WO2021114630A1 - 医学图像样本筛查方法、装置、计算机设备和存储介质 - Google Patents

医学图像样本筛查方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2021114630A1
WO2021114630A1 PCT/CN2020/099328 CN2020099328W WO2021114630A1 WO 2021114630 A1 WO2021114630 A1 WO 2021114630A1 CN 2020099328 W CN2020099328 W CN 2020099328W WO 2021114630 A1 WO2021114630 A1 WO 2021114630A1
Authority
WO
WIPO (PCT)
Prior art keywords
medical image
image sample
lesion
unlabeled
value
Prior art date
Application number
PCT/CN2020/099328
Other languages
English (en)
French (fr)
Inventor
王俊
高鹏
Original Assignee
平安科技（深圳）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 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2021114630A1 publication Critical patent/WO2021114630A1/zh

Links

Images

Classifications

    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H30/00—ICT specially adapted for the handling or processing of medical images
    • G16H30/40—ICT specially adapted for the handling or processing of medical images for processing medical images, e.g. editing
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Definitions

  • This application relates to the field of computer vision technology, and in particular to a method, device, computer equipment and storage medium for screening medical image samples.
  • the detection of lesions and key organs based on medical images is one of the most frequently used tasks in the field of medical impact artificial intelligence-assisted diagnosis and treatment.
  • the actual medical image data collected in clinical practice has complex semantics and target layout. Interval occlusion makes accurate and effective medical imaging target detection extremely difficult.
  • the supervised learning algorithm based on deep learning has achieved certain results in many computer vision application fields. It needs to be based on a large number of labeled training samples.
  • the quality of training data is critical to the performance of the model, which means that the learning is relatively good.
  • Good features and detection models require a large number of labeled samples.
  • the labeling data When it is applied to medical imaging for lesion target detection, the labeling data requires doctors with relevant professional knowledge, and it is often difficult for doctors to have time for special labeling work. Hiring doctors or special labeling technicians to perform labeling results in very expensive labeling. High, the period of interpretation or labeling is very long.
  • the deep convolutional neural network model can process complex medical image data with good feature expression and learning capabilities.
  • target detection needs to be more accurately located
  • the diseased area is used for intelligent auxiliary diagnosis. Therefore, the difficulty index is higher, and a large number of labeled training samples are needed to fully realize its potential.
  • images of different disease types and severity levels are complex, and it is necessary to firstly find the most valuable image samples from the massive unlabeled samples, and provide enough information for the model to learn for use in medical images.
  • Robust deep learning model training for target detection When there is a serious imbalance in the number of lesions, the selection of high-value samples becomes more critical and important.
  • the inventor realizes that the current cost of labeling medical image data is very high, and the time period for image interpretation is very long.
  • a method for screening medical image samples for intelligent screening of unlabeled medical image samples including the following steps:
  • the unlabeled medical image sample set U is predicted one by one, and the prediction result of each medical image sample in the unlabeled medical image sample set U is obtained, and each medical image is judged according to the prediction result.
  • the new medical image lesion samples are verified according to the iteratively updated lesion target detection depth model C, and the iterative update is ended when the performance of the lesion target detection depth model C can no longer continue to label new samples.
  • This application also proposes a medical image sample screening device, which includes a focus target depth model initialization unit, an unlabeled medical image sample annotation value prediction unit, a focus target detection depth model iteration unit, and a model iterative update judgment unit;
  • the lesion target detection depth model initialization unit is used to select the initial labeled sample set L of the current medical image target detection task, and use the Mask-RCNN model to perform model training on the initial labeled sample set L to obtain the current medical image Focus target detection depth model C for focus target detection;
  • the unlabeled medical image sample labeling value prediction unit is configured to predict the unlabeled medical image sample set U one by one according to the lesion target detection depth model C to obtain each medical image in the unlabeled medical image sample set U The prediction result of the sample, and the annotation value of each medical image sample is judged according to the prediction result;
  • the lesion target detection depth model iteration unit is used to select medical image samples with high annotated value from the unlabeled medical image sample set U for annotation confirmation, and detect the current lesion target based on the selected medical image samples with high annotated value Deep model C is iteratively updated;
  • the model iterative update judgment unit is used to verify a new medical image lesion sample according to the iteratively updated lesion target detection depth model C. For example, the performance of the lesion target detection depth model C can no longer continue to label new samples to end the iteration Update, otherwise, the iterative unit of the detection depth model of the focus target continues to iteratively update according to the new sample.
  • the present application also provides a computer device, including a memory and a processor, the memory stores computer-readable instructions, and when the computer-readable instructions are executed by the processor, the execution of the processor includes the following steps:
  • the unlabeled medical image sample set U is predicted one by one, and the prediction result of each medical image sample in the unlabeled medical image sample set U is obtained, and each medical image is judged according to the prediction result.
  • the new medical image lesion samples are verified according to the iteratively updated lesion target detection depth model C, and the iterative update is ended when the performance of the lesion target detection depth model C can no longer continue to label new samples.
  • the present application also provides a storage medium storing computer-readable instructions, which when executed by one or more processors, cause the one or more processors to execute the steps of the medical image sample screening method.
  • this application is based on an active learning strategy. From a large number of unlabeled original medical images, some high-value samples are selected for labeling. It is not necessary to label all the samples. Each time it is selected to improve deep learning. The most valuable samples of the target detection model are added to the training, on the basis of obtaining the ideal task accuracy, effectively reducing the labeling cost and workload, and maximizing the efficiency of manual labeling of the samples. The sample with the largest amount of information is selected to accelerate model training, and the amount of label data used is significantly reduced, which provides a new implementation method for deep learning to reduce data set requirements, realizes efficient data and computing resource utilization, and saves computing resource consumption.
  • FIG. 1 is an implementation environment diagram of a method for screening medical image samples provided in an embodiment
  • Figure 2 is a block diagram of the internal structure of a computer device in an embodiment
  • FIG. 3 is a flowchart of a method for screening medical image samples in an embodiment
  • Figure 4 is a schematic diagram of the Mask-RCNN model structure in an embodiment
  • FIG. 5 is a flow chart of the technical route of a method for screening medical image samples in an embodiment
  • FIG. 6 is a flowchart of the initial focus target detection depth model C predicting the unlabeled medical image sample set U one by one in an embodiment
  • FIG. 7 is a flowchart of calculating the confidence value of the lesion target in an embodiment
  • FIG. 8 is a histogram of the number of sample target detection frames in the confidence detection of the lesion target in an embodiment
  • FIG. 9 is a flowchart of calculating the anti-disturbance stability value of the lesion target instance in an embodiment
  • FIG. 10 is a detection diagram of a target instance of a medical image sample in an embodiment
  • Fig. 11 is an example diagram of two medical image sample images in an embodiment
  • Figure 12 is a structural block diagram of a medical image sample screening device in an embodiment
  • FIG. 13 is a structural block diagram of an unlabeled medical image sample labeled value prediction unit in an embodiment.
  • FIG. 1 is an implementation environment diagram of a medical image sample screening method provided in an embodiment. As shown in FIG. 1, the implementation environment includes a computer device 110 and a terminal 120.
  • the computer device 110 is a test device, for example, a computer device used by a tester, and an automated test tool is installed on the computer device 110, for example, Appium.
  • the test application of the medical image sample screening method is installed on the terminal 120.
  • the tester can send a request to the computer device 110, the request carries an identifier, and the computer device 110 receives the request and obtains the computer device 110 according to the identifier.
  • the script corresponding to the request ID in.
  • an automated testing tool is used to execute the script to test the tested application of the medical image sample screening method on the terminal 120, and obtain the positioning result corresponding to the script.
  • the terminal 120 and the computer device 110 may be smart phones, tablet computers, notebook computers, desktop computers, etc., but are not limited thereto.
  • the computer device 110 and the terminal 110 may be connected via Bluetooth, USB (Universal Serial Bus, Universal Serial Bus) or other communication connection methods, which is not limited in this application.
  • Figure 2 is a schematic diagram of the internal structure of a computer device in an embodiment.
  • the computer device includes a processor, a non-volatile storage medium, a memory, and a network interface connected through a system bus.
  • the non-volatile storage medium of the computer device stores an operating system, a database, and computer-readable instructions.
