Image segmentation method, device, computer equipment, storage media and program product

Abstract

The application discloses an image segmentation method and device, a computer device, a storage medium and a program product, and relates to the technical field of image processing. The method comprises the following steps: determining an under-segmentation mask and an over-segmentation mask corresponding to a sample medical image based on sample diameter annotations corresponding to sample segmentation regions in the sample medical image; training a first image segmentation network based on the sample medical image and the under-segmentation mask, the first image segmentation network being used for under-segmentation region prediction of the sample medical image; and training a second image segmentation network based on the sample medical image and the over-segmentation mask, the second image segmentation network being used for over-segmentation region prediction of the sample medical image. Weak supervision information is generated by using sample diameter annotations, which can reduce manual annotation operations, so that large-scale training sample data can be obtained through less manual annotation, the efficiency of generating supervision information used for training an image segmentation network is improved, and the model training efficiency is further improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image segmentation method, apparatus, computer equipment, storage medium, and program product. Background Technology

[0002] Medical image segmentation is used to segment out parts of medical images that have certain special meanings and extract relevant features, providing a reliable basis for clinical medical and pathological research.

[0003] Medical images may contain various types of images to be segmented, such as organs and tissues in the human body. Computer devices segment medical images by calling relevant image segmentation networks. To implement the image segmentation function of the image segmentation network, it needs to be trained on medical images with labeled segmentation regions. In related technologies, medical experts manually annotate the contours of the segmentation regions to obtain sample images with labeled regions.

[0004] Obviously, manually labeled sample images require annotation by professional medical personnel, making it difficult to obtain sample images, which leads to low training efficiency of image segmentation networks. Summary of the Invention

[0005] This application provides an image segmentation method, apparatus, computer device, storage medium, and program product. The technical solution includes the following aspects.

[0006] On the one hand, an image segmentation method is provided, the method comprising:

[0007] Under-segmentation region prediction is performed on the target segmentation region in the target medical image to obtain a first target probability image. The first target probability image contains the probability value of each pixel in the target medical image belonging to the under-segmentation region, and the under-segmentation region is smaller than the target segmentation region.

[0008] Oversegmentation region prediction is performed on the target segmentation region in the target medical image to obtain a second target probability image. The second target probability image contains the probability that each pixel in the target medical image belongs to an oversegmentation region, and the oversegmentation region is larger than the target segmentation region.

[0009] Based on the first target probability image and the second target probability image, the target segmentation mask corresponding to the target medical image is determined.

[0010] On the other hand, an image segmentation method is provided, the method comprising:

[0011] Based on the sample diameter annotation corresponding to the sample segmentation region in the sample medical image, the undersegmentation mask and oversegmentation mask corresponding to the sample medical image are determined. The undersegmentation mask corresponds to an undersegmentation region smaller than the sample segmentation region, and the oversegmentation mask corresponds to an oversegmentation region larger than the sample segmentation region.

[0012] Based on the sample medical image and the undersegmentation mask, a first image segmentation network is trained, which is used to predict the undersegmentation region of the sample medical image.

[0013] Based on the sample medical image and the oversegmentation mask, a second image segmentation network is trained. The second image segmentation network is used to predict oversegmentation regions in the sample medical image.

[0014] On the other hand, an image segmentation apparatus is provided, the apparatus comprising:

[0015] The first segmentation prediction module is used to predict under-segmented regions in the target segmentation region of the target medical image to obtain a first target probability image. The first target probability image contains the probability value of each pixel in the target medical image belonging to the under-segmented region, and the under-segmented region is smaller than the target segmentation region.

[0016] The second segmentation prediction module is used to predict oversegmentation regions in the target segmentation region of the target medical image to obtain a second target probability image. The second target probability image contains the probability that each pixel in the target medical image belongs to an oversegmentation region, and the oversegmentation region is larger than the target segmentation region.

[0017] The first determining module is used to determine the target segmentation mask corresponding to the target medical image based on the first target probability image and the second target probability image.

[0018] On the other hand, an image segmentation apparatus is provided, the apparatus comprising:

[0019] The second determining module is used to determine the undersegmentation mask and oversegmentation mask corresponding to the sample medical image based on the sample diameter annotation corresponding to the sample segmentation region in the sample medical image, wherein the undersegmentation region corresponding to the undersegmentation mask is smaller than the sample segmentation region, and the oversegmentation region corresponding to the oversegmentation mask is larger than the sample segmentation region.

[0020] The first training module is used to train a first image segmentation network based on the sample medical image and the undersegmentation mask. The first image segmentation network is used to predict the undersegmentation region of the sample medical image.

[0021] The second training module is used to train a second image segmentation network based on the sample medical image and the oversegmentation mask. The second image segmentation network is used to predict oversegmentation regions in the sample medical image.

[0022] On the other hand, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one program, the at least one program being loaded and executed by the processor to implement the image segmentation method as described above.

[0023] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set, or instruction set is stored therein, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the image segmentation method as described above.

[0024] According to another aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the image segmentation method provided in the above-described optional implementation.

[0025] The beneficial effects of the technical solutions provided in this application include at least the following:

[0026] This paper provides a weakly supervised image segmentation method. By using the sample diameter annotations corresponding to the segmentation regions in sample medical images, it constructs supervision information for training a first image segmentation network and a second image segmentation network: undersegmentation masks and oversegmentation masks. This enables the first image segmentation network to predict undersegmented regions and the second image segmentation network to predict oversegmented regions, thus allowing the trained first and second image segmentation networks to be used for segmentation prediction of segmentation regions in medical images. Compared with related technologies that require the use of sample segmentation contour annotations of sample segmentation regions as supervision information, this embodiment uses sample diameter annotations to generate weakly supervised information, which can reduce manual annotation operations. This allows for the acquisition of large-scale training sample data with less manual annotation, improving the efficiency of generating supervision information for training image segmentation networks and further improving model training efficiency. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 A schematic diagram of an implementation environment provided by an exemplary embodiment of this application is shown;

[0029] Figure 2 A flowchart illustrating an exemplary embodiment of the image segmentation method provided in this application is shown;

[0030] Figure 3 A flowchart of an image segmentation method provided by another exemplary embodiment of this application is shown;

[0031] Figure 4 The illustration shows a schematic diagram illustrating the generation process of under-segmented and over-segmented masks in an exemplary embodiment of this application;

[0032] Figure 5 This illustration shows a schematic diagram of the training process of an image segmentation model according to an exemplary embodiment of this application;

[0033] Figure 6 A flowchart of an image segmentation method provided by another exemplary embodiment of this application is shown;

[0034] Figure 7 This illustration shows a schematic diagram of the training process of an image segmentation model as illustrated in another exemplary embodiment of this application;

[0035] Figure 8 This illustration shows a schematic diagram of the training process of an image segmentation model as illustrated in another exemplary embodiment of this application;

[0036] Figure 9 A flowchart illustrating an exemplary embodiment of the image segmentation method provided in this application is shown;

[0037] Figure 10 A flowchart of an image segmentation method provided by another exemplary embodiment of this application is shown;

[0038] Figure 11 A schematic diagram illustrating the principle of an image segmentation process according to an exemplary embodiment of this application is shown;

[0039] Figure 12 This is a schematic diagram illustrating the test results of the effect of the value of λ on the segmentation results on the KiTS19 dataset.

[0040] Figure 13This is a schematic diagram illustrating the test results of the effect of the value of λ on the segmentation results on the LiTS17 dataset.

[0041] Figure 14 This is a structural block diagram of an image segmentation apparatus provided in an exemplary embodiment of this application;

[0042] Figure 15 This is a structural block diagram of an image segmentation apparatus provided in another exemplary embodiment of this application;

[0043] Figure 16 A schematic diagram of the structure of a computer device provided in one embodiment of this application is shown. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0045] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0046] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0047] Computer vision (CV) is a science that studies how to enable machines to "see." More specifically, it refers to using cameras and computers to identify and measure targets instead of the human eye, and then performing image processing to create images more suitable for human observation or transmission to instruments for detection. As a scientific discipline, computer vision studies related theories and technologies, attempting to build artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include image processing, image recognition, image semantic understanding, image retrieval, optical character recognition (OCR), video processing, video semantic understanding, video content / behavior recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), and common biometric recognition technologies such as facial recognition and fingerprint recognition.

[0048] The method described in this application involves using computer vision technology to segment specific regions in medical images. For example, using the solution provided in this application, lesion areas in the liver can be segmented based on computed tomography (CT) images of the liver.

[0049] Figure 1 A schematic diagram of an implementation environment provided by an exemplary embodiment of this application is shown. This implementation environment includes a terminal 110 and a server 120. The terminal 110 and the server 120 communicate via a communication network. Optionally, the communication network can be a wired network or a wireless network, and the communication network can be at least one of a local area network (LAN), a metropolitan area network (MAN), and a wide area network (WAN).

[0050] Terminal 110 is an electronic device with medical image segmentation requirements. This electronic device can be a smartphone, tablet computer, or personal computer, etc., and this embodiment does not limit it. Figure 1 The following example uses Terminal 110, a computer used by medical staff, for illustration.

[0051] In some embodiments, the terminal 110 is equipped with an application that has medical image segmentation capabilities. When it is necessary to segment a scanned medical image, the user inputs the medical image to be segmented into the application, thereby uploading the medical image to the server 120, which then segments the medical image and returns the segmentation result.

