Image Prediction Model Generation Method, Device, Computer Equipment and Storage Medium

Abstract

The present application relates to an image prediction model generation method, device, computer device, and storage medium. The method includes: obtaining a pre-trained multi-view image generation model; obtaining a first training sample, inputting the non-target side medical image of the first sample object into the pre-trained multi-view image generation model to obtain the target side prediction image of the first sample object; performing self-supervised training on the feature extraction model to be trained according to the target side medical image and the target side prediction image of the first sample object to obtain a pre-trained feature extraction model; obtaining a second training sample, using the target side medical image of the second sample object as input and the target side annotation image of the second sample object as supervision information to train the region prediction model to be trained to obtain a pre-trained region prediction model. By using this method, self-supervised adversarial training can be performed based on multi-view images, improving the training efficiency and training accuracy, and achieving an accurate breast lesion segmentation effect.

Description

Technical Field

[0001] The present application relates to the field of computer technologies, and particularly to a method, apparatus, computer device, storage medium, and computer program product for generating an image prediction model. Background Art

[0002] Currently, for the diagnosis of breast cancer, most detections are carried out using the imaging mode of mammography screening. Mammogram images are high-resolution images obtained based on low-energy X-rays, which can detect abnormal lesions or tumors covered by overlapping breast tissues. Therefore, accurate auxiliary diagnosis of mammogram images is very necessary for improving mammography screening.

[0003] In traditional methods, radiologists usually manually delineate the regions of interest in mammogram images, which is time-consuming and laborious and lacks consistency; or traditional computer-aided systems are used for mammogram image segmentation. However, due to the lack of unity in lesion positions, lesion shapes, and appearances, this method will result in inaccurate segmentation results, prone to a large number of false positives, and due to the large difference in the pixel ranges occupied by different lesions, it is also impossible to accurately segment different lesions adaptively. At the same time, in mammogram images, the image information in different perspectives varies greatly.

[0004] Therefore, there is a problem of poor accuracy in mammogram image segmentation in related technologies. Summary of the Invention

[0005] Based on this, it is necessary to provide a method, apparatus, computer device, storage medium, and computer program product for generating an image prediction model that can solve the above problems in view of the above technical problems.

[0006] In a first aspect, the present application provides a method for generating an image prediction model, the method comprising:

[0007] Obtaining a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a predicted image corresponding to the other side according to a medical image corresponding to any side of a preset object;

[0008] Obtaining a first training sample, and inputting the medical image of the non-target side of the first sample object into the pre-trained multi-view image generation model to obtain the predicted image of the target side of the first sample object;

[0009] Performing self-supervised training on a feature extraction model to be trained according to the medical image of the target side and the predicted image of the target side of the first sample object to obtain a pre-trained feature extraction model;

[0010] Obtain a second training sample, use the target-side medical image of the second sample object as input, and the target-side annotation image of the second sample object as supervision information to train the region prediction model to be trained, and obtain a pre-trained region prediction model; the region prediction model to be trained is constructed based on the pre-trained feature extraction model.

[0011] In one embodiment, the obtaining of the pre-trained multi-view image generation model includes:

[0012] Obtain a third training sample; the third training sample includes the medical images corresponding to the two sides of the preset object respectively.

[0013] According to the medical image corresponding to any one side of the preset object, obtain the predicted image corresponding to the other side of the preset object; the medical image corresponding to any one side and the predicted image corresponding to the other side are images under the same view.

[0014] For any one side of the preset object, according to at least two medical images corresponding to the one side under different views, obtain the image features corresponding to the at least two medical images respectively.

[0015] Use the image consistency loss value and the view feature transformation loss value to train the multi-view image generation model to be trained, and obtain the pre-trained multi-view image generation model; the image consistency loss value is the loss value obtained based on the medical image and the predicted image corresponding to the same side under the same view, and the view feature transformation loss value is the loss value obtained based on the image features corresponding to the at least two medical images corresponding to the same side under different views.

[0016] In one embodiment, the self-supervised training of the feature extraction model to be trained according to the target-side medical image and the target-side predicted image of the first sample object to obtain the pre-trained feature extraction model includes:

[0017] Generate negative sample data according to the target-side predicted image of the first sample object; there are lesions in the target side of the first sample object, and there are no lesions in the non-target side of the first sample object.

[0018] Generate positive sample data according to the target-side medical image of the first sample object.

[0019] Determine the positive sample features corresponding to the positive sample data, and determine the negative sample features corresponding to the negative sample data.

