利用第一候选区域的特征信息，预测得到关于目标的最终预测结果，包括：利用第一候选区域的特征信息对第一候选区域进行调整，得到至少一个(例如各个)第一候选区域对应的初始预测结果，其中，第一候选区域对应的初始预测结果包括基于第一候选区域调整得到的目标的初始预测区域；利用至少一个(例如各个)第一候选区域对应的初始预测结果进行优化预测，得到关于目标的最终预测结果。
因此，通过利用第一候选区域的特征信息对第一候选区域进行调整，以得到至少一个(例如各个)第一候选区域对应的初始预测结果，并进一步利用至少一个(例如各个)第一候选区域对应的初始预测结果进行优化预测，得到关于目标的最终预测结果，以此可以目标检测模型对目标进行检测的敏感性，有助于提高目标检测的准确度。
其中，上述的最终预测结果还包括最终预测区域所属的类别的最终置信度；利用最终预测结果与标注结果，调整目标检测模型的参数，包括：基于最终置信度，得到第一类别损失；基于最终预测区域与实际区域之间的偏移量、以及最终预测区域与实际区域的交并比，得到第一回归损失；利用第一类别损失和第一回归损失，调整目标检测模型的参数。
因此，通过获得第一类别损失和第一回归损失，可以基于目标检测模型的分类损失和预测结果的回归损失，实现对目标检测模型的训练。
其中，上述的最终预测结果是对目标检测模型利用第一候选区域进行预测得到初始预测结果并对初始预测结果进行优化预测得到的，初始预测结果包括目标的初始预测区域和初始预测区域所属的类别的初始置信度；上述的利用最终预测结果与标注结果，调整目标检测模型的参数，还包括：基于初始置信度，得到第二类别损失；基于初始预测区域与实际区域之间的偏移量、以及初始预测区域与实际区域的交并比，得到第二回归损失；利用第一类别损失和第一回归损失，调整目标检测模型的参数，包括：利用第一类别损失、第一回归损失、第二类别损失和第二回归损失，调整目标检测模型的参数。
因此，通过获得第一类别损失、第一回归损失、第二类别损失和第二回归损失，可以对目标检测模型的参数的调整，以此实现对目标检测模型的训练。
其中，上述的基于最终预测区域与实际区域之间的偏移量、以及最终预测区域与实际区域的交并比，得到第一回归损失，或基于初始预测区域与实际区域之间的偏移量、以及初始预测区域与实际区域的交并比，得到第二回归损失，包括：利用相应预测区域对应的偏移量，得到相应预测区域的第一偏移量损失；基于相应预测区域对应的交并比，得到相应预测区域的损失权重，利用相应预测区域的损失权重对相应预测区域的第一偏移量损失进行相乘，得到相应预测区域的第二偏移量损失，其中，交并比越大，损失权重越小；基于相应预测区域对应的交并比，得到相应预测区域的GIOU损失；利用第二偏移量损失和GIOU损失，得到相应回归损失。
因此，通过设定交并比越大，损失权重越小。以此可以实现对目标对应的最终预测区域与目标所在的实际区域重合度较低的结果以较大的惩罚，使得目标检测模型在优化定位时的参数更新力度更大，有助于提高目标检测的准确度。另外，通过利用第二偏移量损失和GIOU损失，得到相应回归损失，可以训练后的目标检测模型的预测区域定位更加准确。
其中，上述的样本医学图像为三维医学图像；和/或，预设器官为肺部，目标为结节。
因此，通过限定预设器官为肺部，目标为结节，使得训练后的目标检测模型能够针对肺部的结节进行针对性的检测。
本申请第二方面提供了一种目标检测方法，该检测方法包括：获取包含预设器官的目标医学图像；利用目标检测模型获得目标医学图像的第一特征图，并确定目标的至少一个第一候选区域，并基于第一候选区域和所第一特征图得到关于目标的最终预测结果；其中，目标检测模 型是利用上述的第一方面目标检测模型的训练方法训练得到的，和/或，第一特征图是将对目标医学图像特征提取得到的第二特征图进行预设注意力处理得到的，预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。
因此，通过利用经过上述的目标检测模型的训练方法训练得到的目标检测模型来进行目标检测，能够提升目标检测的准确度和召回率。另外，通过对第二特征图进行预设注意力处理，有助于目标检测模型提取到更准确的关于目标的特征信息，以此能够目标检测的准确度和召回率。
本申请第三方面提供了一种目标检测模型的训练装置，该装置包括获取模块、检测模块和调整模块，其中，获取模块，用于获取包含预设器官的样本医学图像，其中，样本医学图像标注有位于预设器官上的至少一个目标的标注结果，标注结果包括目标所在的实际区域；检测模块，用于利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的最终预测结果，其中，匹配顺序是基于至少一个(例如各个)目标所在的实际区域的尺寸确定的；调整模块，用于利用最终预测结果与标注结果，调整目标检测模型的参数。
本申请第四方面提供了一种目标检测装置，目标检测装置包括：获取模块和检测模块，其中，获取模块用于获取包含预设器官的目标医学图像；检测模块用于利用目标检测模型获得目标医学图像的第一特征图，并确定目标的至少一个第一候选区域，并基于第一候选区域和所第一特征图得到关于目标的最终预测结果；其中，目标检测模型是利用上述的第一方面目标检测模型的训练方法训练得到的，和/或，第一特征图是将对目标医学图像特征提取得到的第二特征图进行预设注意力处理得到的，预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。
本申请第五方面提供了一种电子设备，该电子设备包括相互耦接的存储器和处理器，处理器用于执行存储器中存储的程序指令，以实现上述第一方面描述的目标检测模型的训练方法，或实现上述第二方面描述的目标检测方法。
本申请第六方面提供了一种计算机可读存储介质，其上存储有程序指令，程序指令被处理器执行时实现上述第一方面描述的目标检测模型的训练方法，或实现上述第二方面描述的目标检测方法。
本申请第七方面提供了一种计算机程序产品，包括计算机可读代码，或者承载有计算机可读代码的计算机可读存储介质，当所述计算机可读代码在电子设备的处理器中运行时，所述电子设备中的处理器执行用于实现上述第一方面描述的目标检测模型的训练方法，或实现上述第二方面描述的目标检测方法。
上述方案，在训练的过程中，通过按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并且，通过基于预设器官上的目标所在的实际区域的尺寸来确定匹配顺序，以此使得匹配顺序能够针对目标所在的实际区域的尺寸大小进行调整，使得匹配顺序能够更加适应目标大小不一的情况，有助于提高目标检测模型训练时的召回率，提高目标检测的效果。
应当理解的是，以上的一般描述和后文的细节描述仅是示例性和解释性的，而非限制本申请。
附图说明
此处的附图被并入说明书中并构成本说明书的一部分，这些附图示出了符合本申请的实施例，并与说明书一起用于说明本申请的技术方案。
图1是本申请目标检测模型的训练方法一实施例的第一流程示意图；
图2是本申请目标检测模型的训练方法一实施例的第二流程示意图；
图3是本申请目标检测模型的训练方法另一实施例的第一流程示意图；
图4是本申请目标检测模型的训练方法另一实施例的第二流程示意图；
图5是本申请目标检测模型的训练方法另一实施例的第三流程示意图；
图6是本申请目标检测模型的训练方法又一实施例的流程示意图；
图7是本申请目标检测模型的训练方法中目标检测模型的结构示意图。
图8是本申请目标检测方法实施例的流程示意图；
图9是本申请目标检测模型的训练装置的一结构示意图；
图10是本申请目标检测装置的一结构示意图；
图11是本申请电子设备一实施例的框架示意图；
图12是本申请计算机可读存储介质一实施例的框架示意图。
具体实施方式
下面结合说明书附图，对本申请实施例的方案进行详细说明。
以下描述中，为了说明而不是为了限定，提出了诸如特定系统结构、接口、技术之类的具体细节，以便透彻理解本申请。
本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中字符“/”，一般表示前后关联对象是一种“或”的关系。此外，本文中的“多”表示两个或者多于两个。另外，本文中术语“至少一种”表示多种中的任意一种或多种中的至少两种的任意组合，例如，包括A、B、C中的至少一种，可以表示包括从A、B和C构成的集合中选择的任意一个或多个元素。
请参阅图1，图1是本申请目标检测模型的训练方法一实施例的第一流程示意图。具体而言，可以包括如下步骤：
步骤S11：获取包含预设器官的样本医学图像。
在本申请中，预设器官可以动物、人体的器官。动物的器官例如是狗的肾脏、心脏等等器官。人体的器官例如是肾脏、肺部、心脏等等。在一个具体实施方式中，预设器官是人体的肺部。
样本医学图像可以是二维图像或者是三维图像。三维图像具体可以是对器官扫描得到的三维图像。例如，可以通过电子计算机断层扫描(Computed Tomography，CT)成像技术，进行三维成像，以此得到样本医学图像。二维图像例如是通过超声成像技术、X光成像技术得到的样本医学图像。可以理解的，样本医学图像的成像方法不受限制。
在本申请中，样本医学图像标注有位于预设器官上的至少一个目标的标注结果，标注结果包括目标所在的实际区域。预设器官上的目标可以是器官上存在的特定的物质。例如是肺部的结节，肾脏的囊肿等等。目标所在的实际区域即是该目标在样本医学图像上存在的区域。例如是肺部的结节所在的区域，肾脏的囊肿所在的区域等等。