[0052] Server 120 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0053] In some embodiments, server 120 is used to provide medical image segmentation services for applications installed on terminal 110. Optionally, server 120 is configured with an image segmentation model, which is a Siamese network structure, including a first image segmentation network 121 and a second image segmentation network 122. The first image segmentation network 121 predicts undersegmented regions of the medical image to obtain undersegmented region prediction results; at the same time, the second image segmentation network 122 predicts oversegmented regions of the medical image to obtain oversegmented region prediction results. Then, based on the undersegmented region prediction results and the oversegmented region prediction results, the image segmentation result corresponding to the medical image is generated.

[0054] Of course, in other possible implementations, the image segmentation model can also be deployed on the terminal 110 side, and the terminal 110 can perform image segmentation locally without the need for the server 120. This embodiment does not limit this.

[0055] Optionally, in other possible application scenarios, this application embodiment also provides a training method for an image segmentation model. The corresponding terminal 110 can upload training sample data to the server 120, and the server 120 can train the image segmentation model based on the training sample data and feed back the trained image segmentation model to the terminal 110.

[0056] For ease of explanation, the following embodiments are illustrated using a medical image segmentation method executed by a computer device as an example.

[0057] Please refer to Figure 2 This document illustrates a flowchart of an image segmentation method provided in an exemplary embodiment of this application. This embodiment uses a computer device as an example to illustrate the method, which includes:

[0058] Step 201: Based on the sample diameter annotation corresponding to the sample segmentation region in the sample medical image, determine the undersegmentation mask and oversegmentation mask corresponding to the sample medical image.

[0059] In related technologies, the training samples used when training image segmentation networks are: sample medical images and the corresponding sample region contour annotations in the sample medical images. The sample region contour annotations are used as supervision information in the training process of image segmentation networks. However, the sample region contour annotations often require manual region drawing. For the sample segmentation regions in medical images, they are often convex polygon contours. If a large number of training samples are to be obtained, the amount of manual annotation is obviously large, resulting in low efficiency in obtaining training samples.

[0060] To address the challenge of obtaining training samples during image segmentation network training in related technologies, this application provides a weakly supervised training method. It uses sample medical images, undersegmentation masks, and oversegmentation masks as training datasets, and employs the undersegmentation and oversegmentation masks as weak supervision information during image segmentation network training. Since both undersegmentation and oversegmentation masks can be determined by the diameter annotations corresponding to the segmented regions in the sample medical images, and the diameter annotation only requires manual annotation of the diameter position of the segmented region, this embodiment can obtain a larger set of usable training samples with minimal manual annotation compared to annotating the complete region contour information of the segmented region.

[0061] Taking CT images as an example, the sample diameter annotation is the RECIST annotation. In clinical medicine, CT uses the Response Evaluation Criteria in Solid Tumors (RECIST) to measure tumors or lymph nodes. The image archives and communication systems of corresponding hospitals store a large number of RECIST diameter measurements and corresponding CT images. The RECIST diameter measurements generally include the major and minor diameters, which is a common way for radiologists to highlight target areas. Therefore, in one possible implementation, it is only necessary to obtain the CT image and its corresponding RECIST diameter measurement to generate the undersegmentation mask and oversegmentation mask for each CT image, without the need for other manual annotation processes.

[0062] In an exemplary example, the relationship between the undersegmentation mask, the oversegmentation mask, and the sample segmentation region is as follows:

[0063]

[0064] in, This indicates the under-segmented region in the sample medical image segmentation area. This represents the segmented region of a sample medical image. This represents the oversegmented region of the sample segmentation area in the sample medical image. As can be seen from formula (1), the undersegmented region corresponding to the undersegmentation mask is smaller than the sample segmentation area, and the oversegmented region corresponding to the oversegmentation mask is larger than the sample segmentation area.

[0065] In one possible implementation, a sample medical image and its corresponding sample diameter annotation are acquired. The under-segmented and over-segmented regions corresponding to the sample segmentation regions in the sample medical image are determined by the sample diameter annotation, and then the corresponding under-segmented and over-segmented masks are generated.

[0066] Step 202: Based on the sample medical image and the undersegmentation mask, train a first image segmentation network. The first image segmentation network is used to predict the undersegmentation region of the sample medical image.

[0067] Since undersegmentation and oversegmentation masks are used as weak supervision information, Siamese networks are required to perform undersegmentation and oversegmentation prediction, respectively. In one possible implementation, a first image segmentation network is provided. The first image segmentation network performs undersegmentation prediction on the sample medical image to obtain the undersegmentation prediction result. In order for the undersegmentation prediction result to learn from the undersegmentation mask, the first image segmentation network needs to be trained using the undersegmentation mask and the undersegmentation prediction result. That is, the training dataset corresponding to the first image segmentation network is: the sample medical image and the undersegmentation mask, so as to train the first image segmentation network with the sample medical image and the undersegmentation mask, so that the first image segmentation network can be used to predict undersegmentation regions of the sample medical image.

[0068] Optionally, the specific network structure corresponding to the first image segmentation network can be any network structure with image segmentation function, such as Hopfield Neural Network (HNN), U-Net, ARUNet, SwinTransformer, etc. The specific network structure adopted by the first image segmentation network in this application embodiment does not constitute a limitation.

[0069] Step 203: Based on the sample medical image and the oversegmentation mask, train a second image segmentation network. The second image segmentation network is used to predict the oversegmentation region of the sample medical image.

[0070] Corresponding to the first image segmentation network set up for undersegmented regions, in one possible implementation, a second image segmentation network is also set up for predicting oversegmented regions. The second image segmentation network is used to predict oversegmented regions of the sample medical image to obtain the oversegmentation prediction result. In order for the oversegmentation prediction result to learn from the oversegmentation mask, the oversegmentation mask and the oversegmentation prediction result are used to train the second image segmentation network. That is, the training dataset of the second image segmentation network is: the sample medical image and the oversegmentation mask, so as to train the second image segmentation network with the sample medical image and the oversegmentation mask, so that the second image segmentation network can be used to predict oversegmented regions of the sample medical image.

[0071] Optionally, the specific network structure corresponding to the second image segmentation network can be any network structure with image segmentation function, such as HNN, U-Net, ARUNet, Swin Transformer, etc. The embodiments of this application do not limit the specific network structure adopted by the second image segmentation network.

[0072] Optionally, after the first image segmentation network has the function of predicting under-segmented regions in the medical image and the second image segmentation network has the function of predicting over-segmented regions in the medical image, the segmentation result corresponding to the medical image can be obtained through model integration based on the probability map predicted by the first image segmentation network and the probability map predicted by the second image segmentation network during the model application process. The specific model application process can be referred to in the following embodiment, which will not be elaborated here.

[0073] It should be noted that since the first image segmentation network and the second image segmentation network are twin network structures, the first image segmentation network can have the same neural network structure as the second image segmentation network. For example, both the first image segmentation network and the second image segmentation network can use the U-net network. Alternatively, the first image segmentation network can have different neural network structures than the second image segmentation network. For example, the first image segmentation network can use the U-net network, while the second image segmentation network can use an HNN.

[0074] In this embodiment of the application, the first image segmentation network and the second image segmentation network are trained simultaneously. That is, in each round of training, the same sample medical images are used as training samples.

[0075] In summary, this application provides a weakly supervised image segmentation method. By using the sample diameter annotations corresponding to the segmentation regions in a sample medical image, supervision information for training a first image segmentation network and a second image segmentation network is constructed: undersegmentation masks and oversegmentation masks. This enables the first image segmentation network to predict undersegmented regions and the second image segmentation network to predict oversegmented regions, thus allowing the trained first and second image segmentation networks to be used for segmentation prediction of segmentation regions in medical images. Compared to related technologies that require the use of sample segmentation contour annotations of the segmentation regions as supervision information, this application uses sample diameter annotations to generate weakly supervised information, which reduces manual annotation operations. This allows for the acquisition of large-scale training sample data with less manual annotation, improving the efficiency of generating supervision information for training the image segmentation network and further improving model training efficiency.

[0076] In the process of generating undersegmented and oversegmented masks with weak supervision information, it is first necessary to determine the undersegmented and oversegmented regions based on the sample diameter annotation, and then obtain the corresponding mask images through binarization processing.

[0077] Please refer to Figure 3 This illustration shows a flowchart of an image segmentation method provided in another exemplary embodiment of this application. This embodiment is executed by an example where a computer device is the execution subject of the method, which includes:

[0078] Step 301: Based on the sample diameter annotation, determine the major and minor axis annotations corresponding to the sample segmentation regions in the sample medical image.

[0079] In one possible implementation, the sample diameter annotation includes a major diameter annotation and a minor diameter annotation, wherein the major diameter annotation is the annotation of the longest diameter corresponding to the sample segmentation region, and the minor diameter annotation is the annotation of the shortest diameter corresponding to the sample segmentation region.

[0080] like Figure 4 The diagram illustrates the generation process of undersegmentation and oversegmentation masks according to an exemplary embodiment of this application. The sample medical image 400 includes a sample segmentation region 401, the longest diameter of which is shown as dashed line 402, and the shortest diameter as shown as dashed line 403.

