WO2021027135A1 - 细胞检测模型训练方法、装置、计算机设备及存储介质 - Google Patents
细胞检测模型训练方法、装置、计算机设备及存储介质 Download PDFInfo
- Publication number
- WO2021027135A1 WO2021027135A1 PCT/CN2019/117233 CN2019117233W WO2021027135A1 WO 2021027135 A1 WO2021027135 A1 WO 2021027135A1 CN 2019117233 W CN2019117233 W CN 2019117233W WO 2021027135 A1 WO2021027135 A1 WO 2021027135A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- ratio threshold
- sample
- image
- region
- parameter
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- 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
- G06T2207/30096—Tumor; Lesion
Definitions
- This application relates to the field of cell detection technology, and in particular to a cell detection model training method, device, computer equipment and storage medium.
- Cervical cancer is a cancer that can be detected and cured early. Most of the deaths caused each year occur in areas with low cervical cancer screening rates. Therefore, the detection of cervical cancer cells is very important.
- deep learning neural networks can be used to extract features from abnormal cell images and train models to detect abnormal cells.
- the model trained by deep learning neural network usually selects 0.5 as the intersection ratio threshold, so that the trained model generates a detection frame that is not accurate enough when detecting abnormal cells.
- the first aspect of the present application provides a method for training a cell detection model, the method including:
- a target model is trained using multiple sample images, where the sample images include images containing abnormal cells and images not containing abnormal cells;
- the target intersection ratio threshold includes a first intersection ratio threshold, a second A merge ratio threshold and a third merge ratio threshold, where the third merge ratio threshold is greater than the second merge ratio threshold, and the second merge ratio threshold is greater than the first merge ratio threshold;
- the neural network is trained to obtain a trained cell detection model.
- a second aspect of the present application provides a cell detection model training device, which includes:
- the training module is used to train a target model using multiple sample images for each preset intersection ratio threshold, wherein the sample images include images containing abnormal cells and images not containing abnormal cells;
- the determining module is configured to determine the correct rate of the target model by using the free response receiver operating characteristic curve method, and determine the target intersection ratio threshold value according to the correct rate, wherein the target intersection ratio threshold includes the first intersection ratio The ratio threshold, the second intersection ratio threshold, and the third intersection ratio threshold, the third intersection ratio threshold greater than the second intersection ratio threshold, and the second intersection ratio threshold greater than the first intersection ratio threshold Ratio threshold
- the training module is further configured to train the neural network according to the first intersection ratio threshold and the multiple sample images to obtain the first parameter;
- a sampling module configured to resample the sample image according to the first parameter to obtain a first sample
- the training module is further configured to train the neural network according to the second intersection ratio threshold and the first sample to obtain a second parameter
- the sampling module is further configured to resample the sample image according to the second parameter and the third cross-union ratio threshold to obtain a second sample;
- the training module is also used to train the neural network according to the second sample to obtain a trained cell detection model.
- a third aspect of the present application provides a computer device that includes a processor and a memory, and the processor is configured to implement the cell detection model training method when executing computer-readable instructions stored in the memory.
- a fourth aspect of the present application provides a non-volatile readable storage medium having computer readable instructions stored on the non-volatile readable storage medium, and when the computer readable instructions are executed by a processor, the Training method of cell detection model.
- this application ensures that the accuracy of the trained cell detection model is relatively high by selecting the target intersection ratio threshold, and re-sampling the sample image by increasing the intersection ratio threshold to ensure that the number of positive samples is sufficient. Avoid over-fitting, thereby improving the accuracy of the cell detection model's prediction frame for abnormal cells, that is, improving the accuracy of detecting abnormal cells.
- Fig. 1 is a flowchart of a preferred embodiment of a cell detection model training method disclosed in the present application.
- Fig. 2 is a functional block diagram of a preferred embodiment of a cell detection model training device disclosed in the present application.
- FIG. 3 is a schematic structural diagram of a computer device in a preferred embodiment of the method for training a cell detection model according to the present application.
- the cell detection model training method of the embodiment of the present application is applied to a computer device, and can also be applied to a hardware environment composed of a computer device and a server connected to the computer device through a network, and is executed by the server and the computer device.
- Networks include but are not limited to: wide area network, metropolitan area network or local area network.
- FIG. 1 is a flowchart of a preferred embodiment of a cell detection model training method disclosed in the present application. Among them, according to different needs, the order of the steps in the flowchart can be changed, and some steps can be omitted.
- the computer device uses a plurality of sample images to train the target model for each preset intersection ratio threshold, where the sample images include images containing abnormal cells and images not containing abnormal cells.
- intersection over union refers to the overlap rate between the generated candidate frame and the label frame in the target detection type model.
- the positive and negative samples are mainly determined according to the intersection ratio threshold.
- the intersection ratio threshold is 0.5, which is IOU >0.5 is regarded as a positive sample.
- the bounding box (coordinate information) and category of abnormal cells are marked on the sample image.
- the target model refers to a model trained using a single intersection ratio threshold to detect abnormal cells.
- multiple intersection ratio thresholds can be preset, and then the target model corresponding to each intersection ratio threshold is trained.
- the training process is the training of the neural network.
- the neural network generates multiple candidate regions based on the input sample image, and then compares the bounding box of the candidate region with the bounding box of the marked abnormal cell (labeled box), and compares the intersection
- the candidate area greater than the intersection ratio threshold is determined as a positive sample, and the bounding box coordinates and category of the positive sample are regressed.
- the loss value of the loss function reaches the convergence state, it is determined that the target model is trained.
- intersection ratio threshold the higher the quality of the positive samples, and the better the performance of the trained model.
- the intersection ratio is too high, the number of positive samples will be too small, which will lead to overfitting in training. , The performance of the trained model decreases instead. After multiple target models are trained, each target model can be evaluated to determine the intersection ratio threshold corresponding to the target model with better performance.
- the computer device uses a free response receiver operating characteristic curve method to determine the correct rate of the target model, and determines a target intersection ratio threshold according to the correct rate, wherein the target intersection ratio threshold includes the first intersection ratio Threshold, a second intersection ratio threshold, and a third intersection ratio threshold, the third intersection ratio threshold is greater than the second intersection ratio threshold, and the second intersection ratio threshold is greater than the first intersection ratio threshold Threshold.
- the free-response receiver operating characteristic curve (Free-response Receiver Operating Characteristic Curves, FROC) is a variant of the receiver operating characteristic curve (Receiver Operating Characteristic Curves, ROC).
- AUC Average under The curve (area under the curve) value can determine the correct rate of a model, but the ROC method cannot solve the evaluation of multiple anomalies on an image, while the FROC method can be used to evaluate multiple anomalies on an image.
- the target intersection ratio threshold is used to optimize the sample.
- the correct rate of different target models (cell detection models) trained according to different cross-combination ratio thresholds can be determined by the FROC method. Determine the intersection ratio thresholds corresponding to the three target models with the highest correct rate as the target intersection ratio thresholds: the first intersection ratio threshold, the second intersection ratio threshold, and the third intersection ratio threshold, where the third The intersection ratio threshold is greater than the second intersection ratio threshold, and the second intersection ratio threshold is greater than the first intersection ratio threshold.
- the computer device trains the neural network according to the first intersection ratio threshold and the multiple sample images to obtain the first parameter.
- the neural network includes, but is not limited to: residual convolutional network (residual network, ResNet), feature pyramid network (Feature Pyramid Networks, FPN), and regional candidate network (Region Proposal Network, RPN).
- ResNet residual convolutional network
- FPN feature pyramid network
- RPN regional candidate network
- the neural network can be trained according to the first intersection ratio threshold and the multiple sample images to obtain the first parameter, because the detection performance of the neural network trained by the first intersection ratio threshold It is better to ensure that the correct rate of the subsequent training model is relatively high.
- the training a neural network according to the first intersection ratio threshold and the plurality of sample images to obtain the first parameter includes:
- the feature information use a feature pyramid network to generate a first feature map
- the first feature map use a region candidate network to generate a candidate region
- regression is performed on the bounding box and category of the first region to obtain the first parameter.
- the residual convolution network can be used to perform feature extraction on the image to obtain the feature information of the image, and then the feature information of the image is input to the feature pyramid network to obtain the feature map of the image (the first feature map). ), the regional candidate network processes the feature map to generate multiple candidate regions. At this time, the candidate region may be screened according to the first intersection ratio threshold to obtain the first region, that is, the positive sample.
- the feature map of the image and the first region can be input to the region of interest (ROI) pooling layer to obtain the feature map (second feature map) corresponding to the first region.
- ROI Align interest Area alignment
- the bounding box (coordinates) and category of the first region can be regressed to obtain the parameters (first parameters) of the neural network.