  • the database may store control information sequences.
  • the processor can realize a A method for screening medical image samples.
  • the processor of the computer equipment is used to provide calculation and control capabilities, and supports the operation of the entire computer equipment.
  • Computer readable instructions may be stored in the memory of the computer device, and when the computer readable instructions are executed by the processor, the processor can make the processor execute a method for screening medical image samples.
  • the network interface of the computer device is used to connect and communicate with the terminal.
  • a method for screening medical image samples is proposed.
  • the method can be applied to the aforementioned computer device 110, and specifically can include the following steps 302 to 308:
  • Step 302 Select the initial labeled sample set L for the current medical image focus target detection task, and use the Mask-RCNN model to perform model training on the initial labeled sample set L to obtain the focus target detection for the current medical image focus target detection Depth model C;
  • FIG. 4 is a schematic diagram of the Mask-RCNN model structure in an embodiment, showing the structure of the Mask-RCNN model.
  • the data set D is randomly selected to obtain labels, and an initial sample set L with labels is generated, that is, the initial labeled sample set L.
  • the labeled sample set L is sent to the aforementioned Mask-RCNN model for training, and after the training is completed, a focus target detection depth model C for current medical image focus target detection is formed.
  • the original medical image data set D are all medical images with the same lesion target, for example, all are medical images of a lung lesion.
  • Step 304 Predict the unlabeled medical image sample set U one by one according to the initial focus target detection depth model C, obtain the prediction result of each medical image sample in the unlabeled medical image sample set U, and judge according to the prediction result The annotation value of each medical image sample;
  • the focus target detection depth model C is a depth model based on the Mask-RCNN model that can initially detect the same focus target, using this model C
  • the remaining unlabeled medical image samples in the original medical image data set D that is, the unlabeled medical image sample set U, are predicted one by one.
  • the predicted result is performed on the unlabeled medical image samples
  • the model is not robust enough and the performance is low.
  • Fig. 6 is a flow chart of the initial lesion target detection depth model C predicting the unlabeled medical image sample set U one by one in an embodiment, specifically predicting the unlabeled medical image sample set U one by one to obtain the unlabeled medical image
  • the prediction result of each medical image sample in the sample set U, and judging the label value of each medical image sample according to the prediction result includes steps 602-606:
  • the comprehensive strategy of selecting Uncertainty (predictive uncertainty score) and the overlapping suggestion area near the target instance to averagely intersect and compare to IOU (anti-disturbance stability score) is a comprehensive strategy to measure the performance of each medical image sample in the unlabeled medical image sample. Mark the value.
  • the uncertainty score of the target to be detected in the medical image sample is measured by the preset confidence of the detection frame; the average intersection of the recommended area near the target instance in the medical image sample is higher than the IOU (anti-disturbance stability) Score), the RPN network is used to generate candidate frames of regions that may contain objects in the medical sample image, and then the candidate frames of overlapping regions near the target instance are selected, and the intersection ratio of the target overlapping regions is calculated to obtain the anti-disturbance stability score value index.
  • Step 602 Calculate the confidence value of the lesion target of each medical image sample in the unlabeled medical image sample set U;
  • FIG. 7 is selected as a flowchart for calculating the confidence level of the focus target in an embodiment, and calculating the confidence level of the focus target specifically includes steps 702 to 704:
  • Step 702 According to the focus target detection depth model C, calculate the number of target detection frames of the focus target in each medical image sample in the unlabeled medical image sample set U;
  • the number of target detection frames in each medical image sample can be obtained, that is, the aforementioned lesion target detection depth model that has been initialized and trained is consistent with each medical image sample.
  • the in-depth model trains the marked lesion targets for regional frame selection.
  • Step 704 Calculate the focus target confidence value according to the number of target detection frames in each medical image sample.
  • the uncertainty of the target to be detected in the unlabeled medical image sample is determined by the number of target detection frames, that is, the number of focal targets predicted in the medical image sample, and the greater the number of focal targets predicted in the medical image sample, It means that the higher the uncertainty of the sample, it can be considered that the sample contains more knowledge or information for improving the performance of the focal target detection depth model C.
  • Figure 8 for the number of target detection frames for the confidence level of each stage in the medical image sample.
  • Figure 8 is a histogram of the number of target detection frames for the focus target confidence detection in an embodiment. The histogram shows the number of target detection frames from 0 to 1. The number distribution of unlabeled medical image samples in the confidence interval.
  • a target detection frame with a confidence level of 0 to 0.2 is determined as the background in the sample, and a target detection frame with a confidence level of 0.3 is uncertain as the foreground or background of the sample ,
  • the target detection frame with a confidence of 0.8-1 is confirmed as the foreground in the sample, and the uncertainty index of each medical image sample is defined as Unc(x,L,u), where L represents the labeled sample, u Represents an unlabeled sample.
  • L represents the labeled sample
  • u Represents an unlabeled sample.
  • Step 604 Calculate the anti-disturbance stability value of the lesion target instance of each medical image sample in the unlabeled medical image sample set U;
  • FIG. 9 shown in FIG. 9 is a flowchart of calculating the anti-disturbance stability value of the focus target instance in an embodiment, which specifically includes steps 902 to 906:
  • Step 902 According to the focus target detection depth model C, the RPN network in the Mask-RCNN model is used to generate an area containing the focus target in each medical image sample in the unlabeled medical image sample set U;
  • Step 904 Select two focus target areas that are closest to the focus target instance
  • the first N regions that are most likely to contain the target are retained, preferably the two most likely to contain the target that are closest to the lesion target instance. area.
  • the two medical image samples 101 and 102 are respectively detected through the RPN network to obtain two region candidate frames that may contain objects that are closest to the target instance of the lesion.
  • the medical image sample 101 includes two area candidate frames 1012 and 1013 that are closest to the lesion target instance area 1011, and the medical image sample 102 includes two area candidate frames 1022, 1023 that are closest to the lesion target instance area 1021.
  • Step 906 Calculate the ratio of the intersection and union of the focus target instance and the two focus target areas overlapped, as the anti-disturbance stability value of the focus target instance.
  • the anti-disturbance stability index IOU of the target instance in the medical image by calculating the ratio of the intersection and union of the target instance of the lesion and the overlap of the two lesion target regions, it is used to detect the anti-disturbance stability index IOU of the target instance in the medical image, if the anti-disturbance stability index A larger range indicates that there is more information around the target instance in the medical image that is easy to confuse the model.
  • the labeling value of the sample is very high.
  • the maximum score of the lesion target instance candidate frame 1011 of the medical image sample 101 in the left image in Figure 10 is 0.98
  • the scores of the two closest regional candidate frames 1012 and 1023 are 0.92 and 0.91, respectively
  • the interval of the anti-disturbance stability index IOU The maximum score of the lesion target instance candidate frame 1021 of the medical image sample 102 on the right is 0.98
  • the scores of the two closest regional candidate frames 1022 and 1023 are 0.72 and 0.65, respectively.
  • the lesion target instance candidate frame 1021 and The intersections of the region candidate frames 1022 and 1023 are 0.4 and 0.3, respectively, and the interval of the anti-disturbance stability index IOU is 0.3 to 0.9.
  • the range of the anti-disturbance stability index of the sample 101 is larger, that is, there is more information that can be easily confused around the medical image sample 102 on the right, which can give the focus target detection depth model C. With more learning content, the value of samples is higher for active learning algorithms.
  • Step 606 According to the focus target confidence value and the focus target instance anti-disturbance stability value, combined with the active learning algorithm, the sample label value is calculated.
  • calculating the sample label value according to the focus target confidence and the focus target instance anti-disturbance stability combined with the active learning algorithm includes the following steps:
  • Figure 11 is an example diagram of two unlabeled medical image sample images in one embodiment.
  • the medical sample image shown on the left is calculated through the above-mentioned prediction and calculation of the medical image sample, and the sample labeling value is obtained as 0.06 points, while the right figure shows The medical sample image is calculated through the above-mentioned medical image sample prediction, and the sample labeling value score is 0.34.
  • the unlabeled medical image sample shown on the right has more labeling value and is more effective for the training and improvement of the focal target detection depth model C .
  • Step 306 After selecting the medical image samples with high annotation value from the unlabeled medical image sample set U for annotation confirmation, iteratively update the current focus target detection depth model C;