[0081] Step 302: Based on the major and minor axis annotations, determine the undersegmentation mask corresponding to the sample medical image.

[0082] In one possible implementation, the undersegmented region indicated by the undersegmentation mask is a quadrilateral region. The four vertices of the quadrilateral region are determined by the major axis label and the minor axis label. That is, the diagonals of the quadrilateral region are the longest diameter indicated by the major axis label and the shortest diameter indicated by the minor axis label, respectively.

[0083] In an exemplary example, step 302 may include steps 302A and 302B.

[0084] Step 302A: Based on the major and minor axis annotations, determine the under-segmented regions in the sample medical image.

[0085] In one possible implementation, if it is necessary to generate the corresponding undersegmentation mask, the undersegmentation region corresponding to the sample medical image is first determined according to the longest diameter indicated by the major axis label and the shortest diameter indicated by the minor axis label. Then, the sample medical image is binarized based on the undersegmentation region to obtain the undersegmentation mask corresponding to the sample medical image.

[0086] Step 302B: Set the pixel values ​​of pixels belonging to the undersegmented region in the medical sample image to the first pixel value, and set the pixel values ​​of pixels not belonging to the undersegmented region in the medical sample image to the second pixel value, thereby generating an undersegmentation mask corresponding to the sample medical image.

[0087] Since the undersegmentation mask can clearly indicate the undersegmented region and other regions in the sample medical image, in one possible implementation, the pixel value of the pixel belonging to the undersegmented region is set as the first pixel value, and the pixel value of the pixel not belonging to the undersegmented region is set as the second pixel value, so that the undersegmented region in the sample medical image can be directly determined based on the difference in pixel values.

[0088] Taking a first pixel value of 1 and a second pixel value of 0 as an example, the pixel values ​​of the pixels located in the undersegmented region of the sample medical image are all 1, while the pixel values ​​of the pixels located outside the undersegmented region of the sample medical image are 0.

[0089] like Figure 4 As shown, the undersegmented region 404 is determined by using the longest diameter 402 and the shortest diameter 403 as diagonals. The sample medical image 400 is then binarized based on the undersegmented region 404 to obtain the undersegmented mask 405.

[0090] Step 303: Based on the major axis annotation, determine the oversegmentation mask corresponding to the sample medical image.

[0091] In one possible implementation, the oversegmentation region indicated by the oversegmentation mask is a circular region. If the circular region is to be larger than the sample segmentation region, the longest diameter indicated by the major axis label is used as the diameter to determine the oversegmentation region in the sample medical image, and then the corresponding oversegmentation mask is generated.

[0092] It should be noted that, in actual execution, the execution order of steps 302 and 303 is not limited. Step 302 can be executed first, followed by step 303; or step 303 can be executed first, followed by step 302; or steps 302 and 303 can be executed simultaneously.

[0093] In an exemplary example, step 303 may include steps 303A and 303B.

[0094] Step 303A: Based on the major axis annotation, identify the oversegmented regions in the sample medical image.

[0095] In one possible implementation, once the position of the longest diameter indicated by the major axis label in the sample medical image is determined, the oversegmented region corresponding to the sample medical image can be determined using the longest diameter as the diameter. Then, the sample medical image is binarized based on the oversegmented region to generate the oversegmented mask corresponding to the sample medical image.

[0096] Step 303B: Set the pixel values ​​of pixels belonging to the oversegmented region in the medical sample image to the first pixel value, and set the pixel values ​​of pixels not belonging to the oversegmented region in the medical sample image to the second pixel value, thereby generating an oversegmentation mask corresponding to the sample medical image.

[0097] Since the oversegmentation mask can clearly indicate oversegmented regions and other regions in a sample medical image, in one possible implementation, the pixel value of the pixel belonging to the oversegmentation region is set as the first pixel value, and the pixel value of the pixel not belonging to the oversegmentation region is set as the second pixel value, so that the oversegmentation region in the sample medical image can be directly determined based on the difference in pixel values.

[0098] Taking a first pixel value of 1 and a second pixel value of 0 as an example, the pixel values ​​of pixels belonging to the oversegmented region in the sample medical image are all 1, while the pixel values ​​of pixels not belonging to the undersegmented region in the sample medical image are 0.

[0099] like Figure 4 As shown, the oversegmented region 406 is determined with the longest diameter 402 as the diameter. The sample medical image 400 is binarized based on the oversegmented region 406 to obtain the oversegmented mask 407.

[0100] Step 304: Input the sample medical image into the first image segmentation network to obtain the first sample probability image output by the first image segmentation network. The first sample probability image includes the probability value of each pixel in the sample medical image belonging to the undersegmented region.

[0101] The image segmentation model in this embodiment includes a first image segmentation network and a second image segmentation network. The first image segmentation network is used to predict under-segmented regions of the sample medical image, while the second image segmentation network is used to predict over-segmented regions of the sample medical image.

[0102] Based on the undersegmentation region prediction function of the first image segmentation network, in one possible implementation, a sample medical image is input into the first image segmentation network, which performs undersegmentation region prediction on the sample medical image to determine whether each pixel in the sample medical image belongs to an undersegmented region, thereby outputting a prediction result first sample probability image. The pixel value of each pixel in the first sample probability image represents the probability that the pixel belongs to the undersegmented region. In other words, the pixel value in the first sample probability image is the probability value of the pixel belonging to the undersegmented region.

[0103] It should be noted that for the first sample probability image, if the pixel value (probability value) corresponding to a pixel is larger, it means that the pixel is more likely to belong to an under-segmented region; conversely, if the pixel value (probability value) corresponding to a pixel is smaller, it means that the pixel is less likely to belong to an under-segmented region.

[0104] Step 305: Train the first image segmentation network based on the first sample probability image and the undersegmentation mask.

[0105] In order for the first image segmentation network to accurately predict the undersegmented regions in the sample medical images, the undersegmentation mask corresponding to the sample medical images needs to be used as weak supervision information to supervise the training of the first image segmentation network. Therefore, in one possible implementation, the segmentation loss in the undersegmentation region prediction process can be determined based on the first sample probability image and the undersegmentation mask, and then the first image segmentation network can be trained based on the segmentation loss.

[0106] In an exemplary example, the segmentation loss of the first image segmentation network can be expressed as:

[0107]

[0108] in, This represents the segmentation loss of the first image segmentation network. Let Q represent the probability image of the first sample, and let Q represent the undersegmentation mask. This represents calculating the probability image of the first sample and the undersegmentation mask across the entire image domain. The soft Dice loss.

[0109] Step 306: Input the sample medical image into the second image segmentation network to obtain the second sample probability image output by the second image segmentation network. The second sample probability image includes the probability value of each pixel in the sample medical image belonging to the oversegmented region.

[0110] Based on the oversegmentation region prediction function of the second image segmentation network, in one possible implementation, a sample medical image is input into the second image segmentation network, which performs oversegmentation region prediction on the sample medical image to determine whether each pixel in the sample medical image belongs to an oversegmentation region, thereby outputting a prediction result second sample probability image. The pixel value of each pixel in the second sample probability image represents the probability that the pixel belongs to an oversegmentation region. In other words, the pixel value in the second sample probability image is the probability value of the pixel belonging to an oversegmentation region.

[0111] It should be noted that for the second sample probability image, the larger the pixel value (probability value) corresponding to a pixel, the higher the probability that the pixel belongs to an oversegmented region; conversely, the smaller the pixel value (probability value) corresponding to a pixel, the lower the probability that the pixel belongs to an oversegmented region.

[0112] Step 307: Train the second image segmentation network based on the second sample probability image and the oversegmentation mask.

[0113] In order for the second image segmentation network to accurately predict oversegmented regions in the sample medical images, the oversegmentation mask corresponding to the sample medical images needs to be used as weak supervision information to supervise the training of the second image segmentation network. Therefore, in one possible implementation, the segmentation loss in the oversegmentation region prediction process can be determined based on the second sample probability image and the oversegmentation mask, and then the second image segmentation network can be trained based on the segmentation loss.

[0114] In an exemplary example, the segmentation loss of the second image segmentation network can be expressed as:

[0115]

[0116] in, This represents the segmentation loss of the second image segmentation network. Let C represent the probability image of the second sample, and let C denote the oversegmentation mask. This represents calculating the second sample probability image and the oversegmentation mask across the entire image domain. The soft Dice loss.

[0117] like Figure 5 The diagram illustrates the training process of an image segmentation model according to an exemplary embodiment of this application. The image segmentation model 500 includes a first image segmentation network 501 and a second image segmentation network 502. A sample medical image 503 is input into the first image segmentation network 501 and the second image segmentation network 502, respectively, to obtain a first sample probability image 504 output by the first image segmentation network 501 and a second sample probability image 505 output by the second image segmentation network 502. Based on the undersegmentation mask 506 corresponding to the first sample probability image 504 and the sample medical image 503, an undersegmentation loss 507 is determined, and the first image segmentation network 501 is trained according to the undersegmentation loss 507. Based on the second sample probability image 505 and the oversegmentation mask 508 corresponding to the sample medical image 503, an oversegmentation loss 509 is determined, and the second image segmentation network 502 is trained according to the oversegmentation loss 509.