- the first parameter refers to each parameter of the entire neural network after training the candidate frame selected according to the first intersection ratio threshold.
- the computer device resamples the sample image according to the first parameter to obtain a first sample.
- the sample image is resampled by the trained first parameter, that is, the trained neural network is used to process the feature map of the previous candidate area network, and the generated candidate frame (the first Region) to optimize, obtain a new batch of bounding boxes, the intersection ratio of the optimized bounding box is higher, to ensure that after increasing the intersection ratio threshold, the number of positive samples is sufficient for training, which can improve the accuracy of the prediction box and prevent Overfitting.
- the trained first parameter that is, the trained neural network is used to process the feature map of the previous candidate area network, and the generated candidate frame (the first Region) to optimize, obtain a new batch of bounding boxes, the intersection ratio of the optimized bounding box is higher, to ensure that after increasing the intersection ratio threshold, the number of positive samples is sufficient for training, which can improve the accuracy of the prediction box and prevent Overfitting.
- the re-sampling the sample image according to the first parameter to obtain the first sample includes:
- the third region and the third feature map are determined as the first sample.
- the neural network may optimize the first region according to the first parameter, generate a second region, and filter the second region according to the second intersection ratio threshold,
- the second area whose intersection ratio is greater than the second intersection ratio threshold is determined as the third area
- the feature map of the image and the third area can be input to the region of interest pooling layer to obtain the corresponding The feature map (the third feature map), where ROI Align can be used to achieve uniformity in the size of the feature map.
- the third region and the third feature map are determined as the first sample (positive sample).
- the computer device trains the neural network according to the second intersection ratio threshold and the first sample to obtain a second parameter.
- the bounding box (coordinates) and category of the first sample can be regressed according to the feature map in the first sample to obtain the parameter (second parameter) of the neural network, and the second parameter is Refers to the retraining of the neural network, and each parameter obtained is the parameter obtained after optimizing the first parameter.
- the computer device resamples the sample image according to the second parameter and the third intersection ratio threshold to obtain a second sample.
- the sample image can be re-sampled through the trained second parameter, that is, the previously generated candidate frame (the third region) is optimized using the trained neural network to obtain a new one.
- the intersection ratio of the bounding box obtained after optimization is higher. After increasing the intersection ratio threshold, the number of positive samples is sufficient for training, which can continue to improve the accuracy of the prediction box and prevent overfitting.
- the re-sampling the sample image according to the second parameter and the third intersection ratio threshold to obtain the second sample includes:
- the fifth region and the fourth feature map are determined as the second sample.
- the neural network may optimize the third region based on the second parameter to generate a fourth region, and filter the fourth region based on the third intersection ratio threshold,
- the fourth area whose intersection ratio is greater than the third intersection ratio threshold is determined as the fifth area
- the feature map of the image and the fifth area can be input to the region of interest pooling layer to obtain the corresponding Feature map (fourth feature map), where ROI Align can be used to achieve uniformity of feature map size.
- the fifth region and the fourth feature map are determined as the second sample (positive sample).
- the computer device trains the neural network according to the second sample to obtain a trained cell detection model.
- the frame coordinates and category of the second sample may be regressed according to the feature map of the second sample, and when the loss value of the loss function reaches the convergence state, it is determined to obtain a trained cell detection model.
- the trained model will lead to the model instead The overall performance of the model is reduced.
- the neural network is gradually optimized using three intersection ratio thresholds, and the performance of the trained model can be greatly improved.
- the method further includes:
- the digital pathological image is a high-resolution digital image that is scanned and collected by a fully automatic microscope or optical magnification system, and then a computer is used to automatically perform high-precision multi-field seamless stitching and processing on the obtained image to obtain high-quality visual data for application Images in various fields of pathology.
- the size of the general digital pathology image is too large to directly use a computer for image analysis and processing, it is necessary to use a window to cut the digital pathology image into multiple small image blocks, and then re-analyze these images
- the block is preprocessed to obtain a target image with more obvious cell characteristics and less noise to improve the accuracy of image detection.
- the target image is input into a trained cell detection model to obtain abnormal cell detection results. Including the bounding box, category and probability of abnormal cells.
- the method further includes:
- a cell image mask is generated and mapped to the digital pathological image.
- the maximum between-class variance algorithm (ostu) is an efficient algorithm for binarizing an image, and the original image can be segmented into two parts, the foreground and the background, using a threshold.
- binarization is to set the gray value of the pixels on the image to 0 or 255, that is, the process of presenting the entire image with a clear black and white effect, which can greatly reduce the amount of data in the image, which can highlight The outline of the target.
- the expansion algorithm can merge all the background points in contact with the object into the object, expand the boundary to the outside, and can be used to fill the void in the object.
- the erosion algorithm can eliminate boundary points and shrink the boundary to the inside, which can be used to eliminate small and meaningless objects.
- the open operation refers to the first corrosion operation, then the expansion operation;
- the closed operation refers to the expansion operation first, then the corrosion operation.
- the digital pathology image may be binarized by the maximum between-class variance algorithm to obtain the binary image of the digital pathology image, and then the dilation algorithm and the erosion algorithm may be used to
- the open operation and close operation of the binary image can expand the internal area of the cell and eliminate isolated small dots, so that the cell area in the binary image is more obvious, the outline is clearer, and the cell image mask (cell After generating the cell image mask, the cell image mask can be mapped to the digital pathology image.
- a mask of abnormal cells may be output.
- the preprocessing the multiple image blocks to obtain multiple target images includes:
- gamma correction is performed on the normalized image to obtain the target image.
- normalization refers to RGB (Red Green Blue) normalization.
- RGB Red Green Blue
- gamma correction is to edit the gamma curve of the image to perform non-linear tone editing on the image, detect the dark part and light part in the image signal, and increase the ratio of the two. Large, thereby improving the image contrast effect.
- the sources of different images may be different, the images are often contaminated by random signals (also called noise) of different intensities during image acquisition, which will affect subsequent image analysis and processing. Unpredictable effects are produced. Therefore, it is necessary to perform RGB normalization and gamma correction on these images to reduce the impact of noise on the image and improve the accuracy of image detection.
- a target model can be trained using multiple sample images for each preset intersection ratio threshold, where the sample images include images containing abnormal cells and images that do not contain abnormal cells.
- Image use the free response receiver operating characteristic curve method to determine the correct rate of the target model, and determine the target intersection and ratio threshold according to the correct rate, wherein the target intersection and ratio threshold includes the first intersection and ratio threshold, A second intersection ratio threshold and a third intersection ratio threshold, where the third intersection ratio threshold is greater than the second intersection ratio threshold, and the second intersection ratio threshold is greater than the first intersection ratio threshold; Training the neural network to obtain the first parameter according to the first intersection ratio threshold and the plurality of sample images; and re-sampling the sample image according to the first parameter to obtain the first sample; According to the second intersection ratio threshold and the first sample, the neural network is trained to obtain the second parameter; according to the second parameter and the third intersection ratio threshold, the sample is The image is resampled to obtain a second sample; the neural network is trained according to the second sample to obtain a trained cell
- the accuracy rate of the trained cell detection model is relatively high, and by increasing the intersection ratio threshold, the sample image is resampled to ensure that the number of positive samples is sufficient to avoid overfitting, thereby Improve the accuracy of the cell detection model's prediction frame for abnormal cells, that is, improve the detection accuracy of abnormal cells.
- FIG. 2 is a functional block diagram of a preferred embodiment of a cell detection model training device disclosed in the present application.
- the cell detection model training device runs in a computer device.
- the cell detection model training device may include multiple functional modules composed of program code segments, and the program is a series of computer readable instruction codes.
- the program code of each program segment in the cell detection model training device can be stored in a memory and executed by at least one processor to perform part or all of the steps in the cell detection model training method described in FIG. 1, specifically Reference may be made to the related description in the method described in FIG. 1, which is not repeated here.
- the cell detection model training device can be divided into multiple functional modules according to the functions it performs.
- the functional modules may include: a training module 201, a determination module 202, and a sampling module 203.
- the module referred to in this application refers to a series of computer-readable instruction segments that can be executed by at least one processor and can complete fixed functions, and are stored in a memory.