在一个实施方式中，预设器官为肺部，预设器官上的目标为结节。通过限定预设器官为肺部，目标为结节，使得训练后的目标检测模型能够针对肺部的结节进行针对性的检测。
在一个实施方式中，样本医学图像可以是对初始样本医学图像进行重采样得到的。通过对初始样本医学图像进行重采样，可以使得样本医学图像的分辨率符合要求，有助于提高目标检测的准确度。进一步地，还可以对样本医学图像中的像素值进行归一化操作，方便于后续目标检测模型的训练。
在一个实施方式中，在得到样本医学图像以后，还可以对样本医学图像进行旋转、平移、镜像、缩放等操作，以此实现数据增强，并且能够平衡样本医学图像中的正负样本，扩增数据量，有助于提高目标检测模型的泛化性、降低过拟合的可能。
步骤S12：利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的最终预测结果。
对于目标检测而言，在基于候选区域的目标检测算法中，不管是一阶段检测还是二阶段检测，均需要在训练时，为目标匹配至少一个第一候选区域，以将第一候选区域作为存在目标的样本区域，并利用第一候选区域来得到目标的最终预测结果。第一候选区域例如是一阶段检测或者是二阶段检测算法中的与目标匹配的锚点区域(anchor)。
在本申请中，匹配顺序是基于至少一个(例如各个)目标所在的实际区域的尺寸确定的。基于至少一个(例如各个)目标所在的实际区域的尺寸确定匹配顺序，例如是根据目标所在的实际区域的尺寸大小确定优先匹配的目标，或者是根据目标所在的实际区域的尺寸大小确定与不同大小的目标匹配的第一候选区域的数量。
在一个实施方式中，匹配顺序为：实际区域的尺寸越小，越早对目标进行匹配。因此，通过设置实际区域的尺寸越小的目标越早进行匹配，可以是实际区域的尺寸较小的目标能够优先被匹配，使得实际区域的尺寸较小的目标能够被匹配到更加合适的第一候选区域，以此能够提高目标检测模型训练时的召回率，尤其是小目标的召回率，有助于提高目标检测的效果。
在得到为至少一个(例如各个)目标匹配至少一个第一候选区域后，便可以基于第一候选 区域得到关于目标的最终预测结果。在一个实施方式中，最终预测结果包括预测得到目标的最终预测区域。
具体得到目标的最终预测结果的过程可以是本领域通用的一阶段检测算法的具体过程，或者是二阶段检测算法的具体过程，此处不再赘述。
步骤S13：利用最终预测结果与标注结果，调整目标检测模型的参数。
得到最终预测结果后，便可以根据最终预测结果与标注结果的差异，利用损失函数确定对应的损失值，并根据损失值来调整目标检测模型的参数。
因此，在训练的过程中，通过按照匹配顺序分别为各目标匹配至少一个第一候选区域，并且，通过基于预设器官上的目标所在的实际区域的尺寸来确定匹配顺序，以此使得匹配顺序能够针对目标所在的实际区域的尺寸大小进行调整，使得匹配顺序能够更加适应目标大小不一的情况，有助于提高目标检测模型训练时的召回率，提高目标检测的效果。
请参阅图2，图2是本申请目标检测模型的训练方法一实施例的第二流程示意图。在本实施例中，上述步骤提及的“利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域”具体包括步骤S121至步骤S123。
步骤S121：基于至少一个(例如各个)目标所在的实际区域的尺寸，将至少一个(例如各个)目标划分至不同目标组。
在本实施例中，目标组分别对应的尺寸范围不同。也即，属于不同的目标组的目标所在的实际区域的尺寸所属的尺寸范围各不相同。以此，能够实现基于目标所在的实际区域的尺寸大小，分组对目标进行划分。
例如，对于肺部的结节，可以按照结节的实际区域的尺寸，将每一个结节划分至不同的目标组。具体的，首先可以将将结节按照不同的大小，划分为不同的组，如将小于6毫米的结节分为小结节，6毫米至12毫米的结节为中结节，大于12毫米的结节为大结节。然后，基于大中小结节的实际区域的尺寸，分成3组，小结节分为一组，中结节分为一组，大结节分为一组。
又如，可以直接根据目标所在的实际区域的尺寸大小，将目标划分至不同目标组。在具体分组时，若样本医学图像是二维图像，则目标所在的实际区域可以是面积，若样本医学图像是三维图像，则目标所在的实际区域可以是体积。然后，将目标所在的实际区域的处于第一预设范围的划分为一组，将目标所在的实际区域的处于第二预设范围的划分为一组，以此类推。
步骤S122：基于不同目标组的尺寸范围，确定不同目标组对应的匹配顺序。
通过将至少一个(例如各个)目标划分至不同目标组，实现了基于目标所在的实际区域的尺寸大小，分组对目标进行划分。在此基础上，通过基于不同目标组的尺寸范围，来确定不同目标组对应的匹配顺序，以此可以实现基于至少一个(例如各个)目标所在的实际区域的尺寸确定匹配顺序。
例如，一共有4个目标组，第一目标组的尺寸范围是小于6毫米。第二目标组的尺寸范围是6毫米至10毫米，第三目标组的尺寸范围是大于10毫米至15毫米，第四目标组的尺寸范围是大于15毫米。以此，可以根据目标组的尺寸范围，确定匹配顺序为：第一匹配第一目标组、第二匹配第人目标组、第三匹配第三目标组、第四匹配第四目标组。
步骤S123：按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域。
确定匹配顺序后，即可按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域。具体的，在为每一目标组的目标匹配至少一个第一候选区域时，可以根据按照该目标组内的目标所在的实际区域的尺寸大小顺序，顺序匹配至少一个第一候选区域。例如，可以从小到大进行匹配，或是从小到大进行匹配均可。
因此，通过基于不同目标组的尺寸范围，确定不同目标组对应的匹配顺序，可以实现基于目标所在的实际区域的尺寸大小，分组对目标进行划分，进一步地，通过基于不同目标组的尺寸范围，确定不同目标组对应的匹配顺序，实现了基于目标所在的实际区域的尺寸确定匹配顺序。
在一个实施方式中，在执行上述步骤“按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域”时，对于至少一个(例如每一个)目标组而言，具体可以是按照匹配顺序对至少一个(例如各个)目标组进行如下匹配步骤S1231和步骤S1232(图未示)。
步骤S1231：获取目标组中的至少一个(例如各个)目标分别与样本医学图像的不同锚点 区域之间的匹配程度。
锚点区域可以是在样本医学图像中生成的默认区域。例如，样本医学图像是二维图像时，则可以在样本医学图像的每个像素点都生成4个锚点区域，分别是4*4、8*8、16*16和32*32大小的锚点区域。又如，样本医学图像是三维图像时，则可以在样本医学图像的每个体素上都生成4个锚点区域，分别是4*4*4、8*8*8、16*16*16和32*32*32大小的锚点区域。可以理解的，锚点区域的大小可以根据需要进行设置，此处不做限制。
得到锚点区域以后，可以利用锚点区域来获得每一个锚点区域与目标的匹配程度，进而得到目标组中的至少一个(例如各个)目标分别与样本医学图像的不同锚点区域之间的匹配程度。
在一个具体实施方式中，目标与锚点区域之间的匹配程度为目标与锚点区域之间的重合度。例如，可以将目标所在的实际区域与锚点区域之间交并比(Intersection of Union，IoU)作为目标与锚点区域之间的重合度。又如，可以直接将目标所在的实际区域与锚点区域之间重合大小占锚点区域的比重作为重合度。因此，通过将目标与锚点区域之间的重合度作为目标与锚点区域之间的匹配程度，以此根据重合度的大小确定匹配程度。
步骤S1232：基于匹配程度，为目标组的至少一个(例如各个)目标选出至少一个锚点区域作为目标的第一候选区域。
一般而言，匹配程度越高，表明该锚点区域更加能够反映目标所在的实际区域的真实情况。因此，可以基于匹配程度，为目标组的各目标选出至少一个锚点区域作为目标的第一候选区域。例如，可以为目标组的每一个目标，选择与其匹配程度最高的若干个锚点区域作为第一候选区域。例如，在样本医学图像上生成了1000个的锚点区域，通过计算1000个的锚点区域与某一目标的匹配程度，可以确定1000个的锚点区域与某一目标的匹配程度的高低，以此便可以选择匹配程度最高的6个锚点区域作为与某一目标的第一候选区域。
在一个具体实施方式中，可以从为样本医学图像生成的多个锚点区域中，选择未作为第一候选区域的锚点区域作为待匹配的锚点区域，并获取目标组中的至少一个(例如各个)目标分别与至少一个(例如各个)待匹配的锚点区域之间的匹配程度。例如，在样本医学图像上生成了1000个的锚点区域中，已经有50个锚点区域作为与某些目标的第一候选区域，则后续在为其他的目标选择第一候选区域时，可以从剩下的950个锚点区域中，选择若干锚点区域作为其他目标的第一候选区域。因此，通过选择未作为第一候选区域的锚点区域作为待匹配的锚点区域，并获取目标组中的至少一个(例如各个)目标分别与至少一个(例如各个)待匹配的锚点区域之间的匹配程度，可以选择更多的锚点区域作为第一候选区域。
因此，通过获取目标组中的至少一个(例如各个)目标分别与样本医学图像的不同锚点区域之间的匹配程度，以此便能够基于匹配程度，为目标组的至少一个(例如各个)目标选出至少一个锚点区域作为目标的第一候选区域，实现了对目标组的至少一个(例如每一个)目标确定第一候选区域。