  • the medical image samples with high annotation value in the unlabeled medical image sample set U are obtained through the above steps.
  • the medical image samples with high annotation value can be confirmed according to the experience value, and the medical images with high annotation value can be screened out.
  • the selected sample can be annotated by the expert, the sample after the labeling confirmation is put into the training sample set, and then the initial focus target detection depth model C is updated, which realizes the focus target detection depth model C through the active learning algorithm The purpose of the update.
  • step 308 the new medical image lesion samples are verified according to the iteratively updated lesion target detection depth model C, until the performance of the lesion target detection depth model C can no longer continue to label new samples, the iterative update is ended.
  • Fig. 5 is a flow chart of the technical route of the medical image sample screening method in an embodiment.
  • Learning from the initial labeled sample set L first trains the focus target detection depth model C, and then extracts the unlabeled medical image sample set U Samples are predicted and calculated one by one through the lesion target detection depth model C.
  • the calculation process includes the calculation of the target confidence and the anti-disturbance stability of the target instance for the medical image samples, and intelligently screens the samples that meet the requirements and integrates the samples.
  • the target confidence of the target instance and the anti-disturbance stability of the target instance get the sample value of the active learning algorithm.
  • the medical image sample set confirmed by labeling is used as the training sample to update the initial The lesion target detection depth model C, on the basis of ensuring the accuracy of the task, effectively reduces the labeling cost and workload, and maximizes the labeling efficiency of medical image samples.
  • the above-mentioned medical image sample screening method is carried out on two tasks of OCT imaging lesion detection and CT imaging cerebral hemorrhage detection. It is shown that the method can achieve almost the same performance with only about 66% of the sample size of the complete data set.
  • the method of this application integrates deep learning and active learning based on the good feature expression ability of deep models. From a large number of unlabeled original medical image samples, select high-quality samples. Sample labeling, filtering samples with low quality, without labeling all samples, each selection can add training to the most valuable samples to improve the model, which greatly improves the efficiency of the model.
  • a structural block diagram of a medical image sample screening device is provided.
  • the device 12 may be integrated into the above-mentioned computer device 110, and may specifically include a focus target depth model initialization unit 1201, An unlabeled medical image sample annotated value prediction unit 1202, a focus target detection depth model iterative unit 1203, and a model iterative update judgment unit 1204;
  • the lesion target detection depth model initialization unit 1201 is used to select the initial labeled sample set L of the current medical image target detection task, and use the Mask-RCNN model to perform model training on the initial labeled sample set L to obtain the current medical Focus target detection depth model C for image focus target detection;
  • the unlabeled medical image sample labeling value prediction unit 1202 is configured to predict the unlabeled medical image sample set U one by one according to the lesion target detection depth model C to obtain each medical image in the unlabeled medical image sample set U The prediction result of the image sample, and the label value of each medical image sample is judged according to the prediction result;
  • the lesion target detection depth model iteration unit 1203 is used to select medical image samples with high annotated value from the unlabeled medical image sample set U for annotation confirmation, and compare the current lesion target based on the selected medical image samples with high annotation value.
  • the detection depth model C is iteratively updated;
  • the model iterative update judging unit 1204 is used to verify the new medical image lesion samples according to the iteratively updated lesion target detection depth model C. If the performance of the lesion target detection depth model C can no longer continue to label new samples, it ends Iterative update, otherwise, the iterative unit of the detection depth model of the focus target continues to update iteratively according to the new sample.
  • FIG. 13 is a structural block diagram of an unlabeled medical image sample labeled value prediction unit in an embodiment.
  • the labeling value prediction unit 13 of the unlabeled medical image sample includes a focus target confidence calculation module 1301, a focus target instance anti-disturbance stability calculation module 1302, and an active learning strategy module 1303;
  • the focus target confidence calculation module 1301 is used to calculate the focus target confidence of each medical image sample in the unlabeled medical image sample set U;
  • the lesion target instance anti-disturbance stability calculation module 1302 is used to calculate the lesion target instance anti-disturbance stability calculation module of each medical image sample in the unlabeled medical image sample set U.
  • the active learning strategy module 1303 is configured to calculate the annotation value of the unlabeled medical image sample according to the focus target confidence and the focus target instance anti-disturbance stability combined with an active learning algorithm.
  • a computer device in one embodiment, includes a memory, a processor, and a computer program that is stored on the memory and can run on the processor, and the processor executes the computer The following steps are implemented during the program:
  • the unlabeled medical image sample set U is predicted one by one, and the prediction result of each medical image sample in the unlabeled medical image sample set U is obtained, and each medical image is judged according to the prediction result.
  • the new medical image lesion samples are verified according to the iteratively updated lesion target detection depth model C, and the iterative update is ended when the performance of the lesion target detection depth model C can no longer continue to label new samples.
  • the unlabeled medical image sample set U is predicted one by one according to the initial focus target detection depth model C, and the prediction result of each medical image sample in the unlabeled medical image sample set U is obtained, and Judging the annotation value of each medical image sample according to the prediction result includes the following steps:
  • the sample label value is calculated.
  • the calculation of the confidence value of the lesion target of each medical image sample in the unlabeled medical image sample set U includes the following steps:
  • the focus target detection depth model C calculate the number of target detection frames of the focus target in each medical image sample in the unlabeled medical image sample set U;
  • the confidence value of the lesion target is calculated according to the number of target detection frames in each medical image sample.
  • the calculation of the anti-disturbance stability value of the lesion target instance of each medical image sample in the unlabeled medical image sample set U includes the following steps:
  • the RPN network in the Mask-RCNN model is used to generate the region containing the lesion target in each medical image sample in the unlabeled medical image sample set U;
  • calculating the sample labeling value according to the focus target confidence and the focus target instance anti-disturbance stability combined with an active learning algorithm includes the following steps:
  • a storage medium storing computer-readable instructions.
  • the one or more processors perform the following steps:
  • the unlabeled medical image sample set U is predicted one by one, and the prediction result of each medical image sample in the unlabeled medical image sample set U is obtained, and each medical image is judged according to the prediction result.
  • the new medical image lesion samples are verified according to the iteratively updated lesion target detection depth model C, and the iterative update is ended when the performance of the lesion target detection depth model C can no longer continue to label new samples.
  • the unlabeled medical image sample set U is predicted one by one according to the initial focus target detection depth model C, and the prediction result of each medical image sample in the unlabeled medical image sample set U is obtained, and Judging the annotation value of each medical image sample according to the prediction result includes the following steps:
  • the sample label value is calculated.
  • the calculation of the confidence value of the lesion target of each medical image sample in the unlabeled medical image sample set U includes the following steps:
  • the focus target detection depth model C calculate the number of target detection frames of the focus target in each medical image sample in the unlabeled medical image sample set U;
  • the confidence value of the lesion target is calculated according to the number of target detection frames in each medical image sample.
  • the calculation of the anti-disturbance stability value of the lesion target instance of each medical image sample in the unlabeled medical image sample set U includes the following steps:
  • the RPN network in the Mask-RCNN model is used to generate the region containing the lesion target in each medical image sample in the unlabeled medical image sample set U;
  • calculating the sample labeling value according to the focus target confidence and the focus target instance anti-disturbance stability combined with an active learning algorithm includes the following steps:
  • the computer program can be stored in a computer readable storage medium, and the program can be stored in a computer readable storage medium. When executed, it may include the procedures of the above-mentioned method embodiments.
  • the aforementioned storage medium may be a non-volatile storage medium such as a magnetic disk, an optical disc, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM), etc.
  • the computer-readable storage medium may be non-volatile or volatile.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • Evolutionary Computation (AREA)
  • Public Health (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Radiology & Medical Imaging (AREA)
  • Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