- the training module 201 is used to train a target model using a plurality of sample images for each preset intersection ratio threshold, where the sample images include images containing abnormal cells and images not containing abnormal cells;
- the determining module 202 is configured to determine the correct rate of the target model using a free response receiver operating characteristic curve method, and determine a target intersection ratio threshold according to the correct rate, wherein the target intersection ratio threshold includes a first intersection The union ratio threshold, the second intersection ratio threshold, and the third intersection ratio threshold, the third intersection ratio threshold being greater than the second intersection ratio threshold, and the second intersection ratio threshold greater than the first intersection ratio threshold. And ratio threshold;
- the training module 201 is further configured to train the neural network according to the first intersection ratio threshold and the multiple sample images to obtain first parameters;
- the sampling module 203 is configured to resample the sample image according to the first parameter to obtain a first sample
- the training module 201 is further configured to train the neural network according to the second intersection ratio threshold and the first sample to obtain a second parameter;
- the sampling module 203 is further configured to resample the sample image according to the second parameter and the third intersection ratio threshold to obtain a second sample;
- the training module 201 is also used to train the neural network according to the second sample to obtain a trained cell detection model.
- the training module 201 trains the neural network according to the first intersection ratio threshold and the multiple sample images, and the specific method for obtaining the first parameter is:
- the feature information use a feature pyramid network to generate a first feature map
- the first feature map use a region candidate network to generate a candidate region
- regression is performed on the bounding box and category of the first region to obtain the first parameter.
- the sampling module 203 re-samples the sample image according to the first parameter, and the method for obtaining the first sample is specifically:
- the third region and the third feature map are determined as the first sample.
- the sampling module 203 re-samples the sample image according to the second parameter and the third intersection ratio threshold, and the method for obtaining the second sample is specifically:
- the fifth region and the fourth feature map are determined as the second sample.
- the cell detection model training device may further include:
- the acquisition module is used to acquire digital pathology images
- the cutting module is used to cut the digital pathological image to obtain multiple image blocks
- a preprocessing module configured to preprocess the multiple image blocks to obtain multiple target images
- the input module is used to input a plurality of the target images into the cell detection model to obtain abnormal cell detection results.
- the cell detection model training device may further include:
- the binarization module is used after the acquisition module acquires the digital pathology image, and the cutting module cuts the digital pathology image, and before obtaining multiple image blocks, the digital pathology image is processed by the maximum between-class variance algorithm Perform binarization processing to obtain binarized images;
- the arithmetic module is used to open the binarized image through the expansion algorithm and the corrosion algorithm, and perform the closed calculation on the binarized image through the expansion algorithm and the corrosion algorithm to obtain a more obvious binarized image of the cell area ;
- the generating module is used to generate a cell image mask and map it to the digital pathological image according to the more obvious binary image of the cell area.
- the preprocessing module performs preprocessing on the multiple image blocks, and the specific method for obtaining multiple target images is:
- gamma correction is performed on the normalized image to obtain the target image.
- a target model can be trained using multiple sample images for each preset intersection ratio threshold, where the sample images include images containing abnormal cells and images that do not contain abnormal cells. Image of abnormal cells; using the free response receiver operating characteristic curve method to determine the correct rate of the target model, and determine the target intersection ratio threshold according to the correct rate, wherein the target intersection ratio threshold includes the first intersection The ratio threshold, the second intersection ratio threshold, and the third intersection ratio threshold, the third intersection ratio threshold greater than the second intersection ratio threshold, and the second intersection ratio threshold greater than the first intersection ratio threshold Ratio threshold; according to the first intersection ratio threshold and the plurality of sample images, train the neural network to obtain the first parameter; according to the first parameter, resample the sample image to obtain the first Sample; according to the second intersection and union ratio threshold and the first sample, train the neural network to obtain a second parameter; according to the second parameter and the third intersection and union ratio threshold, The sample image is resampled to obtain a second sample; the neural network is trained according to the second intersection ratio threshold
- the accuracy rate of the trained cell detection model is relatively high, and by increasing the intersection ratio threshold, the sample image is resampled to ensure that the number of positive samples is sufficient to avoid overfitting, thereby Improve the accuracy of the cell detection model's prediction frame for abnormal cells, that is, improve the detection accuracy of abnormal cells.
- FIG. 3 is a schematic structural diagram of a computer device in a preferred embodiment of the method for training a cell detection model according to the present application.
- the computer device 3 includes a memory 31, at least one processor 32, computer readable instructions 33 stored in the memory 31 and executable on the at least one processor 32, and at least one communication bus 34.
- FIG. 3 is only an example of the computer device 3, and does not constitute a limitation on the computer device 3. It may include more or less components than those shown in the figure, or a combination. Some components, or different components, for example, the computer device 3 may also include input and output devices, network access devices, and so on.
- the computer device 3 also includes, but is not limited to, any electronic product that can interact with the user through a keyboard, a mouse, a remote control, a touch panel, or a voice control device, for example, a personal computer, a tablet computer, a smart phone, Personal digital assistants (Personal Digital Assistant, PDA), game consoles, interactive network television (Internet Protocol Television, IPTV), smart wearable devices, etc.
- a keyboard a mouse
- a remote control a touch panel
- a voice control device for example, a personal computer, a tablet computer, a smart phone, Personal digital assistants (Personal Digital Assistant, PDA), game consoles, interactive network television (Internet Protocol Television, IPTV), smart wearable devices, etc.
- PDA Personal Digital Assistant
- IPTV Internet Protocol Television
- smart wearable devices etc.
- the at least one processor 32 may be a central processing unit (Central Processing Unit, CPU), or other general-purpose processors, digital signal processors (Digital Signal Processors, DSPs), application specific integrated circuits (ASICs). ), Field-Programmable Gate Array (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
- the processor 32 can be a microprocessor or the processor 32 can also be any conventional processor, etc.
- the processor 32 is the control center of the computer device 3, and connects the entire computer device 3 by using various interfaces and lines. Parts.
- the memory 31 may be used to store the computer-readable instructions 33 and/or modules/units, and the processor 32 can run or execute the computer-readable instructions and/or modules/units stored in the memory 31, and
- the data stored in the memory 31 is called to realize various functions of the computer device 3.
- the memory 31 may mainly include a storage program area and a storage data area, where the storage program area may store an operating system, an application program required by at least one function (such as a sound playback function, an image playback function, etc.); the storage data area may Data (such as audio data) created according to the use of the computer device 3 and the like are stored.
- the memory 31 may include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a Secure Digital (SD) card, a flash memory card (Flash Card), At least one magnetic disk storage device, flash memory device, or other non-volatile solid state storage device.
- non-volatile memory such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a Secure Digital (SD) card, a flash memory card (Flash Card), At least one magnetic disk storage device, flash memory device, or other non-volatile solid state storage device.
- the memory 31 in the computer device 3 stores multiple instructions to implement a method for training a cell detection model, and the processor 32 can execute the multiple instructions to achieve:
- a target model is trained using multiple sample images, where the sample images include images containing abnormal cells and images not containing abnormal cells;
- the target intersection ratio threshold includes a first intersection ratio threshold, a second A merge ratio threshold and a third merge ratio threshold, where the third merge ratio threshold is greater than the second merge ratio threshold, and the second merge ratio threshold is greater than the first merge ratio threshold;
- the neural network is trained to obtain a trained cell detection model.
- a target model can be trained using multiple sample images for each preset intersection ratio threshold, where the sample images include images containing abnormal cells and images that do not contain abnormal cells.
- the image use the free response receiver operating characteristic curve method to determine the correct rate of the target model, and determine the target intersection and ratio threshold according to the correct rate, wherein the target intersection and ratio threshold includes the first intersection and ratio threshold , A second intersection ratio threshold and a third intersection ratio threshold, the third intersection ratio threshold is greater than the second intersection ratio threshold, and the second intersection ratio threshold is greater than the first intersection ratio threshold
- the neural network is trained to obtain the first parameter; according to the first parameter, the sample image is resampled to obtain the first sample ;
- the neural network is trained to obtain the second parameter; according to the second parameter and the third intersection ratio threshold, the The sample image is re-sampled to obtain a second sample; the neural network is trained according to the second sample to obtain
- the accuracy rate of the trained cell detection model is relatively high, and by increasing the intersection ratio threshold, the sample image is resampled to ensure that the number of positive samples is sufficient to avoid overfitting, thereby Improve the accuracy of the cell detection model's prediction frame for abnormal cells, that is, improve the detection accuracy of abnormal cells.
- the integrated module/unit of the computer device 3 is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a non-volatile readable storage medium.
- this application implements all or part of the processes in the above-mentioned embodiments and methods, and can also be completed by instructing relevant hardware through computer-readable instructions.
- the computer-readable instructions can be stored in a non-volatile memory. In the read storage medium, when the computer-readable instructions are executed by the processor, the steps of the foregoing method embodiments can be implemented.
- the computer-readable instruction includes computer-readable instruction code
- the computer-readable instruction code may be in the form of source code, object code, executable file, or some intermediate form.
- the non-volatile readable medium may include: any entity or device capable of carrying the computer readable instruction code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory).