[0118] In this embodiment, under-segmented regions are determined by the major and minor axis annotations in the sample diameter annotation to generate corresponding under-segmentation masks; and over-segmented regions are determined by the major axis annotations in the sample diameter annotation to generate corresponding over-segmentation masks, so that the under-segmented regions, sample segmented regions, and over-segmented regions satisfy a specific relationship; furthermore, the first image segmentation network predicts under-segmented regions of the sample medical image to obtain a first sample probability image, and trains the first image segmentation network with the first sample probability image and the under-segmentation mask, so that the first image segmentation network has the function of predicting under-segmented regions; simultaneously, the second image segmentation network predicts over-segmented regions of the sample medical image to obtain a second sample probability image, and trains the second image segmentation network with the second sample probability image and the over-segmentation mask, so that the second image segmentation network has the function of predicting over-segmented regions.

[0119] In the above embodiments, different losses are used to train the first image segmentation network and the second image segmentation network respectively. Therefore, the training processes of the two image segmentation networks are independent of each other and cannot learn from each other. Therefore, in one possible implementation, in order to connect the first image segmentation network and the second image segmentation network, a consistency loss is added during the training process.

[0120] Please refer to Figure 6 This illustration shows a flowchart of an image segmentation method provided in another exemplary embodiment of this application. This embodiment is executed by an example where a computer device is the execution subject of the method, which includes:

[0121] Step 601: Based on the sample diameter annotation, determine the major and minor axis annotations corresponding to the sample segmentation regions in the sample medical image.

[0122] Step 602: Based on the major and minor axis annotations, determine the undersegmentation mask corresponding to the sample medical image.

[0123] Step 603: Based on the major axis annotation, determine the oversegmentation mask corresponding to the sample medical image.

[0124] Step 604: Input the sample medical image into the first image segmentation network to obtain the first sample probability image output by the first image segmentation network. The first sample probability image includes the probability value of each pixel in the sample medical image belonging to the undersegmented region.

[0125] The implementation methods for steps 601 to 604 can be referred to the above embodiments, and will not be repeated here.

[0126] Step 605: Determine the first segmentation loss based on the first sample probability image and the second sample probability image.

[0127] The second sample probability image is obtained by the second image segmentation network from the oversegmented regions of the sample medical image. The second sample probability image includes the probability value of each pixel in the sample medical image belonging to the oversegmented region.

[0128] In this embodiment, in addition to using the undersegmentation prediction loss between the first sample probability image and the undersegmentation mask when training the first image segmentation network, a prediction loss between the first sample probability image and the second sample probability image is also introduced in order to enable the first image segmentation network and the second image segmentation network to learn from each other.

[0129] Although the prediction functions of the first image segmentation network and the second image segmentation network are different, the under-segmented and over-segmented regions predicted by them are essentially related to the sample segmentation regions. Therefore, in one possible implementation, after obtaining the first sample probability image output by the first image segmentation network and the second sample probability image output by the second image segmentation network, the first segmentation loss can be determined based on the first sample probability image and the second sample probability image, and used to train the first image segmentation network and the second image segmentation network, so that the first image segmentation network and the second image segmentation network learn from each other.

[0130] In an exemplary example, the formula for calculating the first segmentation loss can be expressed as:

[0131]

[0132] in, Indicates the first segmentation loss. This represents the probability image of the first sample. This represents the probability image of the second sample. This represents calculating the probability images of the first and second samples across the entire image domain. The soft Dice loss.

[0133] Optionally, in addition to soft Dice loss, the first segmentation loss can also be cross-entropy loss. The specific calculation formula for the first segmentation loss is not limited in the embodiments of this application.

[0134] As can be seen from formula (4), it calculates the segmentation loss between the first sample probability image and the second sample probability image over the entire image domain. However, for undersegmented and oversegmented region prediction, the focus should be on undersegmented and oversegmented regions during the loss determination process. Applying consistency loss to regions other than undersegmented and oversegmented regions may introduce unnecessary factors and affect the prediction accuracy of the segmentation network. Therefore, in order to further improve the prediction accuracy of the segmentation network, in one possible implementation, consistency loss is applied only to specific regions.

[0135] In an exemplary example, step 605 may also include steps 605A and 605B.

[0136] Step 605A: Determine the first region image of the target location region in the first sample probability image and the second region image of the target location region in the second sample probability image.

[0137] In one possible implementation, during the determination of the first segmentation loss, a consistency loss is applied only to the region image at the target location region (specific region) in the first sample probability image and the second sample probability image, corresponding to the need to determine the first region image of the target location region in the first sample probability image and the second region image of the target location region in the second sample probability image.

[0138] The target location region is the region that is irrelevant to the prediction of under-segmented and over-segmented regions.

[0139] In an exemplary example, the process of determining the first region image and the second region image may include the following steps 1 to 3.

[0140] 1. Define the over-segmented regions outside the under-segmented regions as the target location regions.

[0141] In this embodiment, a specific region (target location region) is defined as the part outside the under-divided region within the over-divided region, that is, the part outside the quadrilateral region within the circular region.

[0142] Because the under-segmented region is smaller than the sample segmented region, which is smaller than the over-segmented region: For any pixel, if it belongs to an undersegmented region, then the pixel value of that pixel is 1 in the undersegmented mask, oversegmented mask, and sample segmentation mask (the mask used to indicate the sample segmentation region). Q p =M p =C p =1; If any pixel does not belong to the oversegmented region, then the pixel value of that pixel in the undersegmented mask, oversegmented mask, and sample segmentation mask is 0: Q p =M p =C p =0; Therefore, we can conclude that if any pixel belongs to an under-segmented region or an image domain other than an over-segmented region, then the pixel value of that pixel is the same in the under-segmentation mask, the over-segmentation mask, and the sample segmentation mask: Q p =M p =C p In other words, the undersegmentation mask, oversegmentation mask, and sample segmentation mask are consistent outside the circular region (oversegmentation region) and inside the quadrilateral region (undersegmentation region).

[0143] Since the undersegmentation mask, oversegmentation mask, and sample segmentation mask are consistent outside the oversegmentation region and within the undersegmentation region, in order to avoid unnecessary losses in the loss calculation process due to consistent regions, in one possible implementation, consistency loss only needs to be applied to regions that are not consistent, that is, the target location region is the oversegmentation region outside the undersegmentation region.

[0144] 2. Based on the target location region, determine the first region image in the first sample probability image.

[0145] In one possible implementation, once the target location region is determined based on the undersegmented and oversegmented regions corresponding to the sample medical image, the pixels in the first sample probability image located in the target location region can be defined as the first region image.

[0146] 3. Based on the target location region, determine the second region image in the second sample probability image.

[0147] In one possible implementation, once the target location region is determined based on the undersegmented and oversegmented regions corresponding to the sample medical image, the pixels in the second sample probability image located in the target location region can be identified as the first region image.

[0148] Step 605B: Determine the first segmentation loss based on the first region image and the second region image.

[0149] In one possible implementation, once the first region image and the second region image are determined, the first segmentation loss can be determined based on the first region image and the second region image.

[0150] In an exemplary example, the formula for calculating the consistency loss applied to the target location region can be expressed as:

[0151]

[0152] in, Indicates the first segmentation loss. This represents the probability image of the first sample. This represents the probability image of the second sample. This represents finding the probability images of the first and second samples in the image domain. Soft Dice loss on the image domain This indicates the target location area.

[0153] Step 606: Determine the second segmentation loss based on the first sample probability image and the undersegmentation mask.

[0154] As can be seen from formula (2), in order to enable the first image segmentation network to have the function of predicting under-segmented regions, in one possible implementation, the second segmentation loss is determined based on the first sample probability image and the under-segmentation mask to train the first image segmentation network.

[0155] Step 607: Train the first image segmentation network based on the first segmentation loss and the second segmentation loss.

[0156] In an exemplary example, after introducing a consistency loss during the training of the first image segmentation network, the total loss of the first image segmentation network can be expressed as:

[0157]

[0158] in, This represents the total loss of the first image segmentation network. This represents the loss between the undersegmentation mask and the probability image of the first sample. Represents the first sample probability image and the second sample probability image in the image domain. The loss is λ, which is a hyperparameter and can take values ​​in the range of [0.2, 0.8].

[0159] The consistency loss in Equation (6) is over the entire image domain. The loss is calculated above. To avoid the influence of the consistency region on this loss, the consistency loss is applied only to the target location region. The corresponding loss in the reconstruction formula (6) is:

[0160]

[0161] in, This represents the loss of the first image segmentation network. This represents the loss between the undersegmentation mask and the probability image of the first sample. Represents the first sample probability image and the second sample probability image in the image domain. The loss over the target location region, where λ is a hyperparameter and its value can range from [0.2, 0.8].

[0162] As can be seen from formulas (6) and (7), in one possible implementation, the first image segmentation network can be trained together with the first segmentation loss and the second segmentation loss; wherein, the first segmentation loss can be selected from formulas (4) and (5) according to actual needs.

[0163] Step 608: Input the sample medical image into the second image segmentation network to obtain the second sample probability image output by the second image segmentation network. The second sample probability image includes the probability value of each pixel in the sample medical image belonging to the oversegmented region.

[0164] The implementation method of step 608 can be referred to the above embodiment, and will not be repeated here.

[0165] Step 609: Determine the first segmentation loss based on the first sample probability image and the second sample probability image.