在一个具体实施方式中，不同目标组中的目标的第一候选区域的数量不同，且尺寸范围越小的目标组中的目标的第一候选区域的数量越多。例如，一共有4个目标组，第一目标组的尺寸范围是小于6毫米。第二目标组的尺寸范围是6毫米至10毫米，第三目标组的尺寸范围是大于10毫米至15毫米，第四目标组的的尺寸范围是大于15毫米。则可以确定，第一目标组的的目标的第一候选区域的数量最多，是6个；第二目标组的的目标的第一候选区域的数量是4个；第三目标组的的目标的第一候选区域的数量是3个；第四目标组的的目标的第一候选区域的数量是2个。因此，通过确定尺寸范围越小的目标组中的目标的第一候选区域的数量越多，使得在训练时能够利用更多与尺寸范围小的目标匹配的第一候选区域来对目标检测模型进行训练，以此有助于提高目标检测模型对小目标检测的敏感性，以及实际应用时对小目标进行检测的准确度。
在一个具体实施方式中，除了为每个目标匹配至少一个第一候选区域以外，还可以选择部分锚点区域作为第二候选区域，并且将第二部分候选区域作为不存在目标的区域，来对目标检测模型进行训练。例如，可以选择与某一目标的匹配程度在一定区间的锚点区域，作为第二候选区域。当目标与锚点区域之间的匹配程度为目标所在的实际区域与锚点区域之间交并比时，可以将交并比为0.02～0.2的锚点区域作为第二候选区域，来对目标检测模型进行训练，以此平衡正负样本的数量，提高对目标检测模型的训练效果。
在一个实施例中，在步骤“按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域”之前，还可以先执行：为样本医学图像的至少一个(例如各个)位 置点生成不同尺寸的预设数量个锚点区域，其中，预设数量个锚点区域的尺寸是分别基于样本医学图像的预设数量个不同尺度第一特征图确定的。
基于样本医学图像的预设数量个不同尺度的第一特征图，可以是利用目标检测模型的特征提取网络对样本医学图像进行特征提取得到的。例如是利用特征图金字塔网络(Feature Pyramid Networks、FPN)、单步多框目标检测SSD(Single Shot MultiBox Detector)模型中的特征提取网络来获得不同尺度的预设数量个第一特征图。可以理解的，获得不同尺度的预设数量个第一特征图的方法不受限制。
在一个实施方式中，可以对应一个第一特征图，在样本医学图像上生成一种尺寸的锚点区域，以此便可以在样本医学图像上生成预设数量个不同尺寸的锚点区域。在一个具体实施方式中，尺寸越大的锚点区域的具体尺寸，是基于尺寸越小的第一特征图确定的。例如，生成的第一特征图的分别为48*48*48，24*24*24，12*12*12，6*6*6，则可以基于48*48*48尺寸的的第一特征图在样本医学图像上生成4*4*4的锚点区域，基于24*24*24的第一特征图在样本医学图像上生成8*8*8的锚点区域，基于12*12*12尺寸的第一特征图在样本医学图像上生成16*16*16的锚点区域，基于6*6*6尺寸的第一特征图在样本医学图像上生成32*32*32的锚点区域。
在一个具体实施方式中，需要获取第一候选区域对应的特征信息时，可以将第一候选区域对应的第一特征图上的特征信息作为与该第一候选区域对应的特征信息。由于预设数量个锚点区域的尺寸是分别基于样本医学图像的预设数量个不同尺度第一特征图确定的，因此，基于某一第一特征图确定的锚点区域即为该锚点区域对应的第一特征图，以此便能相应地确定与第一候选区域的尺寸对应的第一特征图。例如。锚点区域的尺寸是16*16*16，则与该锚点区域尺寸对应的第一特征图的尺寸为12*12*12尺寸，该锚点区域对应的特征信息即是该锚点区域对应的12*12*12尺寸的第一特征图上的区域的特征信息。
因此，通过基于样本医学图像的预设数量个不同尺度的第一特征图来确定设数量个锚点区域的尺寸，可以生成多个锚点区域。
在一个实施例中，在实际的应用过程中，在利用目标检测模型对目标医学图像进行目标检测时，可以直接将在样本医学图像生成的锚点区域作为第一候选区域，进行目标检测。
请参阅图3，图3是本申请目标检测模型的训练方法另一实施例的第一流程示意图。在本实施例中，具体包括步骤S21至步骤S24。
步骤S21：获取包含预设器官的样本医学图像。
关于本步骤的具体描述，请参阅上述步骤S11，此处不再赘述。
步骤S22：利用目标检测模型获取样本医学图像的预设数量个不同尺度的第一特征图。
在本实施例中，预设数量大于或等于1。
利用目标检测模型获取样本医学图像的预设数量个不同尺度的第一特征图，可以是利用目标检测模型的特征提取网络对样本医学图像进行特征提取得到的。例如是利用特征图金字塔网络(Feature Pyramid Networks、FPN)、单步多框目标检测SSD(Single Shot MultiBox Detector)模型中的特征提取网络来获得不同尺度的预设数量个第一特征图。可以理解的，获得不同尺度的预设数量个第一特征图的方法不受限制。在一个具体实施方式中，特征图金字塔网络的自下而上的部分为残差网络(Residual Network,ResNet)。残差网络例如是ResNet18。
步骤S23：利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的最终预测结果。
关于本步骤的具体描述，请参阅上述步骤S12，此处不再赘述。
在一个实施方式中，步骤S23具体包括：基于第一候选区域和第一特征图，预测得到最终预测结果。对于每一个第一候选区域而言，都可以得确定其在第一特征图上的对应区域，并以此确定第一候选区域在第一特征图上的特征信息，后续便可以基于第一候选区域和第一候选区域在第一特征图上的特征信息，得到最终预测结果。
在一个具体实施方式中，因为在样本医学图像上生成了预设数量个锚点区域，因此第一候选区域可以从样本医学图像的不同尺寸的若干锚点区域中选择得到的。并且，不同尺寸的第一候选区域分别对应不同尺度的第一特征图。具体确定的方法和第一候选区域与不同尺度的第一特征图的对应关系，请参阅上述步骤的相关描述，此处不再赘述。
因此，通过生成不同尺度的预设数量个第一特征图，可以利用不同尺寸的第一特征图来对目标检测模型进行训练，以此能够提高目标检测模型对不同大小的目标的检测效果。
在一个实施方式中，上述的步骤“基于第一候选区域和第一特征图，预测得到最终预测结果”，具体包括步骤S231和步骤S232(图未示)。
步骤S231：基于与第一候选区域的尺寸分别对应的第一特征图，获得第一候选区域的特征信息。
第一候选区域的尺寸对应的第一特征图的具体确定方法，以及获得第一候选区域的特征信息的具体方法，请参阅上述步骤的相关描述，此处不再描述。
步骤S232：利用第一候选区域的特征信息，预测得到关于目标的最终预测结果。
在确定第一候选区域的特征信息后，具体可以利用目标检测算法进行目标检测，以预测得到关于目标的最终预测结果。目标检测算法例如是一阶段检测算法或者是二阶段检测算法，此处不再赘述。
因此，通过确定第一候选区域的尺寸对应的第一特征图，并获得第一候选区域的特征信息，以此便能够利用第一候选区域的特征信息进行目标检测，以得到最终预测结果。
在一个具体实施方式中，步骤“利用第一候选区域的特征信息，预测得到关于目标的最终预测结果”具体包括步骤S2321和步骤S2322(图未示)。
步骤S2321：利用第一候选区域的特征信息对第一候选区域进行调整，得到至少一个(例如各个)第一候选区域对应的初始预测结果。
在本实施方式中，第一候选区域对应的初始预测结果包括基于第一候选区域调整得到的目标的初始预测区域。初始预测区域例如是二阶段检测算法得到的Proposal。也即，目标检测模型能够基于第一候选区域的特征信息，对第一候选区域进行回归(调整)，以此得到第一候选区域对应的初始预测结果。
在一个实施方式中，目标检测模型还能够基于第二候选区域的特征信息，对第二候选区域进行回归，以此得到第二候选区域对应的初始预测结果。
步骤S2322：利用至少一个(例如各个)第一候选区域对应的初始预测结果进行优化预测，得到关于目标的最终预测结果。
利用至少一个(例如各个)第一候选区域对应的初始预测结果进行优化预测，具体可以是将初始预测结果的初始预测区域作为感兴趣区域(region of interest，RoI)，进行再次的预测，具体过程可以参阅二阶段检测算法中的相关过程，此处不再赘述。
因此，通过利用第一候选区域的特征信息对第一候选区域进行调整，以得到至少一个(例如各个)第一候选区域对应的初始预测结果，并进一步利用至少一个(例如各个)第一候选区域对应的初始预测结果进行优化预测，得到关于目标的最终预测结果，以此可以目标检测模型对目标进行检测的敏感性，有助于提高目标检测的准确度。
步骤S24：利用最终预测结果与标注结果，调整目标检测模型的参数。
关于本步骤的具体描述，请参阅上述步骤S13，此处不再赘述。
请参阅图4，图4是本申请目标检测模型的训练方法另一实施例的第二流程示意图。在本实施例中，上述步骤“获取样本医学图像的预设数量个不同尺度的第一特征图”具体包括步骤S221和步骤S222。
步骤S221：对样本医学图像进行特征提取，得到不同尺度的预设数量个第二特征图。