一种医学图像样本筛查方法、装置、计算机设备和存储介质，涉及计算机视觉技术领域，用于对未标注的医学图像样本进行智能筛查，利用Mask-RCNN模型对已标注样本集进行模型训练，以获取病灶目标检测深度模型；根据病灶目标检测深度模型对未标注医学图像样本集进行预测，得到每个医学图像样本的预测结果并判断标注价值；选取标注价值高的医学图像样本进行标注确认后，对病灶目标检测深度模型进行迭代更新，直到病灶目标检测深度模型的性能不能继续标注新的样本时结束迭代更新。这种医学图像样本筛查方法和装置通过模拟医学专家智能学习模式去高效地诊断和决策，智能化程度高，处理速度快，有效地解决了标注效率低的问题。

Description

医学图像样本筛查方法、装置、计算机设备和存储介质
本申请要求申请号为2020104686909，申请日为2020年05月28日，发明创造名称为“医学图像样本筛查方法、装置、计算机设备和存储介质”的专利申请的优先权。
技术领域
本申请涉及计算机视觉技术领域，特别是涉及一种医学图像样本筛查方法、装置、计算机设备和存储介质。
背景技术
基于医学图像进行病灶、关键器官等目标的检测，是医疗影响人工智能辅助诊疗领域使用频率较高的任务之一，临床采集的实际医学影像数据本身具有复杂语义和目标布局，不同类型病变区域之间的遮挡使得准确有效的医学影像目标检测变得异常困难。
目前，通过深度学习为主的监督学习算法在很多计算机视觉应用领域取得了一定的效果，其需要基于大量的标注训练样本，训练数据的质量对于模型的性能影响至关重要，意味着学习到较好的特征和检测模型需要大量的标注样本。在应用到医学影像进行病灶目标检测时，标注数据需要具有相关专业知识的医生，而医生往往很难有时间来进行专门的标注工作，聘请医生或专门的标注技术人员进行标注导致标注的成本很高，判读或标注的周期都很长。
深度卷积神经网络模型凭借良好的特征表达和学习能力可处理复杂医学图像数据，但是，相比基于医学图像实现分类，比如诊断患者是否患病、病情轻重程度分级，目标检测需要进一步准确地定位出病变区域用于智能辅助诊断，因此，难度指数更高，更加需要大量标记的训练样本才能充分发挥其潜力。医学图像智能分析这类领域中，不同疾病类型和严重程度的图像纷繁复杂，需要从海量的未标注样本中优先找出价值最大的图像样本，提供足够的信息给模型学习，以用于医学图像目标检测的稳健深度学习模型训练，当病灶之间存在严重的数量不平衡的情况时，高价值样本的筛选也越发关键和重要。
因此，发明人意识到目前对于医学影像数据进行标注的成本很高，且图像判读的时间周期很长。
发明内容
基于此，有必要针对医学图像样本检测监督学习时，智能筛选样本的模型效率低下，准确率不高的问题，提出一种利用深度卷积神经网络模型智能筛选有价值的医学图像样本，并根据有价值的医学图像样本对模型进行迭代改进，以进一步提高模型智能标注病灶目标的水平。
一种医学图像样本筛查方法，用于对未标注的医学图像样本进行智能筛查，包括如下步骤：
选取针对当前医疗影像病灶目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
根据所述初始病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对 当前的病灶目标检测深度模型C进行迭代更新；
根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，直到病灶目标检测深度模型C的性能不能再继续标注新的样本时，则结束迭代更新。
本申请还提出一种医学图像样本筛查装置，所述装置包括病灶目标深度模型初始化单元、未标注医学图像样本标注价值预测单元、病灶目标检测深度模型迭代单元和模型迭代更新判断单元；
所述病灶目标检测深度模型初始化单元，用于选取当前医疗影像目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
所述未标注医学图像样本标注价值预测单元，用于根据所述病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
所述病灶目标检测深度模型迭代单元，用于选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认，根据选取的标注价值高的医学图像样本对当前的病灶目标检测深度模型C进行迭代更新；
所述模型迭代更新判断单元，用于根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，如病灶目标检测深度模型C的性能不能再继续标注新的样本时结束迭代更新，否则所述病灶目标检测深度模型迭代单元继续根据新的样本进行迭代更新。
本申请还提供一种计算机设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述计算机可读指令被所述处理器执行时，使得所述处理器执行包括如下步骤：
选取针对当前医疗影像病灶目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
根据所述初始病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对当前的病灶目标检测深度模型C进行迭代更新；
根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，直到病灶目标检测深度模型C的性能不能再继续标注新的样本时，则结束迭代更新。
本申请还提供一种存储有计算机可读指令的存储介质，所述计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器执行上述医学图像样本筛查方法的步骤。
与现有技术相比较，本申请基于主动学习策略，从未标注的大量原始医学图像中，通过挑选部分高价值样本进行标注，不需要对所有的样本进行标注，每次都挑选对改进深度学习目标检测模型最有价值的样本加入训练，在获取理想任务精度的基础上，有效地减少了标注代价和工作量，最大化样本人工标注效率。选择信息量最大的样本来加速模型训练，使用标签数据量明显降低，为深度学习降低数据集要求提供了新的实现方法，实现高效的数据和计算资源利用，节省计算资源消耗。
另外，通过结合目标检测模型的预测输出，将主动学习与主流的目标检测模型融合在一起，从而可以显著地节省训练深度神经网络目标检测器的标注成本。在此基础上可训练得到泛化能力更强更准确的医学图像目标检测模型，减少网络过拟合以更好的适应医学应用场景。本方法从已获得的标签中取得的知识以探索决策边界，在有限的计算资源或者标注成本条件下，探索性地主动挖掘抽取高价值的小数据集，以此开展模型训练并做出决定，于复杂多变的状态空间里面，通过模拟医学专家智能学习模式去高效地诊断和决策，智能 化程度高，处理速度快，在保证目标检测性能的同时大大节省了训练目标检测器的数据标注成本，有效地解决了标注效率低的问题。
附图说明
图1为一个实施例中提供的医学图像样本筛查方法的实施环境图；
图2为一个实施例中计算机设备的内部结构框图；
图3为一个实施例中医学图像样本筛查方法的流程图；
图4为一个实施例中Mask-RCNN模型结构示意图；
图5为一个实施例中医学图像样本筛查方法的技术路线流程图；
图6为一个实施例中初始病灶目标检测深度模型C对未标注的医学图像样本集U进行逐一预测的流程图；
图7为一个实施例中计算病灶目标置信度值的流程图；
图8为一个实施例中病灶目标置信度检测的样本目标检测框数量柱状图；
图9为一个实施例中计算病灶目标实例抗扰动稳定度值的流程图；
图10为一个实施例中医学图像样本的目标实例检测图；
图11为一个实施例中两个医学图像样本图像示例图；
图12为一个实施例中医学图像样本筛查装置的结构框图；
图13为一个实施例中未标注医学图像样本标注价值预测单元的结构框图。
本申请目的的实现、功能特点及优点将结合实施例，参照附图做进一步说明。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
图1为一个实施例中提供的医学图像样本筛查方法的实施环境图，如图1所示，在该实施环境中，包括计算机设备110以及终端120。
计算机设备110为测试设备，例如为测试人员使用的电脑等计算机设备，计算机设备110上安装有自动化测试工具，例如可以为Appium。终端120上安装有医学图像样本筛查方法的被测应用，当需要测试时，测试人员可以在计算机设备110发出请求，该请求中携带有标识，计算机设备110接收请求，根据标识获取计算机设备110中与请求标识对应的脚本。然后利用自动化测试工具执行该脚本，对终端120上的医学图像样本筛查方法被测应用进行测试，并获取脚本对应的定位结果。
需要说明的是，终端120以及计算机设备110可为智能手机、平板电脑、笔记本电脑、台式计算机等，但并不局限于此。计算机设备110以及终端110可以通过蓝牙、USB(Universal Serial Bus，通用串行总线)或者其他通讯连接方式进行连接，本申请在此不做限制。
图2为一个实施例中计算机设备的内部结构示意图。如图2所示，该计算机设备包括通过系统总线连接的处理器、非易失性存储介质、存储器和网络接口。其中，该计算机设备的非易失性存储介质存储有操作系统、数据库和计算机可读指令，数据库中可存储有控件信息序列，该计算机可读指令被处理器执行时，可使得处理器实现一种医学图像样本筛查方法。该计算机设备的处理器用于提供计算和控制能力，支撑整个计算机设备的运行。该计算机设备的存储器中可存储有计算机可读指令，该计算机可读指令被处理器执行时，可使得处理器执行一种医学图像样本筛查方法。该计算机设备的网络接口用于与终端连接通信。本领域技术人员可以理解，图2中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定，具体的计算机设备可以包括比图中所示更多或更少的部件，或者组合某些部件，或者具有不同的部件布置。