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Computing Systems (AREA)
- Multimedia (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
一种细胞检测模型训练方法,所述方法包括:针对每个预设的交并比阈值,训练出目标模型;确定目标模型的正确率,并根据正确率确定第一交并比阈值、第二交并比阈值以及第三交并比阈值;根据第一交并比阈值以及多个样本图像,对神经网络进行训练,获得第一参数;根据第一参数,对样本图像进行重新采样,获得第一样本;根据第二交并比阈值以及第一样本,对神经网络进行训练,获得第二参数;根据第二参数以及第三交并比阈值,对样本图像进行重新采样,获得第二样本;根据第二样本,对神经网络进行训练,获得训练好的细胞检测模型。还提供一种细胞检测模型训练装置、计算机设备以及存储介质。能提高对异常细胞的检测精度。
Description
本申请要求于2019年08月15日提交中国专利局,申请号为201910755143.6发明名称为“细胞检测模型训练方法、装置、计算机设备及存储介质”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。
本申请涉及细胞检测技术领域,尤其涉及一种细胞检测模型训练方法、装置、计算机设备及存储介质。
宫颈癌是可以早发现并治愈的癌症,每年造成的死亡病例多数发生在宫颈癌普查率低的地区,因此,对宫颈癌细胞的检测工作非常重要。
目前,可以利用深度学习神经网络对异常细胞图像进行特征提取并训练出模型,以检测异常细胞。但在实践中发现,利用深度学习神经网络训练的模型通常是选取0.5作为交并比阈值,这样训练出来的模型在对异常细胞进行检测的时候生成的检测框不够精确。
因此,如何提高对异常细胞的检测精度是一个亟待解决的技术问题。
发明内容
鉴于以上内容,有必要提供一种细胞检测模型训练方法、装置、计算机设备及存储介质,能够提高对异常细胞的检测精度。
本申请的第一方面提供一种细胞检测模型训练方法,所述方法包括:
针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;
使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;
根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一 参数;
根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;
根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;
根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;
根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
本申请的第二方面提供一种细胞检测模型训练装置,所述装置包括:
训练模块,用于针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;
确定模块,用于使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;
所述训练模块,还用于根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;
采样模块,用于根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;
所述训练模块,还用于根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;
所述采样模块,还用于根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;
所述训练模块,还用于根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
本申请的第三方面提供一种计算机设备,所述计算机设备包括处理器和存储器,所述处理器用于执行所述存储器中存储的计算机可读指令时实现所述的细胞检测模型训练方法。
本申请的第四方面提供一种非易失性可读存储介质,所述非易失性可读存储介质上存储有计算机可读指令,所述计算机可读指令被处理器执行时实现所述的细胞检测模型训练方法。
由以上技术方案可知,本申请通过选取目标交并比阈值,确保训练的细胞检测模型的正确率比较高,并通过提高交并比阈值对所述样本图像进行重新取样,确保正样本数 量足够,避免过度拟合,从而提高细胞检测模型对异常细胞的预测框的精度,即提高了对异常细胞的检测精度。
图1是本申请公开的一种细胞检测模型训练方法的较佳实施例的流程图。
图2是本申请公开的一种细胞检测模型训练装置的较佳实施例的功能模块图。
图3是本申请实现细胞检测模型训练方法的较佳实施例的计算机设备的结构示意图。
本申请实施例的细胞检测模型训练方法应用在计算机设备中,也可以应用在计算机设备和通过网络与所述计算机设备进行连接的服务器所构成的硬件环境中,由服务器和计算机设备共同执行。网络包括但不限于:广域网、城域网或局域网。
请参见图1,图1是本申请公开的一种细胞检测模型训练方法的较佳实施例的流程图。其中,根据不同的需求,该流程图中步骤的顺序可以改变,某些步骤可以省略。
S11、计算机设备针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像。
其中,所述交并比(Intersection over union,IOU)在目标检测类型的模型中指产生的候选框与标注框的交叠率。在设定训练的正负样本(用于训练分类以及对正样本进行坐标回归)时,主要是根据交并比阈值来确定正样本和负样本,比如,选取交并比阈值为0.5,即IOU>0.5的作为正样本。
其中,所述样本图像上标注有异常细胞的边界框(坐标信息)以及类别。
其中,所述目标模型是指使用单个交并比阈值训练出的用来检测异常细胞的模型。
本申请实施例中,可以预先设置多个交并比阈值(比如:0.3,0.4,0.5,0.6,0.7,08,0.9),然后训练出每个交并比阈值对应的目标模型。训练的过程就是对神经网络的训练,神经网络会根据输入的样本图像生成多个候选区域,然后将候选区域的边界框跟标注的异常细胞的边界框(标注框)做比较,将交并比大于交并比阈值的候选区域确定为正样本,对正样本的边界框坐标以及类别进行回归,当损失函数的损失值达到收敛状态时,确定目标模型被训练好。一般来说,交并比阈值越高,正样本的质量越高,训练出来的模型性能越好,但是交并比阈值过高会导致正样本数量过少,从而使得训练出现过度拟合的情况,训练出的模型性能反而下降。在多个目标模型被训练好后,可以对每个目标模型进行评估,以确定性能比较好的目标模型对应的交并比阈值。
S12、计算机设备使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值。
其中,所述自由响应接收器操作特性曲线(Free-response Receiver Operating Characteristic Curves,FROC),是接收器操作特性曲线(Receiver Operating Characteristic Curves,ROC)的一个变种,通过计算ROC曲线的AUC(Area under curve,曲线下面积)值可以确定一个模型的正确率,但是ROC方法不能解决对一幅图像上多个异常进行评价,而使用FROC方法可以对一幅图像上多个异常进行评价。
其中,所述目标交并比阈值用于对样本进行优化。
本申请实施例中,可以通过FROC方法确定根据不同交并比阈值训练出来的不同的目标模型(细胞检测模型)的正确率。将正确率最高的三个目标模型对应的交并比阈值确定为目标交并比阈值:第一交并比阈值、第二交并比阈值以及第三交并比阈值,其中,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值。
S13、计算机设备根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数。
其中,所述神经网络包括但不限于:残差卷积网络(residual network,ResNet)、特征金字塔网络(Feature Pyramid Networks,FPN)以及区域候选网络(Region Proposal Network,RPN)。
本申请实施例中,可以根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数,因为第一交并比阈值训练出的神经网络的探测性能比较好,可以确保后续训练出来的模型的正确率比较高。
具体的,所述根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数包括:
针对每个所述样本图像,使用残差卷积网络对所述样本图像进行特征提取,获得特征信息;
根据所述特征信息,使用特征金字塔网络生成第一特征图;
根据所述第一特征图,使用区域候选网络生成候选区域;
根据所述第一交并比阈值,对所述候选区域进行筛选,获得第一区域;
将所述第一特征图以及所述第一区域输入至感兴趣区域池化层,获得所述第一区域对应的第二特征图;
根据所述第二特征图,对所述第一区域的边界框以及类别进行回归,获得第一参数。
在该可选的实施方式中,可以使用残差卷积网络对图像进行特征提取,获取图像的特征信息,然后将图像的特征信息输入至特征金字塔网络,获得图像的特征图(第一特征图),区域候选网络对特征图进行处理,生成多个候选区域。这时可以根据所述第一交并比阈值对所述候选区域进行筛选,获得第一区域,即正样本。可以将图像的特征图以及第一区域输入至感兴趣区域(region of interest,ROI)池化层,获得第一区域对应的特征图(第二特征图),其中,可以使用ROI Align(感兴趣区域对齐),实现特征图尺寸的统一化。然后,可以根据第一区域的特征图,对第一区域的边界框(坐标)以及类别进行回归,获得神经网络的参数(第一参数)。所述第一参数是指根据第一交并比阈值筛选出来的候选框,进行训练后的整个神经网络的各个参数。
S14、计算机设备根据所述第一参数,对所述样本图像进行重新采样,获得第一样本。
在本申请实施例中,通过训练出来的第一参数对所述样本图像进行重新采样,即使用训练过的神经网络对之前候选区域网络对特征图进行处理,生成的候选框(所述第一区域)进行优化,获得新的一批边界框,优化后的边界框的交并比更高,确保提高交并比阈值后,正样本的数量足够用于训练,可以提高预测框的精度并防止过度拟合。