[0020] The feature extraction model to be trained is trained using the feature consistency loss value and the feature contrast loss value to obtain the pre-trained feature extraction model; the feature consistency loss value includes the loss value obtained based on multiple positive sample features and the loss value obtained based on multiple negative sample features, and the feature contrast loss value is the loss value obtained according to the positive sample features and the negative sample features on the same side under the same perspective.

[0021] In one embodiment, the region detection model to be trained includes the target encoder in the pre-trained feature extraction model and a target decoder to be trained connected to the output position of the target encoder. Using the target-side medical image of the second sample object as the input and the target-side annotated image of the second sample object as the supervision information, the region prediction model to be trained is trained to obtain the pre-trained region prediction model, including:

[0022] The target-side medical image of the second sample object is input into the target encoder in the region detection model to be trained, and the output result is input into the target decoder to be trained to obtain a sample prediction image;

[0023] Based on the difference between the target-side annotated image of the second sample object and the sample prediction image, the target decoder to be trained is trained to obtain a trained target decoder, and the pre-trained region prediction model is obtained according to the trained target decoder.

[0024] In one embodiment, the region detection model to be trained includes the target encoder in the pre-trained feature extraction model and a target classifier to be trained connected to the output position of the target encoder. Using the target-side medical image of the second sample object as the input and the target-side annotated image of the second sample object as the supervision information, the region prediction model to be trained is trained to obtain the pre-trained region prediction model, including:

[0025] The target-side medical image of the second sample object is input into the target encoder in the region detection model to be trained, and the output result is input into the target classifier to be trained to obtain a sample classification result;

[0026] Based on the difference between the reference classification result corresponding to the target-side annotated image of the second sample object and the sample classification result, the target classifier to be trained is trained to obtain a trained target classifier, and the pre-trained region prediction model is obtained according to the trained target classifier.

[0027] In one embodiment, after the step of obtaining the pre-trained region prediction model, the method further includes:

[0028] Obtain the image to be detected; the image to be detected is a medical image of any side of the object to be detected from any perspective.

[0029] Input the image to be detected into the pre-trained region prediction model to obtain a segmentation prediction result; the segmentation prediction result includes a region prediction image, and the region prediction image is used to represent the prediction probabilities corresponding to different lesion regions.

[0030] In a second aspect, the present application further provides an image prediction model generation device, and the device includes:

[0031] An image generation model acquisition module, configured to acquire a pre-trained multi-perspective image generation model; the pre-trained multi-perspective image generation model is used to generate a prediction image corresponding to the other side according to the medical image corresponding to any side of the preset object.

[0032] An image generation module, configured to acquire a first training sample, and input the medical image of the non-target side of the first sample object into the pre-trained multi-perspective image generation model to obtain the prediction image of the target side of the first sample object.

[0033] A feature extraction model acquisition module, configured to perform self-supervised training on the feature extraction model to be trained according to the medical image of the target side and the prediction image of the target side of the first sample object to obtain a pre-trained feature extraction model.

[0034] A region prediction model training module, configured to acquire a second training sample, use the medical image of the target side of the second sample object as input, and use the labeled image of the target side of the second sample object as supervision information to train the region prediction model to be trained to obtain a pre-trained region prediction model; the region prediction model to be trained is constructed based on the pre-trained feature extraction model.

[0035] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the image prediction model generation method described above are implemented.

[0036] In a fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the image prediction model generation method described above are implemented.

[0037] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the steps of the image prediction model generation method described above are implemented.

[0038] The above method, apparatus, computer device, storage medium, and computer program product for generating an image prediction model obtain a pre-trained multi-view image generation model, which is used to generate a predicted image corresponding to one side of a preset object based on a medical image corresponding to the other side. The multi-view images can be used to generate each other to expand the data set. Then, a first training sample is obtained. The medical image of the non-target side of the first sample object is input into the pre-trained multi-view image generation model to obtain the predicted image of the target side of the first sample object. According to the medical image of the target side and the predicted image of the target side of the first sample object, self-supervised training is performed on the feature extraction model to be trained to obtain a pre-trained feature extraction model. Self-supervised adversarial training can be performed based on multi-view images to further obtain a second training sample. Using the medical image of the target side of the second sample object as the input and the labeled image of the target side of the second sample object as the supervision information, the region prediction model to be trained is trained to obtain a pre-trained region prediction model. The region prediction model to be trained is constructed based on the pre-trained feature extraction model. By expanding the data set through the above method and training with a large number of automatically generated data, the training efficiency and training accuracy can be improved. At the same time, self-supervised adversarial training based on multi-view images can avoid false positive situations. Therefore, a subsequent prediction model is constructed based on the self-supervised adversarial training results, providing more information for the training of the subsequent model, saving training time, improving training accuracy, and achieving an accurate breast lesion segmentation effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic flowchart of a method for generating an image prediction model in an embodiment;