如图3所示，在一个实施例中，提出了一种医学图像样本筛查方法，该方法可以应用于上述的计算机设备110中，具体可以包括以下步骤302～308：
步骤302，选取针对当前医疗影像病灶目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
在本实施例中，对于医学图像的检测模型采用目前主流的深度学习目标检测模型Mask-RCNN，图4为一个实施例中Mask-RCNN模型结构示意图，示出了该Mask-RCNN模型的结构。设存在一个原始的医学图像数据集D，对该数据集D进行随机挑选获取标签，并生成带有标签的初始化样本集L，即初始已标注样本集L。再将该已标注样本集L送入到上述Mask-RCNN模型中进行训练，训练结束后形成当前医疗影像病灶目标检测的病灶目标检测深度模型C。应当说明的是，原始的医学图像数据集D均为病灶目标相同的医学图像，比如，均为肺部的病灶医学图像。
步骤304，根据所述初始病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
在本实施例中，在得到病灶目标检测深度模型C后，该病灶目标检测深度模型C即为一个建立在Mask-RCNN模型基础上的、能够初步检测相同病灶目标的深度模型，用该模型C对上述原始的医学图像数据集D中余下的没有打标签的医学图像样本，即未标注医学图像样本集U，进行逐一预测，此处，预测的结果是对未进行标签标注的医学图像样本进行初步的检测，因初始病灶目标深度检测模型C的实例学习较少，模型还不够健壮，性能较低。
图6为一个实施例中初始病灶目标检测深度模型C对未标注的医学图像样本集U进行逐一预测的流程图，具体对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值包括步骤602～606：
在本实施例中，选择Uncertainty(预测不确定性得分)和目标实例附近重叠建议区平均交并比IOU(抗扰动稳定度得分)的综合策略衡量未标注医学图像样本中每个医学图像样本的标注价值。其中，对于医学图像样本中待检测目标的不确定性得分，是通过检测框预置置信度来进行衡量的；对于医学图像样本中目标实例附近重叠建议区平均交并比IOU(抗扰动稳定度得分)，则是采用RPN网络来生成医学样本图像中可能包含物体的区域候选框，再选择目标实例附近的重叠区域候选框，计算目标重叠区域的交并比得到抗扰动稳定度得分价值指标。
步骤602，计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值；
在本实施例中，选择所示图7为一个实施例中计算病灶目标置信度的流程图，计算病灶目标置信度值，具体包括步骤702～704：
步骤702，根据所述病灶目标检测深度模型C，计算未标注的医学图像样本集U中每个医学图像样本中的病灶目标的目标检测框数量；
通过病灶目标检测深度模型C对每个医学图像样本进行检测，可以得到每个医学图像样本中的目标检测框数量，即前述已经经过初始化训练的病灶目标检测深度模型对医学图像样本中每个符合深度模型训练标注的病灶目标进行区域框选。
步骤704，根据每个医学图像样本中的目标检测框数量计算出病灶目标置信度值。
对于未标注医学图像样本中待检测目标的不确定性，是通过目标检测框数量的多少，即病灶目标在医学图像样本中预测的数量，而病灶目标在医学图像样本中预测的数量越多，表示样本的不确定性越高，则可认为该样本中包含更多的改良所述病灶目标检测深度模型C的性能的知识或信息。医学图像样本中各个阶段的置信度的目标检测框数量参见图8所 示，图8为一个实施例中病灶目标置信度检测的样本目标检测框数量柱状图，该柱状图显示了0～1各个置信度区间中未标注医学图像样本的数量分布情况，如置信度为0～0.2区间的目标检测框确定为样本中的背景，置信度为0.3的目标检测框则不确定是样本的前景或背景，置信度为0.8～1区间的目标检测框确认为样本中的前景，将每个医学图像样本的不确定度指标定义为Unc(x,L,u)，其中，L代表已标注样本，u代表未标注样本，对于样本中目标的预测结果越接近0.5的样本，表示当前病灶目标检测深度模型对样本中目标的信息具有较高的不确定性，即样本需要进行标注的价值就越高。
步骤604，计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值；
在本实施例中，所示图9为一个实施例中计算病灶目标实例抗扰动稳定度值的流程图，具体包括步骤902～906：
步骤902，根据所述病灶目标检测深度模型C，采用Mask-RCNN模型中的RPN网络来生成未标注的医学图像样本集U中每个医学图像样本中包含病灶目标的区域；
步骤904，选取与病灶目标实例最接近的两个病灶目标区域；
在本实施例中，采用RPN网络生成医学图像样本中可能包含物体的区域候选框后，保留前N个最可能包含目标的区域，优选为与病灶目标实例最接近的两个最可能包含目标的区域。参见图10所示一个实施例中医学图像样本的目标实例检测图，通过RPN网络对两个医学图像样本101、102分别进行检测得到可能包含物体的两个最接近病灶目标实例的区域候选框，医学图像样本101包含两个最接近病灶目标实例区域1011的区域候选框1012、1013，而医学图像样本102则包含两个最接近病灶目标实例区域1021的区域候选框1022、1023。
步骤906，计算病灶目标实例与两个病灶目标区重叠的交集与并集的比值，作为病灶目标实例的抗扰动稳定度值。
在本实施例中，通过计算病灶目标实例与所述两个病灶目标区域重叠的交集与并集的比值，用于检测医学图像中目标实例的抗扰动稳定度指标IOU，如果抗扰动稳定度指标区间范围更大，则说明医学图像中目标实例的周围存在更多对于模型容易混淆的信息量，这对于主动学习算法来说，样本的标注价值就非常高。如图10中左图医学图像样本101的病灶目标实例候选框1011的最大分数为0.98，两个最接近的区域候选框1012和1023的分数分别为0.92和0.91，抗扰动稳定度指标IOU的区间为0.9～1，右图医学图像样本102的病灶目标实例候选框1021的最大分数为0.98，两个最接近的区域候选框1022和1023的分数分别为0.72和0.65，病灶目标实例候选框1021与区域候选框1022、1023的交集分别为0.4和0.3，其抗扰动稳定度指标IOU的区间为0.3～0.9，相比而言，图10中右图的医学图像样本102相比左图的医学图像样本101的抗扰动稳定度指标的区间范围更大，也就是右图医学图像样本102中的周围存在了更多容易被混淆的信息量，这些信息量能够给所述病灶目标检测深度模型C带来更多的学习内容，对于主动学习算法而言样本的价值更高。
步骤606，根据所述病灶目标置信度值和病灶目标实例抗扰动稳定度值，结合主动学习算法计算出样本标注价值。
在本实施例中，根据所述病灶目标置信度和病灶目标实例抗扰动稳定度，结合主动学习算法计算出样本标注价值包括如下步骤：
选取病灶目标置信度值为0.4～0.7之间的医学图像样本，表达为max Unc(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，Unc表示为置信度；
选取抗扰动稳定度值为0.3～0.9之间的医学图像样本，表达为IOU(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，IOU表示为抗扰动稳定度值；
通过主动学习算法公式max f(x,L,u)＝max Unc(x,L,u)*IOU(x,L,u)β修正计算出每个医学图像样本的标注价值max f(x,L,u)，其中，β是调解所述抗扰动稳定度值比重的参数。
图11为一个实施例中两个未标注医学图像样本图像示例图，左图示出的医学样本图像通过上述医学图像样本的预测计算，得到样本标注价值分为0.06分，而右图示出的医学样本图像通过上述医学图像样本的预测计算，得到样本标注价值得分为0.34，显然，右图示出的未标注医学图像样本更具有标注的价值，对于病灶目标检测深度模型C的训练提升更加有效。
步骤306，选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对当前的病灶目标检测深度模型C进行迭代更新；
在本实施例中，通过上述步骤得到未标注医学图像样本集U中标注价值高的医学图像样本，此处标注价值高的医学图像样本可根据经验值来确认，筛选出标注价值高的医学图像样本后，可由专家对选择的样本进行标注，将标注确认后的样本放入到训练样本集中，然后更新初始的病灶目标检测深度模型C，即实现了通过主动学习算法对病灶目标检测深度模型C进行更新的目的。
步骤308，根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，直到病灶目标检测深度模型C的性能不能再继续标注新的样本时，则结束迭代更新。
图5为一个实施例中医学图像样本筛查方法的技术路线流程图，从初始已标注样本集L中学习首先训练出病灶目标检测深度模型C，然后从未标注的医学图像样本集U中抽取样本，通过病灶目标检测深度模型C对样本进行逐一预测计算，计算过程包括对医学图像样本进行目标置信度和目标实例抗扰动稳定度的计算，并对预测符合要求的样本进行智能筛选，综合样本的目标置信度和目标实例抗扰动稳定度得到主动学习算法的样本价值，最后，经过对样本价值高的样本进行标签的标注确认，经过标签标注确认的医学图像样本集合，作为训练样本更新初始的病灶目标检测深度模型C，在保证任务精度的基础上，有效减少了标注代价和工作量，最大化了医学图像样本的标注效率。