[0166] The first sample probability image is obtained by the first image segmentation network predicting undersegmented regions of the sample medical image. The first sample probability image includes the probability value of each pixel in the sample medical image belonging to the undersegmented region.

[0167] Similar to the training process of the first image segmentation network, a first segmentation loss can also be introduced between the first sample probability image and the second sample probability image during the training of the second image segmentation network, so that the first image segmentation network and the second image segmentation network can learn from each other; in one possible implementation, the first segmentation loss can be determined based on the first sample probability image and the second sample probability image.

[0168] The process of determining the first segmentation loss can be referred to in the above embodiment, and will not be repeated here.

[0169] Step 610: Determine the third segmentation loss based on the second sample probability image and the oversegmentation mask.

[0170] As can be seen from formula (3), in order to enable the second image segmentation network to have the function of oversegmentation region prediction, in one possible implementation, the third segmentation loss is determined based on the second sample probability image and the oversegmentation mask to train the second image segmentation network.

[0171] Step 611: Train the second image segmentation network based on the first segmentation loss and the third segmentation loss.

[0172] In an exemplary example, after introducing a consistency loss during the training of the second image segmentation network, the total loss of the second image segmentation network can be expressed as:

[0173]

[0174] in, This represents the total loss of the first image segmentation network. This represents the loss between the oversegmentation mask and the second sample probability image. Represents the first sample probability image and the second sample probability image in the image domain. The loss is λ, which is a hyperparameter and can take values ​​in the range of [0.2, 0.8].

[0175] like Figure 7 The diagram illustrates the training process of an image segmentation model according to another exemplary embodiment of this application. The image segmentation model 700 includes a first image segmentation network 701 and a second image segmentation network 702. A sample medical image 703 is input into the first image segmentation network 701 and the second image segmentation network 702, respectively, to obtain a first sample probability image 704 output by the first image segmentation network 701 and a second sample probability image 705 output by the second image segmentation network 702. Based on the undersegmentation mask 706 corresponding to the first sample probability image 704 and the sample medical image 703, a second segmentation loss 707 is determined, and based on the first sample probability image 704 and the second image segmentation model 705... The first segmentation loss 710 is determined based on the sample probability image 705, and then the first image segmentation network 701 is trained based on the first segmentation loss 710 and the second segmentation loss 707; similarly, the third segmentation loss 709 is determined based on the oversegmentation mask 708 corresponding to the second sample probability image 705 and the sample medical image 703, and then the second image segmentation network 702 is trained based on the third segmentation loss 709 and the first segmentation loss 710; wherein, the first segmentation loss 710 is the segmentation loss calculated for the first sample probability image 704 and the second sample probability image 705 in the entire image domain.

[0176] The consistency loss in Equation (8) is over the entire image domain. The loss is calculated above. To avoid the influence of the consistency region on this loss, the consistency loss is applied only to the target location region. The loss corresponding to the reconstruction formula (8) is:

[0177]

[0178] in, This represents the total loss of the first image segmentation network. This represents the loss between the oversegmentation mask and the second sample probability image. Represents the first sample probability image and the second sample probability image in the image domain. The loss over the target location region, where λ is a hyperparameter and its value can range from [0.2, 0.8].

[0179] like Figure 8 The diagram illustrates the training process of an image segmentation model according to another exemplary embodiment of this application. The image segmentation model 800 includes a first image segmentation network 801 and a second image segmentation network 802. A sample medical image 803 is input into the first image segmentation network 801 and the second image segmentation network 802, respectively, to obtain a first sample probability image 804 output by the first image segmentation network 801 and a second sample probability image 805 output by the second image segmentation network 802. Based on the undersegmentation mask 806 corresponding to the first sample probability image 804 and the sample medical image 803, a second segmentation loss 807 is determined. When determining the first segmentation loss 812, a first region image 810 of the target location region in the first sample probability image 804 and the second sample probability image 805 are first determined. The first image segmentation network 801 is trained based on the second region image 810 and the second region image 811, and a first segmentation loss 812 is determined based on the first segmentation loss 812 and the second segmentation loss 807. Similarly, a third segmentation loss 809 is determined based on the oversegmentation mask 808 corresponding to the second sample probability image 805 and the sample medical image 803, and a second image segmentation network 802 is trained based on the third segmentation loss 809 and the first segmentation loss 812. The first segmentation loss 812 is the segmentation loss calculated for the first sample probability image 804 and the second sample probability image 805 in the target location region.

[0180] As can be seen from formulas (8) and (9), in one possible implementation, the second image segmentation network can be trained together based on the first segmentation loss and the third segmentation loss; wherein, the first segmentation loss can be selected from formulas (4) and (5) according to actual needs.

[0181] In this embodiment, by adding a consistency loss during the training of the first image segmentation network and the second image segmentation network, the first image segmentation network and the second image segmentation network can learn from each other while making independent predictions, thereby improving the prediction accuracy of the first image segmentation network and the second image segmentation network. In addition, applying the consistency loss only to the target location region can avoid the additional influence of the consistent region, which is beneficial to further improve the prediction accuracy of the image segmentation network.

[0182] The above embodiments mainly describe the training process of the image segmentation network. After the model is trained, a first image segmentation network with undersegmentation region prediction function and a second image segmentation network with oversegmentation prediction function can be obtained. The first image segmentation network and the second image segmentation network can be deployed isomorphically to segment medical images to determine the segmentation regions corresponding to the medical images.

[0183] Please refer to Figure 9 This document illustrates a flowchart of an image segmentation method provided in an exemplary embodiment of this application. This embodiment uses a computer device as an example to illustrate the method, which includes:

[0184] Step 901: Perform undersegmentation region prediction on the target segmentation region in the target medical image to obtain a first target probability image. The first target probability image contains the probability value of each pixel in the target medical image belonging to the undersegmented region, and the undersegmented region is smaller than the target segmentation region.

[0185] The image segmentation model in this embodiment includes two sub-networks: a first image segmentation network and a second image segmentation network. The first image segmentation network is used to segment under-segmented regions in medical images, and the second image segmentation network is used to segment over-segmented regions in medical images. The segmentation result of the segmented regions is then determined through model ensemble.

[0186] Similar to the model training process, the model application process also requires the first image segmentation network and the second image segmentation network to predict the segmentation regions of the target medical image respectively. In one possible implementation, the target medical image is input into the first image segmentation network, which predicts the undersegmented regions of the target segmentation region in the target medical image and determines the probability that each pixel in the target medical image belongs to the undersegmented region, thereby obtaining the undersegmented region prediction result: the first target probability image.

[0187] Step 902: Perform oversegmentation region prediction on the target segmentation region in the target medical image to obtain a second target probability image. The second target probability image contains the probability that each pixel in the target medical image belongs to an oversegmentation region, and the oversegmentation region is greater than the target segmentation region.

[0188] In one possible implementation, the target medical image is input into a second image segmentation network, which performs oversegmentation region prediction on the target segmentation region in the target medical image, determines the probability that each pixel in the target medical image belongs to the oversegmentation region, and thus obtains the oversegmentation region prediction result: the second target probability image.

[0189] Similar to the training process of the first image segmentation network, the relationship between the undersegmented region, the target segmentation region, and the oversegmented region is as follows: the undersegmented region is contained within the target segmentation region, and the target segmentation region is contained within the oversegmented region.

[0190] It should be noted that steps 902 and 903 can be executed simultaneously, or steps 902 can be executed first and then steps 903; or steps 903 can be executed first and then steps 902.

[0191] Step 903: Based on the first target probability image and the second target probability image, determine the target segmentation mask corresponding to the target medical image.

[0192] Since the first target probability image is the prediction result of the under-segmented region of the target segmentation region in the target medical image, and the second target probability image is the prediction result of the over-segmented region of the target segmentation region in the target medical image, and both the under-segmented and over-segmented regions differ from the target segmentation region, it is clear that a single prediction result is not an accurate segmentation result of the target medical image. Therefore, in order to obtain an accurate segmentation result of the target medical image, in one possible implementation, a target segmentation mask corresponding to the target medical image is determined based on the first target probability image and the second target probability image through model ensemble.

[0193] Among them, the target segmentation mask can clearly indicate the target segmentation region in the target medical image.

[0194] In this embodiment, during model training, by using the sample diameter annotations corresponding to the sample segmentation regions in the sample medical images, supervision information for training the first and second image segmentation networks is constructed: undersegmentation masks and oversegmentation masks. This enables the first image segmentation network to predict undersegmented regions and the second image segmentation network to predict oversegmented regions, allowing the trained first and second image segmentation networks to be used for segmentation prediction of segmented regions in medical images. Compared to related technologies that require the use of sample segmentation contour annotations of sample segmentation regions as supervision information, this embodiment uses sample diameter annotations to generate weak supervision information, which reduces manual annotation operations. This allows for the acquisition of large-scale training sample data with less manual annotation, improving the efficiency of generating supervision information for training the image segmentation networks and further improving model training efficiency.

[0195] Since the first target probability image is the prediction result of the under-segmented region of the target segmentation region in the target medical image, and the second target probability image is the prediction result of the over-segmented region of the target segmentation region in the target medical image, based on the relationship between the under-segmented region, the over-segmented region and the target segmentation region, in one possible implementation, the accurate segmentation result of the target medical image can be obtained by averaging the pixel values ​​of the first target probability image and the second target probability image.