对样本医学图像进行特征提取，得到不同尺度的预设数量个第二特征图，具体可以是通过利用特征图金字塔网络(Feature Pyramid Networks、FPN)、单步多框目标检测SSD(Single Shot MultiBox Detector)模型中的特征提取网络来获得，此处不再赘述。
步骤S222：分别对第二特征图进行预设注意力处理，得到第二特征图对应的第一特征图。
在本实施方式中，预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。维度注意力处理例如是目标检测算法中的注意力机制(coordinate attention)处理。特征通道注意力处理例如是目标检测算法中的通道注意力(Channel Attention)处理。此处不再赘述。
因此，通过对第二特征图进行预设注意力处理，有助于目标检测模型提取到更准确的关于目标的特征信息，以此能够目标检测的准确度和召回率。
在一个具体实施方式中，上述步骤S222具体包括步骤S2221和步骤S2222(图未示)。
步骤S2221：获取第二特征图的至少一个(例如各个)维度对应的维度权重，利用至少一个(例如各个)维度对应的维度权重对第二特征图中的至少一个(例如各个)维度特征进行加权处理，得到空间聚焦特征图。
在样本医学图像为二维图像的情况下，可以是第二特征图的X、Y维度对应的维度权重，然后X、Y维度对应的维度权重对第二特征图中的X、Y维度特征进行加权处理，以此得到空间聚焦特征图。在样本医学图像为三维图像的情况下，可以是第二特征图的X、Y、Z维度对应的维度权重，然后X、Y、Z维度对应的维度权重对第二特征图中的X、Y、Z维度特征进行加权处理，以此得到空间聚焦特征图。获得各维度对应的维度权重，可以利用注意力机制(coordinate attention)对第二特征图的各维度的特征信息进行处理后得到的。
在一个具体实施方式中，可以通过以下步骤1和步骤2(图未示)获取第二特征图的至少一个(例如各个)维度对应的维度权重。
步骤1：分别将第二特征图的至少一个(例如各个)维度作为目标维度，对第二特征图进行除目标维度以外的剩余维度的平均池化，得到目标维度上的第三特征图。
例如，样本医学图像是三维图像，则第二特征图也是三维特征图。此时，可以将第二特征图的X、Y、Z维度分别作为为目标维度。在以第二特征图的X维度作为目标维度，然后将Y、Z维度进行平均池化，以此的X维度上的第三特征图。同理，以同样的方法也可以得到Y、Z维度上的第三特征图。
步骤2：利用不同目标维度上的第三特征图，确定第二特征图的至少一个(例如各个)维度对应的维度权重。
在得到至少一个(例如各个)维度上的第三特征图后，利用至少一个(例如各个)维度上的第三特征图确定第二特征图的至少一个(例如各个)维度对应的维度权重。
在一个具体实施方式中，可以将各维度上的第三特征图进行拼接，然后利用带有批标准化和非线性激活的卷积层进行处理，然后卷积层处理后的输出再重新分为各维度上的特征图，随后经过一层卷积和激活后，得到第二特征图的各维度对应的维度权重。例如，在得到X、Y、Z维度上的第三特征图，可以将X、Y、Z维度上的第三特征图进行拼接，然后利用带有批标准化和非线性激活的卷积层进行处理，继而将卷积层处理后的输出再重新分为X、Y、Z维度上的特征图，随后经过一层卷积和激活后，得到第二特征图的X、Y、Z维度对应的维度权重。
在一个实施方式中，可以先将第二特征图在通道维度上切分成若干份，然后分别对每一份进行步骤1和步骤2的处理，在得到每一份的第二特征图的各维度对应的维度权重，再将每一份的结果合并，以此可以得到完整的第二特征图的各维度对应的维度权重。通过将第二特征图在通道维度上切分成若干份，可以降低每次处理的数据量。
因此，通过分别将第二特征图的至少一个(例如各个)维度作为目标维度，对第二特征图进行除目标维度以外的剩余维度的平均池化，得到目标维度上的第三特征图，并利用不同目标维度上的第三特征图，以此可以确定第二特征图的至少一个(例如各个)维度对应的维度权重，有助于目标检测模型提取到更准确的关于目标的特征信息。
步骤S2222：将空间聚焦特征图中的不同通道的特征分为若干份通道特征组，并分别获取通道特征组对应的通道权重，利用通道权重对若干份通道特征组进行加权处理，得到经特征通道注意力处理得到的第一特征图。
将空间聚焦特征图中的不同通道的特征分为若干份通道特征组，例如是，将空间聚焦特征图中的256维的通道特征，划分为四份通道特征组，每一份通道特征组均为64维的通道特征。
在一个具体实施方式中，分别获取通道特征组对应的通道权重具体可以是对至少一份(例如每份)通道特征组进行余弦变换，即对至少一份(例如每一份)通道特征组都做频域通道注意力(Frequency Channel Attention)处理，以此得到与至少一份(例如每一份)通道特征组对应的通道权重。具体的，可以空间聚焦特征图的特征在通道维度上若干等分，对每一份特征乘上余弦级数，得到余弦变换后的频率。随后把这些频率在通道维度上合并，经过一层带激活函数sigmoid激活的全连接层，以此得到通道权重。因此，通过对至少一个(例如每份)通道特征组进行余弦变换，可以提高特征信息的利用率，有助于提高目标检测的准确度。
因此，通过利用第二特征图得到空间聚焦特征图，并进一步地利用空间聚焦特征图得到经特征通道注意力处理得到的第一特征图，有助于目标检测模型在空间维度和通道维度上提取到更准确的关于目标的特征信息。
请参阅图5，图5是本申请目标检测模型的训练方法另一实施例的第三流程示意图。在本实施例中，上述的最终预测结果还包括最终预测区域所属的类别的最终置信度。上述的步骤“利用最终预测结果与标注结果，调整目标检测模型的参数”具体包括步骤S241至步骤S243。
步骤S241：基于最终置信度，得到第一类别损失。
就目标检测而言，对于一个目标，存在着若干个与目标匹配的第一候选区域，基于若干个第一候选区域能相应得到若干个最终预测结果。因此，一个目标会对应多个最终预测结果。此时，可以基于最终置信度，从若干个最终预测结果中选择最优的最终预测结果，以此实现对目标的检测。例如，可以每个最终预测结果的最终置信度，进行非极大值抑制(Non-Maximum Suppression，NMS)处理，以此得到最优的最终预测结果，同时还可以相应确定最优的最终预测结果的分类得分。
然后，可以基于若干个目标的对应的最优的最终预测结果的分类得分，以及第二候选区域对应的最终预测结果的分类得分，计算第一类别损失。例如，可以利用Focal loss函数来得到第一类别损失。
在一个具体实施方式中，可以利用以下公式(1)计算第一类别损失：
Figure PCTCN2022131716-appb-000001
其中，y表示得到的最优的最终预测结果的真实分类信息，y＝1表示最优的最终预测结果为目标，y＝0表示最优的最终预测结果为背景，y′表示最优的最终预测结果的分类得分，γ为损失权重，α为调整权重。
在一个实施方式中，可以将真实分类信息的标签[0,1]软化为[0.1,0.9]，以此能够增强目标检测模型的泛化性能。
步骤S242：基于最终预测区域与实际区域之间的偏移量、以及最终预测区域与实际区域的交并比，得到第一回归损失。
该步骤的最终预测区域可以是步骤S231中确定的一个目标的最优的最终预测结果的最终预测区域。
确定最终预测区域与实际区域之间的偏移量，可以是本领域的通用方法，例如利用smooth-L1损失函数确定最终预测区域与实际区域之间的偏移量。确定最终预测区域与实际区域的交并比方法可以是通用的计算方法，此处不再赘述。
具体的，基于最终预测区域与实际区域之间的偏移量、以及最终预测区域与实际区域的交并比，可以是利用最终预测区域与实际区域的交并比作为调整权重，对最终预测区域与实际区域之间的偏移量进行加权处理，也可以是利用用最终预测区域与实际区域的交并比计算损失值，然后与最终预测区域与实际区域之间的偏移量进行加权求和，得到第一回归损失。
步骤S243：利用第一类别损失和第一回归损失，调整目标检测模型的参数。
在分别得到第一类别损失和第一回归损失后，可以基于这两个损失，确定最终损失，例如是通过加权求和等方式，得到最终损失。并基于最终损失调整目标检测模型的参数。
因此，通过获得第一类别损失和第一回归损失，可以基于目标检测模型的分类损失和预测结果的回归损失，实现对目标检测模型的训练。
在一个实施方式中，最终预测结果是对目标检测模型利用第一候选区域进行预测得到初始预测结果并对初始预测结果进行优化预测得到的。初始预测结果包括目标的初始预测区域和初始预测区域所属的类别的初始置信度。在一个具体实施方式中，初始预测结果还包括与第二候选区域对应的初始预测结果。对应的，上述步骤“利用最终预测结果与标注结果，调整目标检测模型的参数”还包括步骤S234和步骤S235(图未示)。
步骤S244：基于初始置信度，得到第二类别损失。
步骤S245：基于初始预测区域与实际区域之间的偏移量、以及初始预测区域与实际区域的交并比，得到第二回归损失。
关于步骤S244和步骤S245的详细描述，请参阅上述步骤S241和步骤S242，此处不再赘述。
在此情况下，上述步骤“利用第一类别损失和第一回归损失，调整目标检测模型的参数”具体包括：利用第一类别损失、第一回归损失、第二类别损失和第二回归损失，调整目标检测模型的参数。例如，可以利用第一类别损失、第一回归损失得到第一损失，利用第二类别损失和第二回归损失得到第二损失，然后基于第一损失和第二损失，得到最终的损失值，并根据最终的损失值来调整目标检测模型的参数。