上述医学图像样本筛查方法在OCT影像病灶检测、CT影像脑出血检测两个任务上进行实验表明，方法能够仅用完整数据集的大约66％的样本量实现几乎同等的性能。相比一般情况下通过随机挑选的方法挑选样本，本申请方法融合深度学习和主动学习的手段能够基于深度模型良好的特征表达能力，从未标注的大量原始医学图像样本中，通过挑选高质量的样本标注，过滤质量较低的样本，不需要对所有的样本进行标注，每次的挑选都能够对改进提升模型最有价值的样本加入训练，大大提高了模型的效率。以OCT影像病灶检测为例，通过每次逐步添加1000张样本图像进行模型训练的结果对比可以发现，主动学习策略可以在筛选出的8000张样本进行标注后训练，达到与随机挑选12000张训练的模型精度。
如图12所示，在一个实施例中，提供了一种医学图像样本筛查装置的结构框图，该装置12可以集成于上述的计算机设备110中，具体可以包括病灶目标深度模型初始化单元1201、未标注医学图像样本标注价值预测单元1202、病灶目标检测深度模型迭代单元1203和模型迭代更新判断单元1204；
所述病灶目标检测深度模型初始化单元1201，用于选取当前医疗影像目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
所述未标注医学图像样本标注价值预测单元1202，用于根据所述病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
所述病灶目标检测深度模型迭代单元1203，用于选取所述未标注医学图像样本集U 中标注价值高的医学图像样本进行标注确认，根据选取的标注价值高的医学图像样本对当前的病灶目标检测深度模型C进行迭代更新；
所述模型迭代更新判断单元1204，用于根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，如病灶目标检测深度模型C的性能不能再继续标注新的样本时结束迭代更新，否则所述病灶目标检测深度模型迭代单元继续根据新的样本进行迭代更新。
图13为一个实施例中未标注医学图像样本标注价值预测单元的结构框图。所述未标注医学图像样本标注价值预测单元13包括病灶目标置信度计算模块1301、病灶目标实例抗扰动稳定度计算模块1302和主动学习策略模块1303；
所述病灶目标置信度计算模块1301用于计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度；
所述病灶目标实例抗扰动稳定度计算模块1302用于计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度计算模块。
所述主动学习策略模块1303用于根据所述病灶目标置信度和病灶目标实例抗扰动稳定度，结合主动学习算法计算出所述未标注医学图像样本的标注价值。
在一个实施例中，提出了一种计算机设备，所述计算机设备包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现以下步骤：
选取针对当前医疗影像病灶目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
根据所述初始病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对当前的病灶目标检测深度模型C进行迭代更新；
根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，直到病灶目标检测深度模型C的性能不能再继续标注新的样本时，则结束迭代更新。
在一个实施例中，根据所述初始病灶目标检测深度模型C对未标注的医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值包括如下步骤：
计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值；
计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值；
根据所述病灶目标置信度值和病灶目标实例抗扰动稳定度值，结合主动学习算法计算出样本标注价值。
在一个实施例中，所述计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值包括如下步骤：
根据所述病灶目标检测深度模型C，计算未标注的医学图像样本集U中每个医学图像样本中的病灶目标的目标检测框数量；
根据每个医学图像样本中的目标检测框数量计算出病灶目标置信度值。
在一个实施例中，所述计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值包括如下步骤：
根据所述病灶目标检测深度模型C，采用Mask-RCNN模型中的RPN网络来生成未标注的医学图像样本集U中每个医学图像样本中包含病灶目标的区域；
选取与病灶目标实例最接近的两个病灶目标区域；
计算病灶目标实例与两个病灶目标区重叠的交集与并集的比值，作为病灶目标实例的抗扰动稳定度值。
在一个实施例中，根据所述病灶目标置信度和病灶目标实例抗扰动稳定度，结合主动学习算法计算出样本标注价值包括如下步骤：
选取病灶目标置信度值为0.4～0.7之间的医学图像样本，表达为max Unc(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，Unc表示为置信度；
选取抗扰动稳定度值为0.3～0.9之间的医学图像样本，表达为IOU(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，IOU表示为抗扰动稳定度值；
通过主动学习算法公式max f(x,L,u)＝max Unc(x,L,u)*IOU(x,L,u)β修正计算出每个医学图像样本的标注价值max f(x,L,u)，其中，β是调解所述抗扰动稳定度值比重的参数。
在一个实施例中，提出了一种存储有计算机可读指令的存储介质，该计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器执行以下步骤：
选取针对当前医疗影像病灶目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
根据所述初始病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对当前的病灶目标检测深度模型C进行迭代更新；
根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，直到病灶目标检测深度模型C的性能不能再继续标注新的样本时，则结束迭代更新。
在一个实施例中，根据所述初始病灶目标检测深度模型C对未标注的医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值包括如下步骤：
计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值；
计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值；
根据所述病灶目标置信度值和病灶目标实例抗扰动稳定度值，结合主动学习算法计算出样本标注价值。
在一个实施例中，所述计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值包括如下步骤：
根据所述病灶目标检测深度模型C，计算未标注的医学图像样本集U中每个医学图像样本中的病灶目标的目标检测框数量；
根据每个医学图像样本中的目标检测框数量计算出病灶目标置信度值。
在一个实施例中，所述计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值包括如下步骤：
根据所述病灶目标检测深度模型C，采用Mask-RCNN模型中的RPN网络来生成未标注的医学图像样本集U中每个医学图像样本中包含病灶目标的区域；
选取与病灶目标实例最接近的两个病灶目标区域；
计算病灶目标实例与两个病灶目标区重叠的交集与并集的比值，作为病灶目标实例的抗扰动稳定度值。
在一个实施例中，根据所述病灶目标置信度和病灶目标实例抗扰动稳定度，结合主动学习算法计算出样本标注价值包括如下步骤：
选取病灶目标置信度值为0.4～0.7之间的医学图像样本，表达为max Unc(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，Unc表示为置信度；
选取抗扰动稳定度值为0.3～0.9之间的医学图像样本，表达为IOU(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，IOU表示为抗扰动稳定度值；
通过主动学习算法公式max f(x,L,u)＝max Unc(x,L,u)*IOU(x,L,u)β修正计算出每个医学图像样本的标注价值max f(x,L,u)，其中，β是调解所述抗扰动稳定度值比重的参数。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，该计算机程序可存储于一计算机可读取存储介质中，该程序在执行时，可包括如上述各方法的实施例的流程。其中，前述的存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory，ROM)等非易失性存储介质，或随机存储记忆体(Random Access Memory，RAM)等。所述计算机可读存储介质可以是非易失性，也可以是易失性。
以上所述实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对本申请专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请专利的保护范围应以所附权利要求为准。