具体的,所述根据所述第一参数,对所述样本图像进行重新采样,获得第一样本包括:
根据所述第一参数以及所述第一区域,生成第二区域;
根据所述第二交并比阈值,对所述第二区域进行筛选,获得第三区域;
将所述第一特征图以及所述第三区域输入至感兴趣区域池化层,获得所述第三区域对应的第三特征图;
将所述第三区域以及所述第三特征图确定为第一样本。
在该可选的实施方式中,神经网络可以根据所述第一参数对所述第一区域进行优化,生成第二区域,根据所述第二交并比阈值对所述第二区域进行筛选,将交并比大于所述第二交并比阈值的所述第二区域确定为第三区域,可以将图像的特征图以及第三区域输入至感兴趣区域池化层,获得第三区域对应的特征图(第三特征图),其中,可以使用ROI Align,实现特征图尺寸的统一化。将所述第三区域以及所述第三特征图确定为第一样本(正样本)。
S15、计算机设备根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数。
本申请实施例中,可以根据第一样本中的特征图,对第一样本的边界框(坐标)以及类别进行回归,获得神经网络的参数(第二参数),所述第二参数是指对神经网络进行再训练,得到的各个参数,是对所述第一参数进行优化后得到的参数。
S16、计算机设备根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本。
本申请实施例中,可以通过训练出来的第二参数对所述样本图像进行重新采样,即使用训练过的神经网络对之前产生的候选框(所述第三区域)进行优化,获得新的一批边界框,优化后得到的边界框的交并比更高,确保提高交并比阈值后,正样本的数量足够用于训练,可以继续提高预测框的精度并防止过度拟合。
具体的,所述根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本包括:
根据所述第二参数以及所述第三区域,生成第四区域;
根据所述第三交并比阈值,对所述第四区域进行筛选,获得第五区域;
将所述第一特征图以及所述第五区域输入至感兴趣区域池化层,获得所述第五区域对应的第四特征图;
将所述第五区域以及所述第四特征图确定为第二样本。
在该可选的实施方式中,神经网络可以根据所述第二参数对所述第三区域进行优化,生成第四区域,根据所述第三交并比阈值对所述第四区域进行筛选,将交并比大于所述第三交并比阈值的所述第四区域确定为第五区域,可以将图像的特征图以及第五区域输入至感兴趣区域池化层,获得第五区域对应的特征图(第四特征图),其中,可以使用ROI Align,实现特征图尺寸的统一化。将所述第五区域以及所述第四特征图确定为第二样本(正样本)。
S17、计算机设备根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
本申请实施中,可以根据所述第二样本的特征图,对所述第二样本的边框坐标以及类别进行回归,当损失函数的损失值达到收敛状态时,确定获得训练好的细胞检测模型。通过实验,如果继续增加更高的第四交并比阈值对样本图像进行重新采样(对上一个交并比阈值训练出的区域候选网络输出的候选框进行优化)训练出来的模型反而会导致模型的整体性能下降,综合考虑,采用三个交并比阈值对神经网络进行逐步优化,训练出的模型的性能可以得到较大的改善。
作为一种可选的实施方式,所述方法还包括:
获取数字病理图像;
对所述数字病理图像进行切割,获得多个图像块;
对所述多个图像块进行预处理,获得多个目标图像;
将多个所述目标图像输入至所述细胞检测模型中,获得异常细胞检测结果。
其中,所述数字病理图像是通过全自动显微镜或光学放大系统扫描采集得到高分辨数字图像,再应用计算机对得到的图像自动进行高精度多视野无缝隙拼接和处理,获得优质的可视化数据以应用于病理学的各个领域的图像。
在该可选的实施方式中,因为一般的数字病理图像的大小太大,无法直接用计算机做图像分析处理,需要使用划窗将数字病理图像切割成多个小图像块,然后再对这些图像块进行预处理,得到细胞特征更明显,噪声更少的目标图像,以提高对图像检测的准确度,将目标图像输入至训练好的细胞检测模型中,获得异常细胞检测结果,所述检测结果包括异常细胞的边界框、类别以及概率。
作为一种可选的实施方式,所述获取数字病理图像之后,以及所述对所述数字病理图像进行切割,获得多个图像块之前,所述方法还包括:
通过最大类间方差算法对所述数字病理图像进行二值化处理,获得二值化图像;
通过膨胀算法和腐蚀算法对所述二值化图像进行开运算,以及通过膨胀算法和腐蚀算法对所述二值化图像进行闭运算,获得细胞区域更明显的二值化图像;
根据所述细胞区域更明显的二值化图像,生成细胞图像掩膜并映射至所述数字病理图像。
其中,所述最大类间方差算法(ostu)是一种对图像进行二值化的高效算法,可以利用阈值将原图像分割成前景,背景两部分。其中,二值化是将图像上的像素点的灰度值设置为0或255,也就是将整个图像呈现出明显的黑白效果的过程,可以使图像中数据量大为减少,从而能凸显出目标的轮廓。
其中,膨胀算法可以将与物体接触的所有背景点合并到该物体中,使边界向外部扩张,可以用来填补物体中的空洞。其中,腐蚀算法可以消除边界点,使边界向内部收缩,可以用来消除小且无意义的物体。
其中,开运算是指先腐蚀运算,再膨胀运算;闭运算是指先膨胀运算,再腐蚀运算。
在该可选的实施方式中,可以通过最大类间方差算法对所述数字病理图像进行二值化处理,获得所述数字病理图像的二值化图像,然后通过膨胀算法和腐蚀算法对所述二值化图像进行开运算以及闭运算,可以扩充细胞内部区域,并消除孤立的小点,使得所述二值化图像中的细胞区域更加明显,轮廓更加清晰,方便生成细胞图像掩膜(细胞的轮廓),在生成细胞图像掩膜后,可以将细胞图像掩膜映射至所述数字病理图像。可选的,可以在对所述数字病理图像进行异常细胞检测时,输出异常细胞的掩膜。
具体的,所述对所述多个图像块进行预处理,获得多个目标图像包括:
针对每个所述图像块,通过归一化算法对所述图像块的像素进行归一化,获得归一化图 像;
根据预设的伽马阈值,对所述归一化图像进行伽马校正,获得目标图像。
其中,归一化是指RGB(Red Green Blue)归一化,通过对图像的RGB色彩空间进行归一化处理,可以消除一部分光照对图像的影响。
其中,伽马校正(gamma correction)是对图像的伽马曲线进行编辑,以对图像进行非线性色调编辑的方法,检出图像信号中的深色部分和浅色部分,并使两者比例增大,从而提高图像对比度效果。
在该可选的实施方式中,因为不同的图像的来源可能不同,在图像的采集中,图像往往会被不同强度的随机信号(也称噪声)所污染,这会对后续的图像分析和处理产生不可预知的影响,因此,需要对这些图像进行RGB归一化以及伽马校正,以减少噪声对图像的影响,可以提高图像检测的准确性。
在图1所描述的方法流程中,可以针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。可见,通过选取目标交并比阈值,确保训练的细胞检测模型的正确率比较高,并通过提高交并比阈值对所述样本图像进行重新取样,确保正样本数量足够,避免过度拟合,从而提高细胞检测模型对异常细胞的预测框的精度,即提高了对异常细胞的检测精度。
以上所述,仅是本申请的具体实施方式,但本申请的保护范围并不局限于此,对于本领域的普通技术人员来说,在不脱离本申请创造构思的前提下,还可以做出改进,但这些均属于本申请的保护范围。
请参见图2,图2是本申请公开的一种细胞检测模型训练装置的较佳实施例的功能模块图。
在一些实施例中,所述细胞检测模型训练装置运行于计算机设备中。所述细胞检测模型训练装置可以包括多个由程序代码段所组成的功能模块,所述程序是一系列的计算机可读指 令代码。所述细胞检测模型训练装置中的各个程序段的程序代码可以存储于存储器中,并由至少一个处理器所执行,以执行图1所描述的细胞检测模型训练方法中的部分或全部步骤,具体可以参照图1所述方法中的相关描述,在此不再赘述。
本实施例中,所述细胞检测模型训练装置根据其所执行的功能,可以被划分为多个功能模块。所述功能模块可以包括:训练模块201、确定模块202及采样模块203。本申请所称的模块是指一种能够被至少一个处理器所执行并且能够完成固定功能的一系列计算机可读指令段,其存储在存储器中。
训练模块201,用于针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;
确定模块202,用于使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;
所述训练模块201,还用于根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;
采样模块203,用于根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;
所述训练模块201,还用于根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;
所述采样模块203,还用于根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;
所述训练模块201,还用于根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
作为一种可选的实施方式,所述训练模块201根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数的方式具体为:
针对每个所述样本图像,使用残差卷积网络对所述样本图像进行特征提取,获得特征信息;
根据所述特征信息,使用特征金字塔网络生成第一特征图;
根据所述第一特征图,使用区域候选网络生成候选区域;
根据所述第一交并比阈值,对所述候选区域进行筛选,获得第一区域;
将所述第一特征图以及所述第一区域输入至感兴趣区域池化层,获得所述第一区域对应的第二特征图;
根据所述第二特征图,对所述第一区域的边界框以及类别进行回归,获得第一参数。
作为一种可选的实施方式,所述采样模块203根据所述第一参数,对所述样本图像进行重新采样,获得第一样本的方式具体为:
根据所述第一参数以及所述第一区域,生成第二区域;
根据所述第二交并比阈值,对所述第二区域进行筛选,获得第三区域;
将所述第一特征图以及所述第三区域输入至感兴趣区域池化层,获得所述第三区域对应的第三特征图;
将所述第三区域以及所述第三特征图确定为第一样本。
作为一种可选的实施方式,所述采样模块203根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本的方式具体为:
根据所述第二参数以及所述第三区域,生成第四区域;
根据所述第三交并比阈值,对所述第四区域进行筛选,获得第五区域;