因此，通过获得第一类别损失、第一回归损失、第二类别损失和第二回归损失，可以对目标检测模型的参数的调整，以此实现对目标检测模型的训练。
请参阅图6，图6是本申请目标检测模型的训练方法又一实施例的流程示意图。在本实施 例中，上述的基于最终预测区域与实际区域之间的偏移量、以及最终预测区域与实际区域的交并比，得到第一回归损失，得到第二回归损失，或基于初始预测区域与实际区域之间的偏移量、以及初始预测区域与实际区域的交并比，得到第二回归损失，包括步骤S31至步骤S33。
步骤S31：利用相应预测区域对应的偏移量，得到相应预测区域的第一偏移量损失；基于相应预测区域对应的交并比，得到相应预测区域的损失权重，利用相应预测区域的损失权重对相应预测区域的第一偏移量损失进行相乘，得到相应预测区域的第二偏移量损失。
在本实施例中，交并比越大，损失权重越小。以此可以实现对目标对应的最终预测区域与目标所在的实际区域重合度较低的结果以较大的惩罚，使得目标检测模型在优化定位时的参数更新力度更大，有助于提高目标检测的准确度。
在本实施例中，相应预测区域可以是初始预测结果中目标对应的初始预测区域，或者是最终预测结果中目标对应的最终预测区域。相应预测区域对应的交并比，分别可以是初始预测区域与实际区域的交并比，最终预测区域与实际区域的交并比。
在一个具体实施方式中，可以通过以下公式(2)和公式(3)确定第二偏移量损失：
W iou＝(e -iou+0.4)             (2)
Figure PCTCN2022131716-appb-000002
其中，iou为相应预测区域对应的交并比。W iou为相应预测区域的损失权重，公式(3)在基于损失权重W iou进行加权后的smooth-L1损失函数。
通过上述公式(2)和公式(3)，可以分别得到初始预测区域和最终预测区域的第二偏移量损失。
步骤S32：基于相应预测区域对应的交并比，得到相应预测区域的GIOU损失。
在本实施例中，GIOU(Generalized Intersection over Union)损失可以基于一下公式(4)得到：
Figure PCTCN2022131716-appb-000003
其中，A为相应预测区域，B为实际区域，C为相应预测区域与实际区域的最小闭区域。
步骤S33：利用第二偏移量损失和GIOU损失，得到相应回归损失。
具体的，可以利用第二偏移量损失和GIOU损失进行加权相加等方式，得到相应回归损失。例如，可以分别得到第一回归损失或者是第二回归损失。
因此，通过利用第二偏移量损失和GIOU损失，得到相应回归损失，可以训练后的目标检测模型的预测区域定位更加准确。
请参阅图7，图7是本申请目标检测模型的训练方法中目标检测模型的结构示意图。在本实施例中，目标检测模型10包括特征提取模块11、注意力模块12和检测模块13。以下结合目标检测模型10的具体结构，简要描述目标检测模型10的训练方法。
特征提取模块11例如是特征图金字塔网络、SSD模型中的特征提取网络等等。特征提取模块11能够对输入的样本医学图像进行特征提取，获得不同尺度的预设数量个第二特征图。在本实施例中，样本医学图像标注有位于预设器官上的至少一个目标的标注结果，标注结果包括目标所在的实际区域。
注意力模块12可以对至少一个(例如每一个)第二特征图进行预设注意力处理。预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者，以此可以得到第一特征图。在本实施例中，注意力模块12包括空间注意力子模块121和特征注意力子模块122，空间注意力子模块121可以对第二特征图进行维度注意力处理得到的空间聚焦特征图，特征注意力子模块122能够对空间聚焦特征图进行特征通道注意力处理得到的第一特征图。
检测模块13能够按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的最终预测结。在本实施例中，检测模块13包括初始预测子模块131和最终预测子模块132。初始预测子模块131能够利用第一候选区域的特征信息对第一候选区域进行调整，得到至少一个(例如各个)第一候选区域对应的初始预测结果。然后，最终预测子模块132能够利用至少一个(例如各个)第一候选区域对应的初始预测结果进行优化预测，得到关于目标的最终预测结果。例如，最终预测子模块132对初始预测结果进行优化预测，可以是利用最终预测子模块132的感兴趣区域池化(ROI pooling)层进行感兴趣区域池化(ROI pooling)，然后利用两层带非线性激活的全连接层处理，然后再进行检测，以 此得到关于目标的最终预测结果。
请参阅图8，图8是本申请目标检测方法实施例的流程示意图。在本实施例中，目标检测方法具体包括：
步骤S41：获取包含预设器官的目标医学图像。
关于获取包含预设器官的目标医学图像，请参阅上述步骤S11，此处不再赘述。
步骤S42：利用目标检测模型获得目标医学图像的第一特征图，并确定目标的至少一个第一候选区域，并基于第一候选区域和所第一特征图得到关于目标的最终预测结果。
在本实施例中，目标检测模型是利用上述的目标检测模型的训练方法训练得到的。
在本实施例中，在利用目标检测模型进行目标检测时，第一候选区域可以是在目标医学图像上直接生成的锚点区域。另外，基于第一候选区域和所第一特征图得到关于目标的最终预测结果，具体可以是根据第一候选区域在第一特征图上的特征信息进行检测，以此得到关于目标的最终预测结果。确定第一候选区域在第一特征图上的特征信息的基体过程，请参阅上述实施例的相关描述，此处不再赘述。
步骤S43：利用至少一个(例如各个)第一候选区域对应的初始预测结果进行优化预测，得到关于目标的最终预测结果。
关于该步骤的详细描述，请参阅上述步骤S2322以及其他相关描述，此处不再赘述。
因此，通过利用经过上述的目标检测模型的训练方法训练得到的目标检测模型来进行目标检测，能够提升目标检测的准确度和召回率。
在一个具体实施方式中，第一特征图是将对目标医学图像特征提取得到的第二特征图进行预设注意力处理得到的，预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。具体获得第二特征图，以及对第二特征图进行预设注意力处理的相关过程，请参阅上述实施例的相关描述，此处不再赘述。因此，通过对第二特征图进行预设注意力处理，有助于目标检测模型提取到更准确的关于目标的特征信息，以此能够目标检测的准确度和召回率。
请参阅图9，图9是本申请目标检测模型的训练装置的一结构示意图。训练装置90包括获取模块91、检测模块92和调整模块96。获取模块91用于获取包含预设器官的样本医学图像，其中，样本医学图像标注有位于预设器官上的至少一个目标的标注结果，标注结果包括目标所在的实际区域；检测模块92用于利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的最终预测结果，其中，匹配顺序是基于至少一个(例如各个)目标所在的实际区域的尺寸确定的；调整模块96用于利用最终预测结果与标注结果，调整目标检测模型的参数。
其中，上述的匹配顺序为：实际区域的尺寸越小，越早对目标进行匹配。
其中，上述的检测模块92用于利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，包括：基于至少一个(例如各个)目标所在的实际区域的尺寸，将至少一个(例如各个)目标划分至不同目标组，其中，目标组分别对应的尺寸范围不同；基于不同目标组的尺寸范围，确定不同目标组对应的匹配顺序；按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域。
其中，上述的检测模块92用于按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域，包括：按照匹配顺序对至少一个(例如各个)目标组进行如下匹配步骤：获取目标组中的至少一个(例如各个)目标分别与样本医学图像的不同锚点区域之间的匹配程度；基于匹配程度，为目标组的至少一个(例如各个)目标选出至少一个锚点区域作为目标的第一候选区域。
其中，上述的不同目标组中的目标的第一候选区域的数量不同，且尺寸范围越小的目标组中的目标的第一候选区域的数量越多；和/或，目标与锚点区域之间的匹配程度为目标与锚点区域之间的重合度。
其中，上述的检测模块92用于获取目标组中的至少一个(例如各个)目标分别与样本医学图像的不同锚点区域之间的匹配程度，包括：从为样本医学图像生成的多个锚点区域中，选择未作为第一候选区域的锚点区域作为待匹配的锚点区域，并获取目标组中的至少一个(例如各个)目标分别与至少一个(例如各个)待匹配的锚点区域之间的匹配程度；和/或，在上述的检测模块92用于按照匹配顺序分别为至少一个(例如各个)目标组中的目标匹配至少一个第一候选区域之前，检测模块92还用于为样本医学图像的至少一个(例如各个)位置点生成不同尺寸的预设数量个锚点区域，其中，预设数量个锚点区域的尺寸是分别基于样本医学图像的 预设数量个不同尺度第一特征图确定的。
其中，上述的检测模块92用于利用目标检测模型按照匹配顺序分别为至少一个(例如各个)目标匹配至少一个第一候选区域，并基于第一候选区域得到关于目标的预测结果之前，检测模块92还用于利用目标检测模型获取样本医学图像的预设数量个不同尺度的第一特征图，其中，预设数量大于或等于1；上述的检测模块92用于基于第一候选区域得到关于目标的最终预测结果，包括：基于第一候选区域和第一特征图，预测得到最终预测结果。