[0196] Please refer to Figure 10 This illustration shows a flowchart of an image segmentation method provided in another exemplary embodiment of this application. This embodiment is executed by an example where a computer device is the execution subject of the method, which includes:

[0197] Step 1001: Perform undersegmentation region prediction on the target segmentation region in the target medical image to obtain a first target probability image. The first target probability image contains the probability value of each pixel in the target medical image belonging to the undersegmentation region, and the undersegmentation region is smaller than the target segmentation region.

[0198] Step 1002: Perform oversegmentation region prediction on the target segmentation region in the target medical image to obtain a second target probability image. The second target probability image contains the probability that each pixel in the target medical image belongs to an oversegmentation region, and the oversegmentation region is greater than the target segmentation region.

[0199] The implementation methods for steps 1001 and 1002 can be referred to the above embodiments, and will not be repeated here.

[0200] Step 1003: Generate a target segmentation probability image based on the first target probability image and the second target probability image.

[0201] In one possible implementation, by integrating the prediction results of the model—a first target probability image and a second target probability image—a target segmentation probability image corresponding to the target segmentation region can be obtained, and then a corresponding target segmentation mask can be generated based on the target segmentation probability image.

[0202] In an exemplary example, the relationship between the first target probability image, the second target probability image, and the target segmentation probability image can be represented as:

[0203]

[0204] in, This represents the target segmentation probability image (the predicted result of the target segmentation region). This represents the probability image of the first target. This represents the probability image of the second target.

[0205] As can be seen from formula (10), in one possible implementation, by averaging the pixel values ​​corresponding to the same pixel point in the first target probability image and the second target probability image, the target segmentation probability image corresponding to the target segmentation region can be obtained.

[0206] Step 1004: Based on the target segmentation probability image, determine the target segmentation mask corresponding to the target medical image.

[0207] The pixel value of each pixel in the target segmentation probability image can represent the probability that the pixel belongs to the target segmentation region. In order to better distinguish the target segmentation region from other regions, in one possible implementation, the target segmentation probability image can be binarized to obtain a target segmentation mask containing two types of pixel values.

[0208] In an exemplary example, the process of binarizing the target segmentation probability image may include the following steps, that is, step 1004 may include steps 1004A to 1004C.

[0209] Step 1004A: If the pixel value of a pixel in the target segmentation probability image is higher than the probability threshold, the pixel value of the pixel is determined as the first pixel value.

[0210] Since the pixel value of each pixel in the target segmentation probability image can represent the probability that the pixel belongs to the target segmentation region, and the higher the pixel value (probability value), the more likely the pixel is to belong to the target segmentation region, and vice versa, the lower the pixel value (probability value), the lower the probability that the pixel belongs to the target segmentation region, a probability threshold is set in one possible implementation. The target segmentation probability image is binarized by the probability threshold. Based on the relationship between pixel value and target segmentation region, if the pixel value of a pixel in the target segmentation probability image is higher than the probability threshold, it means that it belongs to the target segmentation region, and the pixel value of that pixel can be set as the first pixel value.

[0211] For illustration purposes, the probability threshold can be 0.5, and the first pixel value can be 1.

[0212] Step 1004B: If the pixel value of a pixel in the target segmentation probability image is lower than the probability threshold, the pixel value of the pixel is determined as the second pixel value.

[0213] Conversely, if the pixel value of a pixel in the target segmentation probability image is lower than the probability threshold, it means that it does not belong to the target segmentation region, and the pixel value of that pixel can be set to the second pixel value.

[0214] For illustrative purposes, the second pixel value can be 0.

[0215] Step 1004C: Generate a target segmentation mask based on the first pixel value and the second pixel value.

[0216] In one possible implementation, after binarizing the target segmentation probability image, the pixel values ​​of each pixel are divided into two categories: first pixel value and second pixel value. The image composed of the first pixel value and the second pixel value is the target segmentation mask. From the target segmentation mask, it can be seen that the set of pixels with the first pixel value is the target segmentation region of the target medical image.

[0217] like Figure 11 The diagram illustrates the principle of an image segmentation process according to an exemplary embodiment of this application. The target medical image 1103 is segmented into the first image segmentation network 1101 and the second image segmentation network 1102 in the image segmentation model 1100, resulting in a first target probability image 1104 output by the first image segmentation network 1101 and a second target probability image 1105 output by the second image segmentation network 1102. A target segmentation probability image 1106 is generated based on the first target probability image 1104 and the second target probability image 1105. By binarizing the target segmentation probability image 1106, the target segmentation mask 1107 corresponding to the target medical image 1103 can be obtained.

[0218] Step 1005: Based on the target segmentation mask and the target medical image, generate the target segmentation image corresponding to the target segmentation region.

[0219] In one possible implementation, if it is necessary to segment a target segmented image from a target medical image, the target medical image can be multiplied based on the target segmentation mask to obtain the target segmented image corresponding to the target segmentation region.

[0220] In this embodiment, by integrating the prediction results of the model—the first target probability image and the second target probability image—a target segmentation probability image corresponding to the target segmentation region can be obtained. Then, a corresponding target segmentation mask is generated based on the target segmentation probability image to achieve the image segmentation purpose of the target segmentation region in the target medical image.

[0221] As shown in Table 1, it presents a comparison of the segmentation accuracy of related technical solutions and the image segmentation method provided in this application on the KiTS19 and LiTS17 datasets (the higher the value, the higher the segmentation accuracy).

[0222] Table 1

[0223]

[0224]

[0225] As shown in Table 1, the image segmentation method provided in this application's embodiments outperforms the comparison schemes on both publicly available datasets. Furthermore, when the image segmentation network uses U-Net and Swin-Transformer, the segmentation performance of this application's embodiments can approach the upper limit of full supervision.

[0226] As shown in Table 2, the sensitivity (%) under 0.5 false positives / CT images (FPs / scan) was used as the detection evaluation index to obtain the comparison results between related technologies and the image segmentation method provided in this application (the higher the value, the more accurate the detection result).

[0227] Table 2

[0228] Method Sensitivity 3DCE, 27 Slices 62.48 Deformable Faster-RCNN+VA 69.1 3DCE+CS_Att, 21 slices 71.4 Anchor-Free RPN 68.73 FPN+MSB (weights sharing) 67.0 Improved RetinaNet 72.15 MVP-Net, 9 slices 73.83 MULAN 76.12 MR-CNN 75.92 RECIST-Net 76.14 Oours (with Swin Transformer) 76.43

[0229] As shown in Table 2, although the image segmentation method provided in this application embodiment is not specifically designed for lesion detection, the detection results converted from the segmentation results obtained using the image segmentation method provided in this application embodiment still exceed those of the comparative schemes.

[0230] To further verify the effectiveness of the image segmentation method provided in the application embodiments, Table 3 shows the ablation experiment results on the LiTS17 dataset.

[0231] Table 3

[0232]

[0233] Table 4 shows the ablation experiment results on the KiTS19 dataset.

[0234] Table 4

[0235]

[0236] As shown in Tables 3 and 4, the experiments on both datasets demonstrate that adding... It significantly improved the network's segmentation performance, while reducing regional limitations. This further improves the network's segmentation performance. Furthermore, the segmentation performance of this technique on the KiTS19 dataset approaches the upper bound of fully supervised learning.

[0237] Furthermore, the influence of different λ values ​​on the segmentation results was tested within the range of [0.2, 0.8] to verify the sensitivity of this technical solution to the hyperparameter λ. Figure 12 This is a schematic diagram illustrating the impact of different λ values ​​on segmentation results on the KiTS19 dataset. The horizontal axis represents the value of λ, and the vertical axis represents the Dice Score (Dice coefficient) that measures the segmentation results. Piece 1210 shows the segmentation results for different λ values ​​when the image segmentation network is Swin Transformer; Piece 1220 shows the segmentation results for different λ values ​​when the image segmentation network is U-Net. Figure 13 This shows the test results on the impact of different λ values ​​on segmentation results on the LiTS17 dataset. The horizontal axis represents the value of λ, and the vertical axis represents the Dice coefficient, which measures the segmentation result. Line 1310 represents the segmentation results for different λ values ​​when the image segmentation network is Swin Transformer; line 1320 represents the segmentation results for different λ values ​​when the image segmentation network is U-Net. Figure 12 and Figure 13 It is understood that, within a reasonable range, the image segmentation method provided in this application embodiment is not sensitive to the value of the hyperparameter λ.

[0238] The following are device embodiments of this application. For details not described in detail in the device embodiments, please refer to the above method embodiments.

[0239] Figure 14 This is a structural block diagram of an image segmentation apparatus provided in an exemplary embodiment of this application. The apparatus may include:

[0240] The first segmentation prediction module 1401 is used to predict the under-segmented region of the target segmentation region in the target medical image to obtain a first target probability image. The first target probability image contains the probability value of each pixel in the target medical image belonging to the under-segmented region. The under-segmented region is smaller than the target segmentation region.

[0241] The second segmentation prediction module 1402 is used to predict the oversegmentation region of the target segmentation region in the target medical image to obtain a second target probability image. The second target probability image contains the probability that each pixel in the target medical image belongs to the oversegmentation region, and the oversegmentation region is larger than the target segmentation region.