将所述第一特征图以及所述第五区域输入至感兴趣区域池化层,获得所述第五区域对应的第四特征图;
将所述第五区域以及所述第四特征图确定为第二样本。
作为一种可选的实施方式,所述细胞检测模型训练装置还可以包括:
获取模块,用于获取数字病理图像;
切割模块,用于对所述数字病理图像进行切割,获得多个图像块;
预处理模块,用于对所述多个图像块进行预处理,获得多个目标图像;
输入模块,用于将多个所述目标图像输入至所述细胞检测模型中,获得异常细胞检测结果。
作为一种可选的实施方式,所述细胞检测模型训练装置还可以包括:
二值化模块,用于所述获取模块获取数字病理图像之后,以及所述切割模块对所述数字病理图像进行切割,获得多个图像块之前,通过最大类间方差算法对所述数字病理图像进行二值化处理,获得二值化图像;
运算模块,用于通过膨胀算法和腐蚀算法对所述二值化图像进行开运算,以及通过膨胀算法和腐蚀算法对所述二值化图像进行闭运算,获得细胞区域更明显的二值化图像;
生成模块,用于根据所述细胞区域更明显的二值化图像,生成细胞图像掩膜并映射至所述数字病理图像。
作为一种可选的实施方式,所述预处理模块对所述多个图像块进行预处理,获得多个目标图像的方式具体为:
针对每个所述图像块,通过归一化算法对所述图像块的像素进行归一化,获得归一化图像;
根据预设的伽马阈值,对所述归一化图像进行伽马校正,获得目标图像。
在图2所描述的细胞检测模型训练装置中,可以针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。可见,通过选取目标交并比阈值,确保训练的细胞检测模型的正确率比较高,并通过提高交并比阈值对所述样本图像进行重新取样,确保正样本数量足够,避免过度拟合,从而提高细胞检测模型对异常细胞的预测框的精度,即提高了对异常细胞的检测精度。
如图3所示,图3是本申请实现细胞检测模型训练方法的较佳实施例的计算机设备的结构示意图。所述计算机设备3包括存储器31、至少一个处理器32、存储在所述存储器31中并可在所述至少一个处理器32上运行的计算机可读指令33及至少一条通讯总线34。
本领域技术人员可以理解,图3所示的示意图仅仅是所述计算机设备3的示例,并不构成对所述计算机设备3的限定,可以包括比图示更多或更少的部件,或者组合某些部件,或者不同的部件,例如所述计算机设备3还可以包括输入输出设备、网络接入设备等。
所述计算机设备3还包括但不限于任何一种可与用户通过键盘、鼠标、遥控器、触摸板或声控设备等方式进行人机交互的电子产品,例如,个人计算机、平板电脑、智能手机、个人数字助理(Personal Digital Assistant,PDA)、游戏机、交互式网络电视(Internet Protocol Television,IPTV)、智能式穿戴式设备等。
所述至少一个处理器32可以是中央处理单元(Central Processing Unit,CPU),还可以是其他通用处理器、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、分立硬件组件等。该处理器32可以是微处理器或者该处理器32也可以是任何常规的处理器等,所述处理器32 是所述计算机设备3的控制中心,利用各种接口和线路连接整个计算机设备3的各个部分。
所述存储器31可用于存储所述计算机可读指令33和/或模块/单元,所述处理器32通过运行或执行存储在所述存储器31内的计算机可读指令和/或模块/单元,以及调用存储在存储器31内的数据,实现所述计算机设备3的各种功能。所述存储器31可主要包括存储程序区和存储数据区,其中,存储程序区可存储操作系统、至少一个功能所需的应用程序(比如声音播放功能、图像播放功能等)等;存储数据区可存储根据计算机设备3的使用所创建的数据(比如音频数据)等。此外,存储器31可以包括非易失性存储器,例如硬盘、内存、插接式硬盘,智能存储卡(Smart Media Card,SMC),安全数字(Secure Digital,SD)卡,闪存卡(Flash Card)、至少一个磁盘存储器件、闪存器件、或其他非易失性固态存储器件。
结合图1,所述计算机设备3中的所述存储器31存储多个指令以实现一种细胞检测模型训练方法,所述处理器32可执行所述多个指令从而实现:
针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;
使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;
根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;
根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;
根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;
根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;
根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
具体地,所述处理器32对上述指令的具体实现方法可参考图1对应实施例中相关步骤的描述,在此不赘述。
在图3所描述的计算机设备3中,可以针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练, 获得第一参数;根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。可见,通过选取目标交并比阈值,确保训练的细胞检测模型的正确率比较高,并通过提高交并比阈值对所述样本图像进行重新取样,确保正样本数量足够,避免过度拟合,从而提高细胞检测模型对异常细胞的预测框的精度,即提高了对异常细胞的检测精度。
所述计算机设备3集成的模块/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,可以存储在一个非易失性可读取存储介质中。基于这样的理解,本申请实现上述实施例方法中的全部或部分流程,也可以通过计算机可读指令来指令相关的硬件来完成,所述的计算机可读指令可存储于一非易失性可读存储介质中,该计算机可读指令在被处理器执行时,可实现上述各个方法实施例的步骤。其中,所述计算机可读指令包括计算机可读指令代码,所述计算机可读指令代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述非易失性可读介质可以包括:能够携带所述计算机可读指令代码的任何实体或装置、记录介质、U盘、移动硬盘、磁碟、光盘、计算机存储器、只读存储器(ROM,Read-Only Memory)。
在本申请所提供的几个实施例中,应该理解到,所揭露的系统,装置和方法,可以通过其它的方式实现。例如,以上所描述的装置实施例仅仅是示意性的,例如,所述模块的划分,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。
最后应说明的是,以上实施例仅用以说明本申请的技术方案而非限制,尽管参照较佳实施例对本申请进行了详细说明,本领域的普通技术人员应当理解,可以对本申请的技术方案进行修改或等同替换,而不脱离本申请技术方案的精神和范围。
Claims (20)
- 一种细胞检测模型训练方法,其特征在于,所述方法包括:针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
- 根据权利要求1所述的方法,其特征在于,所述根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数包括:针对每个所述样本图像,使用残差卷积网络对所述样本图像进行特征提取,获得特征信息;根据所述特征信息,使用特征金字塔网络生成第一特征图;根据所述第一特征图,使用区域候选网络生成候选区域;根据所述第一交并比阈值,对所述候选区域进行筛选,获得第一区域;将所述第一特征图以及所述第一区域输入至感兴趣区域池化层,获得所述第一区域对应的第二特征图;根据所述第二特征图,对所述第一区域的边界框以及类别进行回归,获得第一参数。
- 根据权利要求2所述的方法,其特征在于,所述根据所述第一参数,对所述样本图像进行重新采样,获得第一样本包括:利用所述第一参数对所述第一区域进行优化,生成第二区域;根据所述第二交并比阈值,对所述第二区域进行筛选,获得第三区域;将所述第一特征图以及所述第三区域输入至感兴趣区域池化层,获得所述第三区域对应 的第三特征图;将所述第三区域以及所述第三特征图确定为第一样本。
- 根据权利要求3所述的方法,其特征在于,所述根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本包括:根据所述第二参数以及所述第三区域,生成第四区域;根据所述第三交并比阈值,对所述第四区域进行筛选,获得第五区域;将所述第一特征图以及所述第五区域输入至感兴趣区域池化层,获得所述第五区域对应的第四特征图;将所述第五区域以及所述第四特征图确定为第二样本。
- 根据权利要求1至4中任一项所述的方法,其特征在于,所述方法还包括:获取数字病理图像;对所述数字病理图像进行切割,获得多个图像块;对所述多个图像块进行预处理,获得多个目标图像;将多个所述目标图像输入至所述细胞检测模型中,获得异常细胞检测结果。
- 根据权利要求5所述的方法,其特征在于,所述获取数字病理图像之后,以及所述对所述数字病理图像进行切割,获得多个图像块之前,所述方法还包括:通过最大类间方差算法对所述数字病理图像进行二值化处理,获得二值化图像;通过膨胀算法和腐蚀算法对所述二值化图像进行开运算,以及通过膨胀算法和腐蚀算法对所述二值化图像进行闭运算,获得细胞区域更明显的二值化图像;根据所述细胞区域更明显的二值化图像,生成细胞图像掩膜并映射至所述数字病理图像。
- 根据权利要求5所述的方法,其特征在于,所述对所述多个图像块进行预处理,获得多个目标图像包括:针对每个所述图像块,通过归一化算法对所述图像块的像素进行归一化,获得归一化图像;根据预设的伽马阈值,对所述归一化图像进行伽马校正,获得目标图像。