其中，上述的检测模块92用于获取样本医学图像的预设数量个不同尺度的第一特征图，包括：对样本医学图像进行特征提取，得到不同尺度的预设数量个第二特征图；分别对第二特征图进行预设注意力处理，得到第二特征图对应的第一特征图，其中，预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。
其中，上述的检测模块92用于对第二特征图进行预设注意力处理，得到第二特征图对应的第一特征图，包括：获取第二特征图的至少一个(例如各个)维度对应的维度权重，利用至少一个(例如各个)维度对应的维度权重对第二特征图中的至少一个(例如各个)维度特征进行加权处理，得到空间聚焦特征图；将空间聚焦特征图中的不同通道的特征分为若干份通道特征组，并分别获取通道特征组对应的通道权重，利用通道权重对若干份通道特征组进行加权处理，得到经特征通道注意力处理得到的第一特征图。
其中，上述的检测模块92用于获取第二特征图的至少一个(例如各个)维度对应的维度权重，包括：分别将第二特征图的至少一个(例如各个)维度作为目标维度，对第二特征图进行除目标维度以外的剩余维度的平均池化，得到目标维度上的第三特征图；利用不同目标维度上的第三特征图，确定第二特征图的至少一个(例如各个)维度对应的维度权重；和/或，分别获取通道特征组对应的通道权重，包括：对至少一份(例如每一份)通道特征组进行余弦变换，以得到与至少一份(例如每一份)通道特征组对应的通道权重。
其中，上述的至少第一候选区域是从样本医学图像的不同尺寸的若干锚点区域中选择得到的，不同尺寸的第一候选区域分别对应不同尺度的第一特征图；上述的检测模块92用于基于第一候选区域和第一特征图，预测得到最终预测结果，包括：基于与第一候选区域的尺寸分别对应的第一特征图，获得第一候选区域的特征信息；利用第一候选区域的特征信息，预测得到关于目标的最终预测结果。
其中，上述的检测模块92用于利用第一候选区域的特征信息，预测得到关于目标的最终预测结果，包括：利用第一候选区域的特征信息对第一候选区域进行调整，得到至少一个(例如各个)第一候选区域对应的初始预测结果，其中，第一候选区域对应的初始预测结果包括基于第一候选区域调整得到的目标的初始预测区域；利用至少一个(例如各个)第一候选区域对应的初始预测结果进行优化预测，得到关于目标的最终预测结果。
其中，上述的最终预测结果还包括最终预测区域所属的类别的最终置信度；上述的调整模块96用于用最终预测结果与标注结果，调整目标检测模型的参数，包括：基于最终置信度，得到第一类别损失；基于最终预测区域与实际区域之间的偏移量、以及最终预测区域与实际区域的交并比，得到第一回归损失；利用第一类别损失和第一回归损失，调整目标检测模型的参数。
其中，上述的最终预测结果是对目标检测模型利用第一候选区域进行预测得到初始预测结果并对初始预测结果进行优化预测得到的，初始预测结果包括目标的初始预测区域和初始预测区域所属的类别的初始置信度；上述的调整模块96用于利用最终预测结果与标注结果，调整目标检测模型的参数，还包括：基于初始置信度，得到第二类别损失；基于初始预测区域与实际区域之间的偏移量、以及初始预测区域与实际区域的交并比，得到第二回归损失；上述的调整模块96用于利用第一类别损失和第一回归损失，调整目标检测模型的参数，包括：利用第一类别损失、第一回归损失、第二类别损失和第二回归损失，调整目标检测模型的参数。
其中，上述的调整模块96用于基于最终预测区域与实际区域之间的偏移量、以及最终预测区域与实际区域的交并比，得到第一回归损失，或基于初始预测区域与实际区域之间的偏移量、以及初始预测区域与实际区域的交并比，得到第二回归损失，包括：利用相应预测区域对应的偏移量，得到相应预测区域的第一偏移量损失；基于相应预测区域对应的交并比，得到相应预测区域的损失权重，利用相应预测区域的损失权重对相应预测区域的第一偏移量损失进行相乘，得到相应预测区域的第二偏移量损失，其中，交并比越大，损失权重越小；基于相应预测区域对应的交并比，得到相应预测区域的GIOU损失；利用第二偏移量损失和GIOU损失， 得到相应回归损失。
其中，上述的样本医学图像为三维医学图像；和/或，预设器官为肺部，目标为结节。
请参阅图10，图10是本申请目标检测装置的一结构示意图。目标检测装置100包括获取模块101、检测模块102。获取模块101用于获取包含预设器官的目标医学图像；检测模块102用于利用目标检测模型获得目标医学图像的第一特征图，并确定目标的至少一个第一候选区域，并基于第一候选区域和所第一特征图得到关于目标的最终预测结果；其中，目标检测模型100是利用上述目标检测模型的训练方法训练得到的，和/或，第一特征图是将对目标医学图像特征提取得到的第二特征图进行预设注意力处理得到的，预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。
本申请方法步骤的执行主体可以为硬件执行，或者通过处理器运行计算机可执行代码的方式执行。
请参阅图11，图11是本申请电子设备一实施例的框架示意图。电子设备110包括相互耦接的存储器111和处理器112，处理器112用于执行存储器111中存储的程序指令，以实现上述任一目标检测模型的训练方法实施例的步骤，或实现上述任一目标检测方法实施例中的步骤。在一个具体的实施场景中，电子设备110可以包括但不限于：微型计算机、服务器，此外，电子设备110还可以包括笔记本电脑、平板电脑等移动设备，在此不做限定。
具体而言，处理器112用于控制其自身以及存储器111以实现上述任一图像分割模型的训练方法实施例的步骤，或实现上述任一图像分割方法实施例中的步骤。处理器112还可以称为CPU(Central Processing Unit，中央处理单元)。处理器112可能是一种集成电路芯片，具有信号的处理能力。处理器112还可以是通用处理器、数字信号处理器(Digital Signal Processor,DSP)、专用集成电路(Application Specific Integrated Circuit,ASIC)、现场可编程门阵列(Field-Programmable Gate Array,FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。另外，处理器112可以由集成电路芯片共同实现。
请参阅图12，图12为本申请计算机可读存储介质一实施例的框架示意图。计算机可读存储介质120存储有能够被处理器运行的程序指令121，程序指令121用于实现上述任一目标检测模型的训练方法实施例的步骤，或实现上述任一目标检测方法实施例中的步骤。
其中，计算机可读存储介质可为易失性存储介质或非易失性存储介质。
在一些实施例中，本公开实施例提供的装置具有的功能或包含的模块可以用于执行上文方法实施例描述的方法，其具体实现可以参照上文方法实施例的描述，为了简洁，这里不再赘述。
上文对各个实施例的描述倾向于强调各个实施例之间的不同之处，其相同或相似之处可以互相参考，为了简洁，本文不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的方法和装置，可以通过其它的方式实现。例如，以上所描述的装置实施方式仅仅是示意性的，例如，模块或单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性、机械或其它的形式。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)或处理器(processor)执行本申请各个实施方式方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
本公开实施例还提供一种计算机程序产品，该计算机程序产品承载有程序代码，包括计算机可读代码，或者承载有计算机可读代码的计算机可读存储介质，当所述计算机可读代码在电 子设备的处理器中运行时，所述程序代码包括的指令可用于执行上述方法实施例中所述方法的步骤，具体可参见上述方法实施例，在此不再赘述。
其中，上述计算机程序产品可以具体通过硬件、软件或其结合的方式实现。在一个可选实施例中，所述计算机程序产品具体体现为计算机存储介质，在另一个可选实施例中，计算机程序产品具体体现为软件产品，例如软件开发包(Software Development Kit，SDK)等等。
若本申请技术方案涉及个人信息，应用本申请技术方案的产品在处理个人信息前，已明确告知个人信息处理规则，并取得个人自主同意。若本申请技术方案涉及敏感个人信息，应用本申请技术方案的产品在处理敏感个人信息前，已取得个人单独同意，并且同时满足“明示同意”的要求。例如，在摄像头等个人信息采集装置处，设置明确显著的标识告知已进入个人信息采集范围，将会对个人信息进行采集，若个人自愿进入采集范围即视为同意对其个人信息进行采集；或者在个人信息处理的装置上，利用明显的标识/信息告知个人信息处理规则的情况下，通过弹窗信息或请个人自行上传其个人信息等方式获得个人授权；其中，个人信息处理规则可包括个人信息处理者、个人信息处理目的、处理方式以及处理的个人信息种类等信息。