Claims (20)

  1. 一种医学图像样本筛查方法，其中，用于对未标注的医学图像样本进行智能筛查，包括如下步骤：
    选取针对当前医疗影像病灶目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
    根据所述初始病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
    选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对当前的病灶目标检测深度模型C进行迭代更新；
    根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，直到病灶目标检测深度模型C的性能不能再继续标注新的样本时，则结束迭代更新。
  2. 如权利要求1所述的医学图像样本筛查方法，其中，根据所述初始病灶目标检测深度模型C对未标注的医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值包括如下步骤：
    计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值；
    计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值；
    根据所述病灶目标置信度值和病灶目标实例抗扰动稳定度值，结合主动学习算法计算出样本标注价值。
  3. 如权利要求2所述的医学图像样本筛查方法，其中，所述计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值包括如下步骤：
    根据所述病灶目标检测深度模型C，计算未标注的医学图像样本集U中每个医学图像样本中的病灶目标的目标检测框数量；
    根据每个医学图像样本中的目标检测框数量计算出病灶目标置信度值。
  4. 如权利要求3所述的医学图像样本筛查方法，其中，
    所述未标注医学图像样本中待检测目标的不确定性，通过所述病灶目标在医学图像样本中预测的数量，其中，
    病灶目标在医学图像样本中预测的数量越多，表示样本的不确定性越高。
  5. 如权利要求3所述的医学图像样本筛查方法，其中，所述计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值包括如下步骤：
    根据所述病灶目标检测深度模型C，采用Mask-RCNN模型中的RPN网络来生成未标注的医学图像样本集U中每个医学图像样本中包含病灶目标的区域；
    选取与病灶目标实例最接近的两个病灶目标区域；
    计算病灶目标实例与两个病灶目标区重叠的交集与并集的比值，作为病灶目标实例的抗扰动稳定度值。
  6. 如权利要求5所述的医学图像样本筛查方法，其中，采用RPN网络生成医学图像样本中可能包含物体的区域候选框后，保留前N个最可能包含目标的区域，其中，与病灶目标实例最接近的两个最可能包含目标的区域。
  7. 如权利要求5所述的医学图像样本筛查方法，其中，通过计算所述病灶目标实例与所述两个病灶目标区域重叠的交集与并集的比值，用于检测医学图像中目标实例的抗扰 动稳定度指标IOU。
  8. 如权利要求5所述的医学图像样本筛查方法，其中，根据所述病灶目标置信度和病灶目标实例抗扰动稳定度，结合主动学习算法计算出样本标注价值包括如下步骤：
    选取病灶目标置信度值为0.4～0.7之间的医学图像样本，表达为max Unc(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，Unc表示为置信度；
    选取抗扰动稳定度值为0.3～0.9之间的医学图像样本，表达为IOU(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，IOU表示为抗扰动稳定度值；
    通过主动学习算法公式max f(x,L,u)＝max Unc(x,L,u)*IOU(x,L,u)β修正计算出每个医学图像样本的标注价值max f(x,L,u)，其中，β是调解所述抗扰动稳定度值比重的参数。
  9. 如权利要求1所述的医学图像样本筛查方法，其中，所述选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对当前的病灶目标检测深度模型C进行迭代更新包括如下步骤：
    选取所述未标注医学图像样本集U中标注价值高的医学图像样本，并进行标注确认；
    将标注确认后的样本放入到训练样本集中；
    通过主动学习算法对病灶目标检测深度模型C进行更新。
  10. 一种医学图像样本筛查装置，其中，所述装置包括病灶目标深度模型初始化单元、未标注医学图像样本标注价值预测单元、病灶目标检测深度模型迭代单元和模型迭代更新判断单元；
    所述病灶目标检测深度模型初始化单元，用于选取当前医疗影像目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
    所述未标注医学图像样本标注价值预测单元，用于根据所述病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
    所述病灶目标检测深度模型迭代单元，用于选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认，根据选取的标注价值高的医学图像样本对当前的病灶目标检测深度模型C进行迭代更新；
    所述模型迭代更新判断单元，用于根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，如病灶目标检测深度模型C的性能不能再继续标注新的样本时结束迭代更新，否则所述病灶目标检测深度模型迭代单元继续根据新的样本进行迭代更新。
  11. 一种计算机设备，包括存储器和处理器，所述存储器中存储有计算机可读指令，所述计算机可读指令被所述处理器执行时，使得所述处理器执行如下步骤：
    选取针对当前医疗影像病灶目标检测任务的初始已标注样本集L，利用Mask-RCNN模型对所述初始已标注样本集L进行模型训练，以获取当前医疗影像病灶目标检测的病灶目标检测深度模型C；
    根据所述初始病灶目标检测深度模型C对未标注医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值；
    选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对当前的病灶目标检测深度模型C进行迭代更新；
    根据迭代更新后的病灶目标检测深度模型C对新的医学图像病灶样本进行验证，直 到病灶目标检测深度模型C的性能不能再继续标注新的样本时，则结束迭代更新。
  12. 如权利要求11所述的计算机设备，其中，根据所述初始病灶目标检测深度模型C对未标注的医学图像样本集U进行逐一预测，得到所述未标注医学图像样本集U中每个医学图像样本的预测结果，并根据预测结果判断每个医学图像样本的标注价值包括如下步骤：
    计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值；
    计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值；
    根据所述病灶目标置信度值和病灶目标实例抗扰动稳定度值，结合主动学习算法计算出样本标注价值。
  13. 如权利要求12所述的计算机设备，其中，所述计算未标注的医学图像样本集U中每个医学图像样本的病灶目标置信度值包括如下步骤：
    根据所述病灶目标检测深度模型C，计算未标注的医学图像样本集U中每个医学图像样本中的病灶目标的目标检测框数量；
    根据每个医学图像样本中的目标检测框数量计算出病灶目标置信度值。
  14. 如权利要求13所述的计算机设备，其中，所述未标注医学图像样本中待检测目标的不确定性，通过所述病灶目标在医学图像样本中预测的数量，
    其中，病灶目标在医学图像样本中预测的数量越多，表示样本的不确定性越高。
  15. 如权利要求13所述的计算机设备，其中，所述计算未标注的医学图像样本集U中每个医学图像样本的病灶目标实例抗扰动稳定度值包括如下步骤：
    根据所述病灶目标检测深度模型C，采用Mask-RCNN模型中的RPN网络来生成未标注的医学图像样本集U中每个医学图像样本中包含病灶目标的区域；
    选取与病灶目标实例最接近的两个病灶目标区域；
    计算病灶目标实例与两个病灶目标区重叠的交集与并集的比值，作为病灶目标实例的抗扰动稳定度值。
  16. 如权利要求15所述的计算机设备，其中，采用RPN网络生成医学图像样本中可能包含物体的区域候选框后，保留前N个最可能包含目标的区域，其中，与病灶目标实例最接近的两个最可能包含目标的区域。
  17. 如权利要求15所述的计算机设备，其中，通过计算所述病灶目标实例与所述两个病灶目标区域重叠的交集与并集的比值，用于检测医学图像中目标实例的抗扰动稳定度指标IOU。
  18. 如权利要求15所述的计算机设备，其中，根据所述病灶目标置信度和病灶目标实例抗扰动稳定度，结合主动学习算法计算出样本标注价值包括如下步骤：
    选取病灶目标置信度值为0.4～0.7之间的医学图像样本，表达为max Unc(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，Unc表示为置信度；
    选取抗扰动稳定度值为0.3～0.9之间的医学图像样本，表达为IOU(x,L,u)，其中，L表示为已标注样本，u表示为未标注样本，x为选取的医学图像样本，IOU表示为抗扰动稳定度值；
    通过主动学习算法公式max f(x,L,u)＝max Unc(x,L,u)*IOU(x,L,u)β修正计算出每个医学图像样本的标注价值max f(x,L,u)，其中，β是调解所述抗扰动稳定度值比重的参数。
  19. 如权利要求11所述的计算机设备，其中，所述选取所述未标注医学图像样本集U中标注价值高的医学图像样本进行标注确认后，对当前的病灶目标检测深度模型C进行迭代更新包括如下步骤：
    选取所述未标注医学图像样本集U中标注价值高的医学图像样本，并进行标注确认；
    将标注确认后的样本放入到训练样本集中；
    通过主动学习算法对病灶目标检测深度模型C进行更新。
  20. 一种存储有计算机可读指令的存储介质，所述计算机可读指令被一个或多个处理器执行时，使得一个或多个处理器执行如权利要求1至9中任一项权利要求所述医学图像样本筛查方法的步骤。
PCT/CN2020/099328 2020-05-28 2020-06-30 医学图像样本筛查方法、装置、计算机设备和存储介质 WO2021114630A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010468690.9 2020-05-28
CN202010468690.9A CN111666993A (zh) 2020-05-28 2020-05-28 医学图像样本筛查方法、装置、计算机设备和存储介质

Publications (1)

Publication Number Publication Date
WO2021114630A1 true WO2021114630A1 (zh) 2021-06-17

Family

ID=72384924

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/099328 WO2021114630A1 (zh) 2020-05-28 2020-06-30 医学图像样本筛查方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN111666993A (zh)
WO (1) WO2021114630A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113487617A (zh) * 2021-07-26 2021-10-08 推想医疗科技股份有限公司 数据处理方法、装置、电子设备以及存储介质
CN113782221A (zh) * 2021-09-16 2021-12-10 平安科技（深圳）有限公司 基于自训练学习的疾病预测装置、设备及存储介质
CN113793301A (zh) * 2021-08-19 2021-12-14 首都医科大学附属北京同仁医院 基于稠密卷积网络模型的眼底影像分析模型的训练方法
CN114271763A (zh) * 2021-12-20 2022-04-05 合肥中纳医学仪器有限公司 一种基于Mask RCNN的胃癌早期识别方法、系统、装置
CN115393323A (zh) * 2022-08-26 2022-11-25 数坤(上海)医疗科技有限公司 一种目标区域获得方法、装置、设备和存储介质
CN116797889A (zh) * 2023-08-24 2023-09-22 青岛美迪康数字工程有限公司 医学影像识别模型的更新方法、装置和计算机设备
CN117115568A (zh) * 2023-10-24 2023-11-24 浙江啄云智能科技有限公司 数据筛选方法、装置、设备及存储介质
CN117152138A (zh) * 2023-10-30 2023-12-01 陕西惠宾电子科技有限公司 一种基于无监督学习的医学图像肿瘤目标检测方法
CN117876910A (zh) * 2024-03-06 2024-04-12 西北工业大学 基于主动学习的无人机目标检测关键数据筛选方法