- 一种细胞检测模型训练装置,其特征在于,所述细胞检测模型训练装置包括:训练模块,用于针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;确定模块,用于使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第 二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;所述训练模块,还用于根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;采样模块,用于根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;所述训练模块,还用于根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;所述采样模块,还用于根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;所述训练模块,还用于根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
- 一种计算机设备,其特征在于,所述计算机设备包括处理器和存储器,所述处理器用于执行存储器中存储的至少一个计算机可读指令以实现以下步骤:针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
- 根据权利要求9所述的计算机设备,其特征在于,在所述根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数时,所述处理器执行所述至少一个计算机可读指令以实现以下步骤:针对每个所述样本图像,使用残差卷积网络对所述样本图像进行特征提取,获得特征信息;根据所述特征信息,使用特征金字塔网络生成第一特征图;根据所述第一特征图,使用区域候选网络生成候选区域;根据所述第一交并比阈值,对所述候选区域进行筛选,获得第一区域;将所述第一特征图以及所述第一区域输入至感兴趣区域池化层,获得所述第一区域对应的第二特征图;根据所述第二特征图,对所述第一区域的边界框以及类别进行回归,获得第一参数。
- 根据权利要求10所述的计算机设备,其特征在于,在所述根据所述第一参数,对所述样本图像进行重新采样,获得第一样本时,所述处理器执行所述至少一个计算机可读指令以实现以下步骤:利用所述第一参数对所述第一区域进行优化,生成第二区域;根据所述第二交并比阈值,对所述第二区域进行筛选,获得第三区域;将所述第一特征图以及所述第三区域输入至感兴趣区域池化层,获得所述第三区域对应的第三特征图;将所述第三区域以及所述第三特征图确定为第一样本。
- 根据权利要求11所述的计算机设备,其特征在于,在所述根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本时,所述处理器执行所述至少一个计算机可读指令以实现以下步骤:根据所述第二参数以及所述第三区域,生成第四区域;根据所述第三交并比阈值,对所述第四区域进行筛选,获得第五区域;将所述第一特征图以及所述第五区域输入至感兴趣区域池化层,获得所述第五区域对应的第四特征图;将所述第五区域以及所述第四特征图确定为第二样本。
- 根据权利要求9至12中任一项所述的计算机设备,其特征在于,所述处理器执行所述至少一个计算机可读指令以实现以下步骤:获取数字病理图像;对所述数字病理图像进行切割,获得多个图像块;对所述多个图像块进行预处理,获得多个目标图像;将多个所述目标图像输入至所述细胞检测模型中,获得异常细胞检测结果。
- 根据权利要求13所述的计算机设备,其特征在于,在所述获取数字病理图像之后,以及所述对所述数字病理图像进行切割,获得多个图像块之前,所述处理器执行所述至少一个计算机可读指令以实现以下步骤:通过最大类间方差算法对所述数字病理图像进行二值化处理,获得二值化图像;通过膨胀算法和腐蚀算法对所述二值化图像进行开运算,以及通过膨胀算法和腐蚀算法对所述二值化图像进行闭运算,获得细胞区域更明显的二值化图像;根据所述细胞区域更明显的二值化图像,生成细胞图像掩膜并映射至所述数字病理图像。
- 一种非易失性可读存储介质,其特征在于,所述非易失性可读存储介质存储有至少一个计算机可读指令,所述至少一个计算机可读指令被处理器执行时实现以下步骤:针对每个预设的交并比阈值,使用多个样本图像训练出目标模型,其中,所述样本图像包括含有异常细胞的图像以及不含有异常细胞的图像;使用自由响应接收器操作特性曲线方法确定所述目标模型的正确率,并根据所述正确率确定目标交并比阈值,其中,所述目标交并比阈值包括第一交并比阈值、第二交并比阈值以及第三交并比阈值,所述第三交并比阈值大于所述第二交并比阈值,所述第二交并比阈值大于所述第一交并比阈值;根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数;根据所述第一参数,对所述样本图像进行重新采样,获得第一样本;根据所述第二交并比阈值以及所述第一样本,对所述神经网络进行训练,获得第二参数;根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本;根据所述第二样本,对所述神经网络进行训练,获得训练好的细胞检测模型。
- 根据权利要求15所述的存储介质,其特征在于,在所述根据所述第一交并比阈值以及所述多个样本图像,对神经网络进行训练,获得第一参数时,所述至少一个计算机可读指令被处理器执行以实现以下步骤:针对每个所述样本图像,使用残差卷积网络对所述样本图像进行特征提取,获得特征信息;根据所述特征信息,使用特征金字塔网络生成第一特征图;根据所述第一特征图,使用区域候选网络生成候选区域;根据所述第一交并比阈值,对所述候选区域进行筛选,获得第一区域;将所述第一特征图以及所述第一区域输入至感兴趣区域池化层,获得所述第一区域对应的第二特征图;根据所述第二特征图,对所述第一区域的边界框以及类别进行回归,获得第一参数。
- 根据权利要求16所述的存储介质,其特征在于,在所述根据所述第一参数,对所述样本图像进行重新采样,获得第一样本时,所述至少一个计算机可读指令被处理器执行以 实现以下步骤:利用所述第一参数对所述第一区域进行优化,生成第二区域;根据所述第二交并比阈值,对所述第二区域进行筛选,获得第三区域;将所述第一特征图以及所述第三区域输入至感兴趣区域池化层,获得所述第三区域对应的第三特征图;将所述第三区域以及所述第三特征图确定为第一样本。
- 根据权利要求17所述的存储介质,其特征在于,在所述根据所述第二参数以及所述第三交并比阈值,对所述样本图像进行重新采样,获得第二样本时,所述至少一个计算机可读指令被处理器执行以实现以下步骤:根据所述第二参数以及所述第三区域,生成第四区域;根据所述第三交并比阈值,对所述第四区域进行筛选,获得第五区域;将所述第一特征图以及所述第五区域输入至感兴趣区域池化层,获得所述第五区域对应的第四特征图;将所述第五区域以及所述第四特征图确定为第二样本。
- 根据权利要求15至18中任一项所述的存储介质,其特征在于,所述至少一个计算机可读指令被处理器执行以实现以下步骤:获取数字病理图像;对所述数字病理图像进行切割,获得多个图像块;对所述多个图像块进行预处理,获得多个目标图像;将多个所述目标图像输入至所述细胞检测模型中,获得异常细胞检测结果。
- 根据权利要求19所述的存储介质,其特征在于,在所述获取数字病理图像之后,以及所述对所述数字病理图像进行切割,获得多个图像块之前,所述至少一个计算机可读指令被处理器执行以实现以下步骤:通过最大类间方差算法对所述数字病理图像进行二值化处理,获得二值化图像;通过膨胀算法和腐蚀算法对所述二值化图像进行开运算,以及通过膨胀算法和腐蚀算法对所述二值化图像进行闭运算,获得细胞区域更明显的二值化图像;根据所述细胞区域更明显的二值化图像,生成细胞图像掩膜并映射至所述数字病理图像。
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910755143.6A CN110705583B (zh) | 2019-08-15 | 2019-08-15 | 细胞检测模型训练方法、装置、计算机设备及存储介质 |
CN201910755143.6 | 2019-08-15 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021027135A1 true WO2021027135A1 (zh) | 2021-02-18 |
Family
ID=69193240
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2019/117233 WO2021027135A1 (zh) | 2019-08-15 | 2019-11-11 | 细胞检测模型训练方法、装置、计算机设备及存储介质 |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN110705583B (zh) |
WO (1) | WO2021027135A1 (zh) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113052217A (zh) * | 2021-03-15 | 2021-06-29 | 上海云从汇临人工智能科技有限公司 | 预测结果标识及其模型训练方法、装置及计算机存储介质 |
CN113128408A (zh) * | 2021-04-21 | 2021-07-16 | 深圳大学 | 一种物品检测方法、装置、终端及存储介质 |
CN113780270A (zh) * | 2021-03-23 | 2021-12-10 | 京东鲲鹏(江苏)科技有限公司 | 目标检测方法和装置 |
CN113837194A (zh) * | 2021-09-23 | 2021-12-24 | 北京百度网讯科技有限公司 | 图像处理方法、图像处理装置、电子设备以及存储介质 |
CN116434066A (zh) * | 2023-04-17 | 2023-07-14 | 安徽农业大学 | 基于深度学习的大豆豆荚考种方法、系统及装置 |
CN117036227A (zh) * | 2022-09-21 | 2023-11-10 | 腾讯科技(深圳)有限公司 | 数据处理方法、装置、电子设备、介质及程序产品 |
CN118658159A (zh) * | 2024-08-19 | 2024-09-17 | 中国科学院心理研究所 | 间接免疫荧光图像的处理方法及系统 |
Families Citing this family (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111275191B (zh) * | 2020-02-26 | 2023-11-28 | 上海商汤智能科技有限公司 | 检测细胞的方法及装置、电子设备和存储介质 |
CN111401387B (zh) * | 2020-03-10 | 2023-09-19 | 腾讯科技(深圳)有限公司 | 异常样本构建方法、装置、计算机设备和存储介质 |
CN111523452B (zh) * | 2020-04-22 | 2023-08-25 | 北京百度网讯科技有限公司 | 用于检测图像中人体位置的方法和装置 |