Claims (22)

  1. 一种目标检测模型的训练方法，其特征在于，所述训练方法包括：
    获取包含预设器官的样本医学图像，其中，所述样本医学图像标注有位于所述预设器官上的至少一个目标的标注结果，所述标注结果包括所述目标所在的实际区域；
    利用所述目标检测模型按照匹配顺序分别为至少一个所述目标匹配至少一个第一候选区域，并基于所述第一候选区域得到关于所述目标的最终预测结果，其中，所述匹配顺序是基于至少一个所述目标所在的实际区域的尺寸确定的；
    利用所述最终预测结果与所述标注结果，调整所述目标检测模型的参数。
  2. 根据权利要求1所述的方法，其特征在于，所述匹配顺序为：所述实际区域的尺寸越小，越早对所述目标进行所述匹配。
  3. 根据权利要求1或2所述的方法，其特征在于，所述利用所述目标检测模型按照匹配顺序分别为至少一个所述目标匹配至少一个第一候选区域，包括：
    基于至少一个所述目标所在的实际区域的尺寸，将至少一个所述目标划分至不同目标组，其中，所述目标组分别对应的尺寸范围不同；
    基于不同所述目标组的尺寸范围，确定不同所述目标组对应的所述匹配顺序；
    按照所述匹配顺序分别为至少一个所述目标组中的目标匹配至少一个第一候选区域。
  4. 根据权利要求3所述的方法，其特征在于，所述按照所述匹配顺序分别为至少一个所述目标组中的目标匹配至少一个第一候选区域，包括：
    按照所述匹配顺序对至少一个所述目标组进行如下匹配步骤：
    获取所述目标组中的至少一个所述目标分别与所述样本医学图像的不同锚点区域之间的匹配程度；
    基于所述匹配程度，为所述目标组的至少一个所述目标选出至少一个所述锚点区域作为所述目标的第一候选区域。
  5. 根据权利要求4所述的方法，其特征在于，不同所述目标组中的目标的所述第一候选区域的数量不同，且所述尺寸范围越小的所述目标组中的目标的第一候选区域的数量越多；
    和/或，所述目标与所述锚点区域之间的匹配程度为所述目标与所述锚点区域之间的重合度。
  6. 根据权利要求4所述的方法，其特征在于，所述获取所述目标组中的至少一个所述目标分别与所述样本医学图像的不同锚点区域之间的匹配程度，包括：从为所述样本医学图像生成的多个锚点区域中，选择未作为所述第一候选区域的锚点区域作为待匹配的锚点区域，并获取所述目标组中的至少一个所述目标分别与至少一个所述待匹配的锚点区域之间的匹配程度；
    和/或，在所述按照所述匹配顺序分别为至少一个所述目标组中的目标匹配至少一个第一候选区域之前，所述方法还包括：为所述样本医学图像的至少一个位置点生成不同尺寸的预设数量个所述锚点区域，其中，所述预设数量个锚点区域的尺寸是分别基于所述样本医学图像的预设数量个不同尺度的第一特征图确定的。
  7. 根据权利要求1至6任一项所述的方法，其特征在于，所述利用所述目标检测模型按照匹配顺序分别为至少一个所述目标匹配至少一个第一候选区域，并基于所述第一候选区域得到关于所述目标的预测结果之前，所述方法还包括：
    利用所述目标检测模型获取所述样本医学图像的预设数量个不同尺度的第一特征图，其中，所述预设数量大于或等于1；
    所述基于所述第一候选区域得到关于所述目标的最终预测结果，包括：
    基于所述第一候选区域和所述第一特征图，预测得到所述最终预测结果。
  8. 根据权利要求7所述的方法，其特征在于，所述获取所述样本医学图像的预设数量个不同尺度的第一特征图，包括：
    对所述样本医学图像进行特征提取，得到预设数量个不同尺度的第二特征图；
    分别对所述第二特征图进行预设注意力处理，得到所述第二特征图对应的第一特征图，其中，所述预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。
  9. 根据权利要求8所述的方法，其特征在于，所述对所述第二特征图进行预设注意力处理，得到所述第二特征图对应的第一特征图，包括：
    获取所述第二特征图的至少一个维度对应的维度权重，利用所述至少一个维度对应的维度权重对所述第二特征图中的至少一个维度特征进行加权处理，得到空间聚焦特征图；
    将所述空间聚焦特征图中的不同通道的特征分为若干份通道特征组，并分别获取所述通道特征组对应的通道权重，利用所述通道权重对所述若干份通道特征组进行加权处理，得到经所述特征通道注意力处理得到的所述第一特征图。
  10. 根据权利要求9所述的方法，其特征在于，所述获取所述第二特征图的至少一个维度对应的维度权重，包括：
    分别将所述第二特征图的至少一个维度作为目标维度，对所述第二特征图进行除所述目标维度以外的剩余维度的平均池化，得到所述目标维度上的第三特征图；
    利用不同所述目标维度上的所述第三特征图，确定所述第二特征图的至少一个维度对应的维度权重；
    和/或，所述分别获取所述通道特征组对应的通道权重，包括：
    对至少一份所述通道特征组进行余弦变换，以得到与至少一份所述通道特征组对应的通道权重。
  11. 根据权利要求7所述的方法，其特征在于，所述至少一个第一候选区域是从所述样本医学图像的不同尺寸的若干锚点区域中选择得到的，不同尺寸的所述第一候选区域分别对应不同尺度的所述第一特征图；
    所述基于所述第一候选区域和所述第一特征图，预测得到所述最终预测结果，包括：
    基于与所述第一候选区域的尺寸分别对应的所述第一特征图，获得所述第一候选区域的特征信息；
    利用所述第一候选区域的特征信息，预测得到关于所述目标的最终预测结果。
  12. 根据权利要求11所述的方法，其特征在于，所述利用所述第一候选区域的特征信息，预测得到关于所述目标的最终预测结果，包括：
    利用所述第一候选区域的特征信息对所述第一候选区域进行调整，得到至少一个第一候选区域对应的初始预测结果，其中，所述第一候选区域对应的初始预测结果包括基于所述第一候选区域调整得到的所述目标的初始预测区域；
    利用所述至少一个第一候选区域对应的初始预测结果进行优化预测，得到关于所述目标的最终预测结果。
  13. 根据权利要求1所述的方法，其特征在于，所述最终预测结果还包括所述最终预测区域所属的类别的最终置信度；所述利用所述最终预测结果与所述标注结果，调整所述目标检测模型的参数，包括：
    基于所述最终置信度，得到第一类别损失；
    基于所述最终预测区域与实际区域之间的偏移量、以及所述最终预测区域与实际区域的交并比，得到第一回归损失；
    利用所述第一类别损失和第一回归损失，调整目标检测模型的参数。
  14. 根据权利要求13所述的方法，其特征在于，所述最终预测结果是对所述目标检测模型利用所述第一候选区域进行预测得到初始预测结果并对所述初始预测结果进行优化预测得到的，所述初始预测结果包括所述目标的初始预测区域和所述初始预测区域所属的类别的初始置信度；
    所述利用所述最终预测结果与所述标注结果，调整所述目标检测模型的参数，还包括：
    基于所述初始置信度，得到第二类别损失；
    基于所述初始预测区域与实际区域之间的偏移量、以及所述初始预测区域与实际区域的交并比，得到第二回归损失；
    所述利用所述第一类别损失和第一回归损失，调整目标检测模型的参数，包括：
    利用所述第一类别损失、第一回归损失、第二类别损失和第二回归损失，调整目标检测模型的参数。
  15. 根据权利要求13或14所述的方法，其特征在于，所述基于所述最终预测区域与实际区域之间的偏移量、以及所述最终预测区域与实际区域的交并比，得到第一回归损失，或基于所述初始预测区域与实际区域之间的偏移量、以及所述初始预测区域与实际区域的交并比，得到第二回归损失，包括：
    利用相应预测区域对应的所述偏移量，得到所述相应预测区域的第一偏移量损失；基于所述相应预测区域对应的交并比，得到所述相应预测区域的损失权重，利用所述相应预测区域的损失权重对所述相应预测区域的第一偏移量损失进行相乘，得到所述相应预测区域的第二偏移量损失，其中，所述交并比越大，所述损失权重越小；
    基于所述相应预测区域对应的交并比，得到所述相应预测区域的GIOU损失；
    利用所述第二偏移量损失和所述GIOU损失，得到相应回归损失。
  16. 根据权利要求1至15任一项所述的方法，其特征在于，所述样本医学图像为三维医学图像；和/或，
    所述预设器官为肺部，所述目标为结节。
  17. 一种目标检测方法，其特征在于，包括：
    获取包含预设器官的目标医学图像；
    利用目标检测模型获得所述目标医学图像的第一特征图，并确定所述目标的至少一个第一候选区域，并基于所述第一候选区域和所第一特征图得到关于所述目标的最终预测结果；
    其中，所述目标检测模型是利用权利要求1至16任一项所述的方法训练得到的，和/或，所述第一特征图是将对所述目标医学图像特征提取得到的第二特征图进行预设注意力处理得到的，所述预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。
  18. 一种目标检测模型的训练装置，其特征在于，包括：
    获取模块，用于获取包含预设器官的样本医学图像，其中，所述样本医学图像标注有位于所述预设器官上的至少一个目标的标注结果，所述标注结果包括所述目标所在的实际区域；
    检测模块，用于利用所述目标检测模型按照匹配顺序分别为至少一个所述目标匹配至少一个第一候选区域，并基于所述第一候选区域得到关于所述目标的最终预测结果，其中，所述匹配顺序是基于至少一个所述目标所在的实际区域的尺寸确定的；
    调整模块，用于利用所述最终预测结果与所述标注结果，调整所述目标检测模型的参数。
  19. 一种目标检测装置，其特征在于，包括：
    获取模块，用于获取包含预设器官的目标医学图像；
    检测模块，用于利用目标检测模型获得所述目标医学图像的第一特征图，并确定所述目标的至少一个第一候选区域，并基于所述第一候选区域和所第一特征图得到关于所述目标的最终预测结果；
    其中，所述目标检测模型是利用权利要求1至16任一项所述的方法训练得到的，和/或，所述第一特征图是将对所述目标医学图像特征提取得到的第二特征图进行预设注意力处理得到的，所述预设注意力处理包括维度注意力处理和特征通道注意力处理中的一者或多者。
  20. 一种电子设备，其特征在于，包括相互耦接的存储器和处理器，所述处理器用于执行所述存储器中存储的程序指令，以实现权利要求1至16任一项所述的目标检测模型的训练方法，或实现权利要求17所述的目标检测方法。
  21. 一种计算机可读存储介质，其上存储有程序指令，其特征在于，所述程序指令被处理器执行时实现权利要求1至16任一项所述的目标检测模型的训练方法，或实现权利要求17所述的目标检测方法。
  22. 一种计算机程序产品，包括计算机可读代码，或者承载有计算机可读代码的计算机可读存储介质，当所述计算机可读代码在电子设备的处理器中运行时，所述电子设备中的处理器执行用于实现权利要求1至16任一项所述的目标检测模型的训练方法，或实现权利要求17所述的目标检测方法。
PCT/CN2022/131716 2022-01-24 2022-11-14 目标检测模型的训练方法及对应的检测方法 WO2023138190A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210080240.1 2022-01-24
CN202210080240.1A CN114429459A (zh) 2022-01-24 2022-01-24 目标检测模型的训练方法及对应的检测方法