[0040] Figure 2 It is a schematic diagram of training a multi-view image generation model in an embodiment;

[0041] Figure 3 It is a schematic diagram of training a feature extraction model in an embodiment;

[0042] Figure 4 It is a schematic diagram of training a region prediction model in an embodiment;

[0043] Figure 5 It is a schematic flowchart of another method for generating an image prediction model in an embodiment;

[0044] Figure 6 It is a structural block diagram of an apparatus for generating an image prediction model in an embodiment;

[0045] Figure 7 It is an internal structure diagram of a computer device in an embodiment. Detailed implementation manners

[0046] To make the objectives, technical solutions and advantages of the present application clearer and more understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0047] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in the present application are all information and data that have been authorized by the user or fully authorized by all parties; correspondingly, the present application also provides a corresponding user authorization entry for the user to choose to authorize or choose to refuse.

[0048] In one embodiment, as Figure 1 shown, a method for generating an image prediction model is provided. In this embodiment, an example is given where this method is applied to a server. It can be understood that this method can also be applied to a terminal, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0049] Step 101, obtain a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a predicted image corresponding to the other side according to a medical image corresponding to any side of a preset object;

[0050] As an example, the preset object may have two parts on both sides with a symmetric relationship or a similar symmetric relationship, such as the left breast and the right breast in a human object.

[0051] As an example, the medical image may be a mammogram image. Based on this mammogram image, abnormal lesions or tumors covered by overlapping breast tissues can be detected. The medical image corresponding to any side of the preset object may be the left mammogram image or the right mammogram image of a human object. The left mammogram image or the right mammogram image may respectively have images collected from multiple perspectives, such as CC position (head-to-tail position from left to right) images, MLO position (internal and external oblique position from left to right) images, etc.

[0052] Among them, the multi-view image generation model can be trained based on the mammogram images of normal breasts to obtain a pre-trained multi-view image generation model, and its training process can be as Figure 2 shown. For example, the CC position image of the left breast without lesions, the MLO position image of the left breast without lesions, the CC position image of the right breast without lesions, and the MLO position image of the right breast without lesions of the same human object can be used as training samples, that is, the third training samples.

[0053] In practical applications, a pre-trained multi-view image generation model can be obtained. This pre-trained multi-view image generation model can be used to generate a predicted image corresponding to the other side of the breast from the medical image corresponding to any one side of the breast in a human object, such as a real mammogram image, to generate a predicted image corresponding to the other side of the breast from the medical image corresponding to any one side of the breast in a human object, such as an unreal mammogram image. Furthermore, a feature extraction model can be further self-supervised adversarially trained based on the pre-trained multi-view image generation model.

[0054] Step 102: Obtain a first training sample. Input the medical image of the non-target side of the first sample object into the pre-trained multi-view image generation model to obtain the predicted image of the target side of the first sample object.

[0055] Among them, there is a lesion in the target side of the first sample object, and there is no lesion in the non-target side of the first sample object. For example, a human object with a unilateral breast lesion can be used as the first sample object.

[0056] As an example, the first training sample can include the medical image of the target side and the medical image of the non-target side of the first sample object, such as the real mammogram image of the side with the lesion and the real mammogram image of the side without the lesion.

[0057] In a specific implementation, the medical image of the non-target side of the first sample object, such as the real image of the side without the lesion, can be input into the pre-trained multi-view image generation model to obtain the predicted image of the target side of the first sample object, such as the unreal image of the side with the lesion.

[0058] Step 103: Perform self-supervised training on the feature extraction model to be trained according to the medical image of the target side and the predicted image of the target side of the first sample object to obtain a pre-trained feature extraction model.

[0059] After obtaining the predicted image of the target side of the first sample object, self-supervised training can be performed on the feature extraction model to be trained according to the medical image of the target side and the predicted image of the target side of the first sample object, such as the real image of the side with the lesion and the generated unreal image, and then a pre-trained feature extraction model can be obtained.

[0060] In an example, for the feature extraction model, its training process can be as Figure 3 shown. The trained encoders in the pre-trained multi-view image generation model can be used. For example, for the encoder of the left breast and the encoder of the right breast, according to the real CC view and MLO view images (i.e., medical images in multiple views) of the breast without the lesion, generate the unreal CC view and MLO view images (i.e., predicted images) of the breast with the lesion.

[0061] Step 104: Obtain a second training sample. Use the target-side medical image of the second sample object as the input and the target-side annotation image of the second sample object as the supervision information to train the region prediction model to be trained, and obtain a pre-trained region prediction model. The region prediction model to be trained is constructed based on the pre-trained feature extraction model.