[0242] The first determining module 1403 is used to determine the target segmentation mask corresponding to the target medical image based on the first target probability image and the second target probability image.

[0243] Optionally, the first determining module 1403 includes:

[0244] The generation unit is configured to generate a target segmentation probability image based on the first target probability image and the second target probability image;

[0245] The first determining unit is used to determine the target segmentation mask corresponding to the target medical image based on the target segmentation probability image.

[0246] Optionally, the first determining unit is further configured to:

[0247] If the pixel value of a pixel in the target segmentation probability image is higher than the probability threshold, the pixel value of the pixel is determined as the first pixel value.

[0248] If the pixel value of a pixel in the target segmentation probability image is lower than the probability threshold, the pixel value of the pixel is determined as the second pixel value.

[0249] The target segmentation mask is generated based on the first pixel value and the second pixel value.

[0250] Optionally, the first generating unit is further configured to:

[0251] The target segmentation probability image is obtained by averaging the pixel values ​​corresponding to the same pixel point in the first target probability image and the second target probability image.

[0252] Optionally, the device further includes:

[0253] The generation module is used to generate a target segmentation image corresponding to the target segmentation region based on the target segmentation mask and the target medical image.

[0254] In summary, this application provides a weakly supervised image segmentation method. By using the sample diameter annotations corresponding to the segmentation regions in a sample medical image, supervision information for training a first image segmentation network and a second image segmentation network is constructed: undersegmentation masks and oversegmentation masks. This enables the first image segmentation network to predict undersegmented regions and the second image segmentation network to predict oversegmented regions, thus allowing the trained first and second image segmentation networks to be used for segmentation prediction of segmentation regions in medical images. Compared to related technologies that require the use of sample segmentation contour annotations of the segmentation regions as supervision information, this application uses sample diameter annotations to generate weakly supervised information, which reduces manual annotation operations. This allows for the acquisition of large-scale training sample data with less manual annotation, improving the efficiency of generating supervision information for training the image segmentation network and further improving model training efficiency.

[0255] Figure 15 This is a structural block diagram of an image segmentation apparatus provided in another exemplary embodiment of this application. The apparatus includes:

[0256] The second determining module 1501 is used to determine the undersegmentation mask and oversegmentation mask corresponding to the sample medical image based on the sample diameter annotation corresponding to the sample segmentation region in the sample medical image, wherein the undersegmentation region corresponding to the undersegmentation mask is smaller than the sample segmentation region, and the oversegmentation region corresponding to the oversegmentation mask is larger than the sample segmentation region.

[0257] The first training module 1502 is used to train a first image segmentation network based on the sample medical image and the undersegmentation mask. The first image segmentation network is used to predict the undersegmentation region of the sample medical image.

[0258] The second training module 1503 is used to train a second image segmentation network based on the sample medical image and the oversegmentation mask. The second image segmentation network is used to predict oversegmentation regions in the sample medical image.

[0259] Optionally, the second determining module 1501 includes:

[0260] The second determining unit is used to determine the major axis label and minor axis label corresponding to the sample segmentation region in the sample medical image based on the sample diameter label, wherein the major axis label is the label of the longest diameter corresponding to the sample segmentation region, and the minor axis label is the label of the shortest diameter corresponding to the sample segmentation region;

[0261] The third determining unit is used to determine the undersegmentation mask corresponding to the sample medical image based on the major axis label and the minor axis label;

[0262] The fourth determining unit is used to determine the oversegmentation mask corresponding to the sample medical image based on the major axis annotation.

[0263] Optionally, the third determining unit is further configured to:

[0264] Based on the major axis annotation and the minor axis annotation, the under-segmented region in the sample medical image is determined;

[0265] The pixel values ​​of pixels belonging to the undersegmented region in the medical sample image are set to the first pixel value, and the pixel values ​​of pixels not belonging to the undersegmented region in the medical sample image are set to the second pixel value, thereby generating the undersegmented mask corresponding to the sample medical image.

[0266] Optionally, the fourth determining unit is further configured to:

[0267] Based on the major axis annotation, the oversegmented region in the sample medical image is determined;

[0268] The pixel values ​​of pixels belonging to the oversegmented region in the medical sample image are set to the first pixel value, and the pixel values ​​of pixels not belonging to the oversegmented region in the medical sample image are set to the second pixel value, thereby generating the oversegmentation mask corresponding to the sample medical image.

[0269] Optionally, the first training module 1502 includes:

[0270] The first segmentation prediction unit is used to input the sample medical image into the first image segmentation network to obtain a first sample probability image output by the first image segmentation network. The first sample probability image includes the probability value of each pixel in the sample medical image belonging to the under-segmented region.

[0271] The first training unit is used to train the first image segmentation network based on the first sample probability image and the undersegmentation mask.

[0272] Optionally, the device further includes:

[0273] The third determining module is used to determine the first segmentation loss based on the first sample probability image and the second sample probability image. The second sample probability image is obtained by the second image segmentation network predicting oversegmented regions of the sample medical image. The second sample probability image includes the probability value of each pixel in the sample medical image belonging to the oversegmented region.

[0274] The first training unit is also used for:

[0275] Based on the first sample probability image and the undersegmentation mask, determine the second segmentation loss;

[0276] The first image segmentation network is trained based on the first segmentation loss and the second segmentation loss.

[0277] Optionally, the second training module 1503 includes:

[0278] The second segmentation prediction unit is used to input the sample medical image into the second image segmentation network to obtain a second sample probability image output by the second image segmentation network. The second sample probability image includes the probability value of each pixel in the sample medical image belonging to the over-segmented region.

[0279] The second training unit is used to train the second image segmentation network based on the second sample probability image and the oversegmentation mask.

[0280] Optionally, the device further includes:

[0281] The fourth determining module is used to determine a first segmentation loss based on the first sample probability image and the second sample probability image. The first sample probability image is obtained by the first image segmentation network predicting undersegmented regions of the sample medical image. The first sample probability image includes the probability value of each pixel in the sample medical image belonging to the undersegmented region.

[0282] The second training unit is also used for:

[0283] Based on the second sample probability image and the oversegmentation mask, a third segmentation loss is determined;

[0284] The second image segmentation network is trained based on the first segmentation loss and the third segmentation loss.

[0285] Optionally, the fourth determining module includes:

[0286] The fifth determining unit is used to determine a first region image of the target location region in the first sample probability image and a second region image of the target location region in the second sample probability image;

[0287] The sixth determining unit is used to determine the first segmentation loss based on the first region image and the second region image.

[0288] Optionally, the fifth determining unit is further configured to:

[0289] The over-segmented region outside the under-segmented region is determined as the target location region;

[0290] Based on the target location region, determine the first region image in the first sample probability image;

[0291] Based on the target location region, the second region image in the second sample probability image is determined.

[0292] In summary, this application provides a weakly supervised image segmentation method. By using the sample diameter annotations corresponding to the segmentation regions in a sample medical image, supervision information for training a first image segmentation network and a second image segmentation network is constructed: undersegmentation masks and oversegmentation masks. This enables the first image segmentation network to predict undersegmented regions and the second image segmentation network to predict oversegmented regions, thus allowing the trained first and second image segmentation networks to be used for segmentation prediction of segmentation regions in medical images. Compared to related technologies that require sample segmentation contour annotations of the segmentation regions as supervision information, this application uses sample diameter annotations to generate weakly supervised information, which reduces manual annotation operations. This allows for the acquisition of large-scale training sample data with less manual annotation, improving the efficiency of generating supervision information for training the image segmentation network and further improving model training efficiency.

[0293] Please refer to Figure 16 This illustration shows a schematic diagram of a computer device according to an embodiment of this application. This computer device can be used to implement the image segmentation method performed by the computer device provided in the above embodiment. The computer device 1600 includes a central processing unit (CPU) 1601, a system memory 1604 including random access memory (RAM) 1602 and read-only memory (ROM) 1603, and a system bus 1605 connecting the system memory 1604 and the central processing unit 1601. The computer device 1600 also includes a basic input / output system (I / O) 1606 to facilitate information transfer between various devices within the computer, and a mass storage device 1607 for storing the operating system 1613, application programs 1614, and other program modules 1615.

[0294] The basic input / output system 1606 includes a display 1608 for displaying information and an input device 1609 for user input, such as a mouse or keyboard. Both the display 1608 and the input device 1609 are connected to the central processing unit 1601 via an input / output controller 1610 connected to the system bus 1605. The basic input / output system 1606 may also include the input / output controller 1610 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 1610 also provides output to a display screen, printer, or other types of output devices.

[0295] The mass storage device 1607 is connected to the central processing unit 1601 via a mass storage controller (not shown) connected to the system bus 1605. The mass storage device 1607 and its associated computer-readable media provide non-volatile storage for the computer device 1600. That is, the mass storage device 1607 may include computer-readable media (not shown) such as a hard disk or a CD-ROM (Compact Disc Read-Only Memory) drive.

[0296] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technologies, CD-ROM, DVD (Digital Video Disc) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage media are not limited to the above-mentioned types. The system memory 1604 and mass storage device 1607 described above can be collectively referred to as memory.

[0297] According to various embodiments of this application, the computer device 1600 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1600 can be connected to a network 1612 via a network interface unit 1611 connected to the system bus 1605, or the network interface unit 1611 can be used to connect to other types of networks or remote computer systems (not shown).