CN111681738B (zh) * | 2020-06-09 | 2024-05-07 | 平安科技(深圳)有限公司 | 基于病理切片扫描和分析一体化方法、装置、设备及介质 |
CN111524137B (zh) * | 2020-06-19 | 2024-04-05 | 平安科技(深圳)有限公司 | 基于图像识别的细胞识别计数方法、装置和计算机设备 |
CN112184733B (zh) * | 2020-09-30 | 2024-07-23 | 平安科技(深圳)有限公司 | 一种宫颈异常细胞检测装置及方法 |
CN112581438B (zh) * | 2020-12-10 | 2022-11-08 | 腾讯医疗健康(深圳)有限公司 | 切片图像识别方法、装置和存储介质及电子设备 |
CN113160159A (zh) * | 2021-04-13 | 2021-07-23 | 王永彬 | Hpv检测及病理分析系统和方法 |
CN113295692B (zh) * | 2021-05-25 | 2024-06-21 | 郑州中普医疗器械有限公司 | 基于细胞核dna和tbs双分析法的细胞分析方法、计算机设备、存储介质 |
CN114359899B (zh) * | 2021-12-09 | 2022-09-20 | 首都医科大学附属北京天坛医院 | 细胞共培养模型及细胞模型构建方法、计算机设备及存储介质 |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016665A (zh) * | 2017-02-16 | 2017-08-04 | 浙江大学 | 一种基于深度卷积神经网络的ct肺结节检测方法 |
CN109800778A (zh) * | 2018-12-03 | 2019-05-24 | 浙江工业大学 | 一种基于难分样本挖掘的Faster RCNN目标检测方法 |
CN109919230A (zh) * | 2019-03-10 | 2019-06-21 | 西安电子科技大学 | 基于循环特征金字塔的医学图像肺结节检测方法 |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108399362B (zh) * | 2018-01-24 | 2022-01-07 | 中山大学 | 一种快速行人检测方法及装置 |
CN109613006A (zh) * | 2018-12-22 | 2019-04-12 | 中原工学院 | 一种基于端到端神经网络的织物疵点检测方法 |
-
2019
- 2019-08-15 CN CN201910755143.6A patent/CN110705583B/zh active Active
- 2019-11-11 WO PCT/CN2019/117233 patent/WO2021027135A1/zh active Application Filing
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107016665A (zh) * | 2017-02-16 | 2017-08-04 | 浙江大学 | 一种基于深度卷积神经网络的ct肺结节检测方法 |
CN109800778A (zh) * | 2018-12-03 | 2019-05-24 | 浙江工业大学 | 一种基于难分样本挖掘的Faster RCNN目标检测方法 |
CN109919230A (zh) * | 2019-03-10 | 2019-06-21 | 西安电子科技大学 | 基于循环特征金字塔的医学图像肺结节检测方法 |
Non-Patent Citations (1)
Title |
---|
ZHAOWEI CAI, NUNO VASCONCELOS: "Cascade R-CNN: Delving into High Quality Object Detection", 2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, 18 June 2018 (2018-06-18), pages 6154 - 6162, XP033473531, DOI: :10.1109/CVPR.2018.00644 * |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113052217A (zh) * | 2021-03-15 | 2021-06-29 | 上海云从汇临人工智能科技有限公司 | 预测结果标识及其模型训练方法、装置及计算机存储介质 |
CN113780270A (zh) * | 2021-03-23 | 2021-12-10 | 京东鲲鹏(江苏)科技有限公司 | 目标检测方法和装置 |
CN113128408A (zh) * | 2021-04-21 | 2021-07-16 | 深圳大学 | 一种物品检测方法、装置、终端及存储介质 |
CN113128408B (zh) * | 2021-04-21 | 2023-09-22 | 深圳大学 | 一种物品检测方法、装置、终端及存储介质 |
CN113837194A (zh) * | 2021-09-23 | 2021-12-24 | 北京百度网讯科技有限公司 | 图像处理方法、图像处理装置、电子设备以及存储介质 |
CN113837194B (zh) * | 2021-09-23 | 2024-05-10 | 北京百度网讯科技有限公司 | 图像处理方法、图像处理装置、电子设备以及存储介质 |
CN117036227A (zh) * | 2022-09-21 | 2023-11-10 | 腾讯科技(深圳)有限公司 | 数据处理方法、装置、电子设备、介质及程序产品 |
CN116434066A (zh) * | 2023-04-17 | 2023-07-14 | 安徽农业大学 | 基于深度学习的大豆豆荚考种方法、系统及装置 |
CN116434066B (zh) * | 2023-04-17 | 2023-10-13 | 安徽农业大学 | 基于深度学习的大豆豆荚考种方法、系统及装置 |
CN118658159A (zh) * | 2024-08-19 | 2024-09-17 | 中国科学院心理研究所 | 间接免疫荧光图像的处理方法及系统 |
Also Published As
Publication number | Publication date |
---|---|
CN110705583B (zh) | 2024-03-15 |
CN110705583A (zh) | 2020-01-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
WO2021027135A1 (zh) | 细胞检测模型训练方法、装置、计算机设备及存储介质 | |
CN109978890B (zh) | 基于图像处理的目标提取方法、装置及终端设备 | |
WO2018103608A1 (zh) | 一种文字检测方法、装置及存储介质 | |
CN111145209B (zh) | 一种医学图像分割方法、装置、设备及存储介质 | |
CN110942074B (zh) | 字符切分识别方法、装置、电子设备、存储介质 | |
CN108229485B (zh) | 用于测试用户界面的方法和装置 | |
CN107403130A (zh) | 一种字符识别方法及字符识别装置 | |
CN110348522B (zh) | 一种图像检测识别方法及系统、电子设备、图像分类网络优化方法及系统 | |
CN111259878A (zh) | 一种检测文本的方法和设备 | |
WO2021139447A1 (zh) | 一种宫颈异常细胞检测装置及方法 | |
CN113537037A (zh) | 路面病害识别方法、系统、电子设备及存储介质 | |
CN112396050B (zh) | 图像的处理方法、设备以及存储介质 | |
CN110738030A (zh) | 表格重建方法、装置、电子设备及存储介质 | |
US12106474B2 (en) | Method of determining a distribution of stem cells in a cell image, electronic device, and storage medium | |
CN114298985B (zh) | 缺陷检测方法、装置、设备及存储介质 | |
CN118037694A (zh) | 一种基于自适应窗口的菌落计数方法、装置 | |
Salvi et al. | cyto‐Knet: An instance segmentation approach for multiple myeloma plasma cells using conditional kernels | |
Schüffler et al. | Overcoming an annotation hurdle: Digitizing pen annotations from whole slide images | |
CN111163332A (zh) | 视频色情度检测方法、终端及介质 | |
KR102026280B1 (ko) | 딥 러닝을 이용한 씬 텍스트 검출 방법 및 시스템 | |
CN116030472A (zh) | 文字坐标确定方法及装置 | |
CN116486467A (zh) | 眼睛检测框的确定方法、装置、设备及存储介质 | |
WO2022126923A1 (zh) | Asc-us诊断结果分流方法、装置、计算机设备及存储介质 | |
CN113573137A (zh) | 视频画布边界检测方法、系统、终端设备及存储介质 | |
CN114445751A (zh) | 视频关键帧图像轮廓特征的提取方法和装置 |
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: 19941362 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 19941362 Country of ref document: EP Kind code of ref document: A1 |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 19941362 Country of ref document: EP Kind code of ref document: A1 |