Publications (1)

Publication Number Publication Date
WO2023138190A1 true WO2023138190A1 (zh) 2023-07-27

Family

ID=81313424

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/131716 WO2023138190A1 (zh) 2022-01-24 2022-11-14 目标检测模型的训练方法及对应的检测方法

Country Status (2)

Country Link
CN (1) CN114429459A (zh)
WO (1) WO2023138190A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117036985A (zh) * 2023-10-09 2023-11-10 武汉工程大学 一种面向视频卫星图像的小目标检测方法及装置

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114429459A (zh) * 2022-01-24 2022-05-03 上海商汤智能科技有限公司 目标检测模型的训练方法及对应的检测方法
CN115131655B (zh) * 2022-09-01 2022-11-22 浙江啄云智能科技有限公司 目标检测模型的训练方法、装置与目标检测方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109658419A (zh) * 2018-11-15 2019-04-19 浙江大学 一种医学图像中小器官的分割方法
CN111539947A (zh) * 2020-04-30 2020-08-14 上海商汤智能科技有限公司 图像检测方法及相关模型的训练方法和相关装置、设备
WO2020164493A1 (zh) * 2019-02-14 2020-08-20 腾讯科技（深圳）有限公司 一种医学图像区域过滤方法、装置及存储介质
US20210343016A1 (en) * 2019-05-22 2021-11-04 Tencent Technology (Shenzhen) Company Limited Medical image processing method and apparatus, electronic medical device, and storage medium
CN114429459A (zh) * 2022-01-24 2022-05-03 上海商汤智能科技有限公司 目标检测模型的训练方法及对应的检测方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109658419A (zh) * 2018-11-15 2019-04-19 浙江大学 一种医学图像中小器官的分割方法
WO2020164493A1 (zh) * 2019-02-14 2020-08-20 腾讯科技（深圳）有限公司 一种医学图像区域过滤方法、装置及存储介质
US20210343016A1 (en) * 2019-05-22 2021-11-04 Tencent Technology (Shenzhen) Company Limited Medical image processing method and apparatus, electronic medical device, and storage medium
CN111539947A (zh) * 2020-04-30 2020-08-14 上海商汤智能科技有限公司 图像检测方法及相关模型的训练方法和相关装置、设备
CN114429459A (zh) * 2022-01-24 2022-05-03 上海商汤智能科技有限公司 目标检测模型的训练方法及对应的检测方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117036985A (zh) * 2023-10-09 2023-11-10 武汉工程大学 一种面向视频卫星图像的小目标检测方法及装置
CN117036985B (zh) * 2023-10-09 2024-02-06 武汉工程大学 一种面向视频卫星图像的小目标检测方法及装置

Also Published As

Publication number Publication date
CN114429459A (zh) 2022-05-03

Similar Documents

Publication Publication Date Title
WO2023138190A1 (zh) 目标检测模型的训练方法及对应的检测方法
CN111401201B (zh) 一种基于空间金字塔注意力驱动的航拍图像多尺度目标检测方法
Xue et al. An application of transfer learning and ensemble learning techniques for cervical histopathology image classification
WO2021179498A1 (zh) 目标检测方法及其模型的训练方法、装置及电子设备
WO2021128825A1 (zh) 三维目标检测及模型的训练方法及装置、设备、存储介质
Nurmaini et al. Accurate detection of septal defects with fetal ultrasonography images using deep learning-based multiclass instance segmentation
Thian et al. Deep learning systems for pneumothorax detection on chest radiographs: a multicenter external validation study
WO2023035586A1 (zh) 图像检测方法、模型训练方法、装置、设备、介质及程序
WO2023155494A1 (zh) 图像检测及训练方法、相关装置、设备、介质和程序产品
CN112614133B (zh) 一种无锚点框的三维肺结节检测模型训练方法及装置
US11449717B2 (en) System and method for identification and localization of images using triplet loss and predicted regions
WO2022242127A1 (zh) 图像特征提取方法、装置、电子设备及存储介质
WO2023115797A1 (zh) 图像分割方法、装置、设备及存储介质
Zhou et al. Learn fine-grained adaptive loss for multiple anatomical landmark detection in medical images
Tang et al. Lesion segmentation and RECIST diameter prediction via click-driven attention and dual-path connection
Li et al. Residual attention learning network and SVM for malaria parasite detection
CN110414562B (zh) X光片的分类方法、装置、终端及存储介质
Wang et al. Renal lesion classification in kidney CT images by seven-layer convolution neural network
Lantang et al. Investigation of the efficiency of an interconnected convolutional neural network by classifying medical images
CN117974672B (zh) 基于全局信息的乳腺超声肿瘤病变区域检测方法及装置
CN114549853A (zh) 图像处理方法及相关模型训练方法、装置、设备
Pant et al. X-rays imaging analysis for early diagnosis of thoracic disorders using capsule neural network: a deep learning approach
Subashini et al. Detection of Covid-19 using Mask R-CNN
WO2024045819A1 (zh) 病变区域确定方法、模型训练方法及装置
Lu Convolutional Neural Network (CNN) for COVID-19 Lung CT Scans Classification Detection

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22921614

Country of ref document: EP

Kind code of ref document: A1