[0298] The memory also includes one or more programs stored in the memory and configured to be executed by one or more central processing units 1601.

[0299] This application also provides a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the image segmentation method provided in any of the above exemplary embodiments.

[0300] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the image segmentation method provided in the optional implementation described above.

[0301] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0302] The above description is merely an optional embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An image segmentation method, characterized in that, The method includes: Based on the sample diameter annotation corresponding to the sample segmentation region in the sample medical image, the major axis annotation and minor axis annotation corresponding to the sample segmentation region are determined. The major axis annotation is the annotation of the longest diameter corresponding to the sample segmentation region, and the minor axis annotation is the annotation of the shortest diameter corresponding to the sample segmentation region. Based on the major axis label and the minor axis label, an undersegmentation mask corresponding to the sample medical image is determined, wherein the undersegmentation region corresponding to the undersegmentation mask is smaller than the sample segmentation region; Based on the major axis annotation, an oversegmentation mask corresponding to the sample medical image is determined, wherein the oversegmentation region corresponding to the oversegmentation mask is larger than the sample segmentation region; Based on the sample medical image and the undersegmentation mask, a first image segmentation network is trained, which is used to predict the undersegmentation region of the sample medical image. Based on the sample medical image and the oversegmentation mask, a second image segmentation network is trained. The second image segmentation network is used to predict oversegmentation regions in the sample medical image.

2. The method according to claim 1, characterized in that, The step of determining the undersegmentation mask corresponding to the sample medical image based on the major axis annotation and the minor axis annotation includes: Based on the major axis annotation and the minor axis annotation, the under-segmented region in the sample medical image is determined; The pixel values ​​of pixels belonging to the undersegmented region in the sample medical image are set to the first pixel value, and the pixel values ​​of pixels not belonging to the undersegmented region in the sample medical image are set to the second pixel value, thereby generating the undersegmented mask corresponding to the sample medical image.

3. The method according to claim 1, characterized in that, The step of determining the oversegmentation mask corresponding to the sample medical image based on the major axis annotation includes: Based on the major axis annotation, the oversegmented region in the sample medical image is determined; The pixel values ​​of pixels belonging to the oversegmented region in the sample medical image are set to the first pixel value, and the pixel values ​​of pixels not belonging to the oversegmented region in the sample medical image are set to the second pixel value, thereby generating the oversegmentation mask corresponding to the sample medical image.

4. The method according to any one of claims 1 to 3, characterized in that, The step of training the first image segmentation network based on the sample medical image and the undersegmentation mask includes: The sample medical image is input into the first image segmentation network to obtain a first sample probability image output by the first image segmentation network. The first sample probability image includes the probability value of each pixel in the sample medical image belonging to the under-segmented region. The first image segmentation network is trained based on the first sample probability image and the undersegmentation mask.

5. The method according to claim 4, characterized in that, The method further includes: Based on the first sample probability image and the second sample probability image, a first segmentation loss is determined. The second sample probability image is obtained by the second image segmentation network predicting oversegmented regions of the sample medical image. The second sample probability image includes the probability value of each pixel in the sample medical image belonging to the oversegmented region. The step of training the first image segmentation network based on the first sample probability image and the undersegmentation mask includes: Based on the first sample probability image and the undersegmentation mask, determine the second segmentation loss; The first image segmentation network is trained based on the first segmentation loss and the second segmentation loss.

6. The method according to any one of claims 1 to 3, characterized in that, The step of training a second image segmentation network based on the sample medical image and the oversegmentation mask includes: The sample medical image is input into the second image segmentation network to obtain a second sample probability image output by the second image segmentation network. The second sample probability image includes the probability value of each pixel in the sample medical image belonging to the over-segmented region. The second image segmentation network is trained based on the second sample probability image and the oversegmentation mask.

7. The method according to claim 6, characterized in that, The method further includes: Based on the first sample probability image and the second sample probability image, a first segmentation loss is determined. The first sample probability image is obtained by the first image segmentation network predicting under-segmented regions of the sample medical image. The first sample probability image includes the probability value of each pixel in the sample medical image belonging to the under-segmented region. The step of training the second image segmentation network based on the second sample probability image and the oversegmentation mask includes: Based on the second sample probability image and the oversegmentation mask, a third segmentation loss is determined; The second image segmentation network is trained based on the first segmentation loss and the third segmentation loss.

8. The method according to claim 7, characterized in that, The determination of the first segmentation loss based on the first sample probability image and the second sample probability image includes: A first region image of the target location region in the first sample probability image and a second region image of the target location region in the second sample probability image are determined. The first segmentation loss is determined based on the first region image and the second region image.

9. The method according to claim 8, characterized in that, The process of determining a first region image of the target location region in the first sample probability image and a second region image of the target location region in the second sample probability image includes: The over-segmented region outside the under-segmented region is determined as the target location region; Based on the target location region, determine the first region image in the first sample probability image; Based on the target location region, the second region image in the second sample probability image is determined.

10. An image segmentation method, characterized in that, The method includes: A first target probability image is obtained by predicting under-segmented regions in the target segmentation region of the target medical image through a first image segmentation network. The first target probability image contains the probability value of each pixel in the target medical image belonging to the under-segmented region, and the under-segmented region is smaller than the target segmentation region. A second image segmentation network is used to predict oversegmentation regions in the target segmentation region of the target medical image to obtain a second target probability image. The second target probability image contains the probability that each pixel in the target medical image belongs to an oversegmentation region, and the oversegmentation region is larger than the target segmentation region. Based on the first target probability image and the second target probability image, determine the target segmentation mask corresponding to the target medical image; The first image segmentation network and the second image segmentation network are trained using the method described in any one of claims 1 to 9.

11. The method according to claim 10, characterized in that, The step of determining the target segmentation mask corresponding to the target medical image based on the first target probability image and the second target probability image includes: Based on the first target probability image and the second target probability image, a target segmentation probability image is generated; Based on the target segmentation probability image, the target segmentation mask corresponding to the target medical image is determined.

12. The method according to claim 11, characterized in that, The step of determining the target segmentation mask corresponding to the target medical image based on the target segmentation probability image includes: If the pixel value of a pixel in the target segmentation probability image is higher than the probability threshold, the pixel value of the pixel is determined as the first pixel value. If the pixel value of a pixel in the target segmentation probability image is lower than the probability threshold, the pixel value of the pixel is determined as the second pixel value. The target segmentation mask is generated based on the first pixel value and the second pixel value.

13. The method according to claim 11, characterized in that, The step of generating a target segmentation probability image based on the first target probability image and the second target probability image includes: The target segmentation probability image is obtained by averaging the pixel values ​​corresponding to the same pixel point in the first target probability image and the second target probability image.

14. The method according to any one of claims 10 to 13, characterized in that, The method further includes: Based on the target segmentation mask and the target medical image, a target segmentation image corresponding to the target segmentation region is generated.

15. An image segmentation apparatus, characterized in that, The device includes: The second determining module is used to determine the major axis label and minor axis label corresponding to the sample segmentation region based on the sample diameter label corresponding to the sample segmentation region in the sample medical image. The major axis label is the label of the longest diameter corresponding to the sample segmentation region, and the minor axis label is the label of the shortest diameter corresponding to the sample segmentation region. Based on the major axis label and the minor axis label, an undersegmentation mask corresponding to the sample medical image is determined, wherein the undersegmentation region corresponding to the undersegmentation mask is smaller than the sample segmentation region; Based on the major axis annotation, an oversegmentation mask corresponding to the sample medical image is determined, wherein the oversegmentation region corresponding to the oversegmentation mask is larger than the sample segmentation region; The first training module is used to train a first image segmentation network based on the sample medical image and the undersegmentation mask. The first image segmentation network is used to predict the undersegmentation region of the sample medical image. The second training module is used to train a second image segmentation network based on the sample medical image and the oversegmentation mask. The second image segmentation network is used to predict oversegmentation regions in the sample medical image.

16. An image segmentation apparatus, characterized in that, The device includes: The first segmentation prediction module is used to predict the under-segmented region of the target segmentation region in the target medical image through the first image segmentation network to obtain a first target probability image. The first target probability image contains the probability value of each pixel in the target medical image belonging to the under-segmented region, and the under-segmented region is smaller than the target segmentation region. The second segmentation prediction module is used to predict oversegmentation regions in the target segmentation region of the target medical image through a second image segmentation network to obtain a second target probability image. The second target probability image contains the probability that each pixel in the target medical image belongs to an oversegmentation region, and the oversegmentation region is larger than the target segmentation region. The first determining module is used to determine the target segmentation mask corresponding to the target medical image based on the first target probability image and the second target probability image; The first image segmentation network and the second image segmentation network are trained using the apparatus described in claim 15.

17. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one program, the at least one program being loaded and executed by the processor to implement the image segmentation method as described in any one of claims 1 to 9, or to implement the image segmentation method as described in any one of claims 10 to 14.

18. A computer-readable storage medium, characterized in that, The readable storage medium stores at least one program, which is loaded and executed by a processor to implement the image segmentation method as described in any one of claims 1 to 9, or to implement the image segmentation method as described in any one of claims 10 to 14.

19. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the image segmentation method as described in any one of claims 1 to 9, or to implement the image segmentation method as described in any one of claims 10 to 14.

Images