Families Citing this family (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112163634B (zh) * 2020-10-14 2023-09-05 平安科技（深圳）有限公司 实例分割模型样本筛选方法、装置、计算机设备及介质
CN112614570B (zh) * 2020-12-16 2022-11-25 上海壁仞智能科技有限公司 样本集标注、病理图像分类、分类模型构建方法及装置
CN113378767A (zh) * 2021-06-28 2021-09-10 浙江新再灵科技股份有限公司 一种基于自适应学习的电梯预警控制方法及其系统
CN113434718A (zh) * 2021-06-29 2021-09-24 联仁健康医疗大数据科技股份有限公司 确定关联图像的方法、装置、电子设备及存储介质
CN113435522A (zh) * 2021-06-30 2021-09-24 平安科技（深圳）有限公司 图像分类方法、装置、设备及存储介质
CN113469981B (zh) * 2021-07-09 2022-07-08 数坤(北京)网络科技股份有限公司 图像处理方法、装置及存储介质
CN113435540A (zh) * 2021-07-22 2021-09-24 中国人民大学 一种类分布不匹配时的图像分类方法、系统、介质和设备
CN113723467A (zh) * 2021-08-05 2021-11-30 武汉精创电子技术有限公司 用于缺陷检测的样本收集方法、装置和设备
CN114334167A (zh) * 2021-12-31 2022-04-12 医渡云（北京）技术有限公司 医学数据挖掘方法及装置、存储介质、电子设备
CN114758197B (zh) * 2022-06-15 2022-11-01 深圳瀚维智能医疗科技有限公司 数据筛选方法、装置及计算机可读存储介质
CN115631386B (zh) * 2022-12-19 2023-03-10 天津医之本医疗科技有限公司 一种基于机器学习的病理图像分类方法和系统
CN116935388B (zh) * 2023-09-18 2023-11-21 四川大学 一种皮肤痤疮图像辅助标注方法与系统、分级方法与系统

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107563123A (zh) * 2017-09-27 2018-01-09 百度在线网络技术（北京）有限公司 用于标注医学图像的方法和装置
US10140709B2 (en) * 2017-02-27 2018-11-27 International Business Machines Corporation Automatic detection and semantic description of lesions using a convolutional neural network
CN108922599A (zh) * 2018-06-27 2018-11-30 西南交通大学 一种基于mil的医学图像病灶点精确标注方法
CN110097559A (zh) * 2019-04-29 2019-08-06 南京星程智能科技有限公司 基于深度学习的眼底图像病灶区域标注方法
CN110599451A (zh) * 2019-08-05 2019-12-20 平安科技（深圳）有限公司 医学图像病灶检测定位方法、装置、设备及存储介质
CN111028224A (zh) * 2019-12-12 2020-04-17 广西医准智能科技有限公司 数据标注、模型训练和图像处理方法、装置及存储介质
CN111128373A (zh) * 2019-12-11 2020-05-08 上海交通大学 一种通过协同主动学习模型的乳腺癌数据标注方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108764281A (zh) * 2018-04-18 2018-11-06 华南理工大学 一种基于半监督自步学习跨任务深度网络的图像分类方法
CN109492686A (zh) * 2018-11-01 2019-03-19 郑州云海信息技术有限公司 一种图片标注方法与系统
CN109784391B (zh) * 2019-01-04 2021-01-05 杭州比智科技有限公司 基于多模型的样本标注方法及装置
CN111160469B (zh) * 2019-12-30 2023-08-08 湖南大学 一种目标检测系统的主动学习方法
CN111191732B (zh) * 2020-01-03 2021-05-14 天津大学 一种基于全自动学习的目标检测方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10140709B2 (en) * 2017-02-27 2018-11-27 International Business Machines Corporation Automatic detection and semantic description of lesions using a convolutional neural network
CN107563123A (zh) * 2017-09-27 2018-01-09 百度在线网络技术（北京）有限公司 用于标注医学图像的方法和装置
CN108922599A (zh) * 2018-06-27 2018-11-30 西南交通大学 一种基于mil的医学图像病灶点精确标注方法
CN110097559A (zh) * 2019-04-29 2019-08-06 南京星程智能科技有限公司 基于深度学习的眼底图像病灶区域标注方法
CN110599451A (zh) * 2019-08-05 2019-12-20 平安科技（深圳）有限公司 医学图像病灶检测定位方法、装置、设备及存储介质
CN111128373A (zh) * 2019-12-11 2020-05-08 上海交通大学 一种通过协同主动学习模型的乳腺癌数据标注方法
CN111028224A (zh) * 2019-12-12 2020-04-17 广西医准智能科技有限公司 数据标注、模型训练和图像处理方法、装置及存储介质

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113487617A (zh) * 2021-07-26 2021-10-08 推想医疗科技股份有限公司 数据处理方法、装置、电子设备以及存储介质
CN113793301A (zh) * 2021-08-19 2021-12-14 首都医科大学附属北京同仁医院 基于稠密卷积网络模型的眼底影像分析模型的训练方法
CN113793301B (zh) * 2021-08-19 2023-07-21 首都医科大学附属北京同仁医院 基于稠密卷积网络模型的眼底影像分析模型的训练方法
CN113782221A (zh) * 2021-09-16 2021-12-10 平安科技（深圳）有限公司 基于自训练学习的疾病预测装置、设备及存储介质
CN114271763B (zh) * 2021-12-20 2024-05-28 合肥中纳医学仪器有限公司 一种基于Mask RCNN的胃癌早期识别方法、系统、装置
CN114271763A (zh) * 2021-12-20 2022-04-05 合肥中纳医学仪器有限公司 一种基于Mask RCNN的胃癌早期识别方法、系统、装置
CN115393323A (zh) * 2022-08-26 2022-11-25 数坤(上海)医疗科技有限公司 一种目标区域获得方法、装置、设备和存储介质
CN116797889A (zh) * 2023-08-24 2023-09-22 青岛美迪康数字工程有限公司 医学影像识别模型的更新方法、装置和计算机设备
CN116797889B (zh) * 2023-08-24 2023-12-08 青岛美迪康数字工程有限公司 医学影像识别模型的更新方法、装置和计算机设备
CN117115568B (zh) * 2023-10-24 2024-01-16 浙江啄云智能科技有限公司 数据筛选方法、装置、设备及存储介质
CN117115568A (zh) * 2023-10-24 2023-11-24 浙江啄云智能科技有限公司 数据筛选方法、装置、设备及存储介质
CN117152138A (zh) * 2023-10-30 2023-12-01 陕西惠宾电子科技有限公司 一种基于无监督学习的医学图像肿瘤目标检测方法
CN117152138B (zh) * 2023-10-30 2024-01-16 陕西惠宾电子科技有限公司 一种基于无监督学习的医学图像肿瘤目标检测方法
CN117876910A (zh) * 2024-03-06 2024-04-12 西北工业大学 基于主动学习的无人机目标检测关键数据筛选方法

Also Published As

Publication number Publication date
CN111666993A (zh) 2020-09-15

Similar Documents

Publication Publication Date Title
WO2021114630A1 (zh) 医学图像样本筛查方法、装置、计算机设备和存储介质
JP7279015B2 (ja) マンモグラフィにおける密度の評価
KR101818074B1 (ko) 인공지능 기반 의료용 자동 진단 보조 방법 및 그 시스템
EP3982292A1 (en) Method for training image recognition model, and method and apparatus for image recognition
KR101908680B1 (ko) 약한 지도 학습 기반의 기계 학습 방법 및 그 장치
CN112102266B (zh) 基于注意力机制的脑梗死医学影像分类模型的训练方法
CN114945951A (zh) 用于图像分类的不确定性引导的半监督神经网络训练
KR102289277B1 (ko) 복수의 의료 영상 판독 알고리듬들에 대한 평가 스코어를 생성하는 의료 영상 판독 지원 장치 및 방법
JP2021535484A (ja) 自動的な腫瘍検出及び分類のためのシステム
US11915822B2 (en) Medical image reading assistant apparatus and method for adjusting threshold of diagnostic assistant information based on follow-up examination
CN110046707B (zh) 一种神经网络模型的评估优化方法和系统
Fotaki et al. Artificial intelligence in cardiac MRI: is clinical adoption forthcoming?
CN111583199A (zh) 样本图像标注方法、装置、计算机设备和存储介质
CN112884060A (zh) 图像标注方法、装置、电子设备及存储介质
CN114266896A (zh) 图像标注方法、模型训练方法及装置、电子设备、介质
CN111340794A (zh) 冠状动脉狭窄的量化方法及装置
US11682135B2 (en) Systems and methods for detecting and correcting orientation of a medical image
CN113705595A (zh) 异常细胞转移程度的预测方法、装置和存储介质
CN115760831A (zh) 一种图像处理模型的训练方法和系统
US11263481B1 (en) Automated contrast phase based medical image selection/exclusion
CN112634255B (zh) 脑部病灶检测模型的建立方法、装置以及计算机设备
KR102360615B1 (ko) 내시경 영상에 대한 복수의 의료 영상 판독 알고리듬들을 이용하는 의료 영상 판독 지원 장치 및 방법
CN113222989A (zh) 一种图像分级方法、装置、存储介质及电子设备
US20240184854A1 (en) Method and apparatus for training image recognition model, and image recognition method and apparatus
Remy-Jardin et al. Artificial Intelligence-Based Detection of Pulmonary Vascular Disease

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: 20898107

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: 20898107

Country of ref document: EP

Kind code of ref document: A1