[0062] Among them, there are lesions in the target side of the second sample object. For example, a human object with breast lesions can be used as the second sample object.

[0063] As an example, the second training sample may include the target-side medical image and the target-side annotation image of the second sample object, such as the real mammogram image on the side with lesions and the annotation image obtained after lesion area annotation of the real mammogram image.

[0064] In practical applications, a region prediction model to be trained can be constructed based on the pre-trained feature extraction model. By obtaining the second training sample, the target-side medical image of the second sample object can be used as the input, such as the real image on any side with lesions, and the target-side annotation image of the second sample object can be used as the supervision information, such as the annotation image obtained after lesion area annotation of the real image on any side with lesions, to train the region prediction model to be trained, and then a pre-trained region prediction model can be obtained. Its training process can be as Figure 4 shown, so that in the model application stage, the pre-trained region prediction model can be used to achieve accurate segmentation of breast lesions for mammogram images.

[0065] In an example, in the model application stage, the medical image of the side of the object to be detected containing lesions at any perspective, such as the CC view image or the MLO view image of any side of the breast with lesions, can be used as the image to be detected. Then, the image to be detected can be input into the pre-trained region prediction model to obtain a segmentation prediction result. The prediction probabilities corresponding to different lesion regions can be obtained based on the region prediction image in the segmentation prediction result. For example, the output of the segmentation task may include the segmentation probability of breast cancer, the segmentation probability of breast masses, the segmentation probability of breast calcifications, the segmentation probability of breast vascular calcifications, etc., and may also include the segmentation probability of breast nipples, the segmentation probability of breast lymph nodes, etc.

[0066] Compared with the traditional method, the technical solution of this embodiment can accurately segment breast lesions and avoid false positives by pre-training a region prediction model for mammogram segmentation. By generating multi-view images of mammograms, it can not only expand the dataset but also facilitate the complementarity of image information from different perspectives. Then, self-supervised adversarial training can be performed on the predicted images generated from multi-view images and the real medical images to obtain a pre-trained feature extraction model, which provides more information for the training of the subsequent region prediction model, saves training time, improves the accuracy of training results, and can achieve the same training accuracy with less labeled data as using a large amount of labeled data.

[0067] In the above method for generating an image prediction model, by obtaining a pre-trained multi-view image generation model, multi-view images can be mutually generated to expand the dataset. Then, the first training sample is obtained, and the non-target-side medical image of the first sample object is input into the pre-trained multi-view image generation model to obtain the target-side predicted image of the first sample object. Based on the target-side medical image and the target-side predicted image of the first sample object, self-supervised training is performed on the feature extraction model to be trained to obtain a pre-trained feature extraction model. Self-supervised adversarial training can be carried out based on multi-view images to obtain the second training sample. Using the target-side medical image of the second sample object as the input and the target-side labeled image of the second sample object as the supervision information, the region prediction model to be trained is trained to obtain a pre-trained region prediction model. By expanding the dataset through the above method and training with a large number of automatically generated data, the training efficiency and training accuracy can be improved. At the same time, self-supervised adversarial training based on multi-view images can avoid false positives, so as to construct a subsequent prediction model based on the self-supervised adversarial training results, which provides more information for the training of the subsequent model, saves training time, and improves the training accuracy.

[0068] In one embodiment, the obtaining of the pre-trained multi-view image generation model may include the following steps:

[0069] Obtain the third training sample; according to the medical image corresponding to either side of the preset object, obtain the predicted image corresponding to the other side of the preset object; for either side of the preset object, according to at least two medical images corresponding to the side from different perspectives, obtain the image features corresponding to each of the at least two medical images; use the image consistency loss value and the perspective feature transformation loss value to train the multi-view image generation model to be trained to obtain the pre-trained multi-view image generation model.

[0070] Among them, the third training sample may include the medical images corresponding to both sides of the preset object, such as the left mammogram or the right mammogram of a human object.

[0071] As an example, the image consistency loss value can be the loss value obtained based on the corresponding medical image and the predicted image on the same side under the same perspective; the perspective feature transformation loss value can be the loss value obtained based on the image features corresponding to at least two medical images on the same side under different perspectives.

[0072] In practical applications, the mammogram images of normal breasts can be collected to obtain the third training sample, such as mammogram images without lesions. Then, methods such as the threshold method and the random walk method can be used to segment the breast boundary region of the obtained sample mammogram images to determine the breast boundary region in the images, as Figure 2 shown. Furthermore, the network structure of the encoder and the decoder can be adopted to generate the right breast CC view image (i.e., the predicted image corresponding to the other side) according to the left breast CC view image (i.e., the medical image corresponding to any side), that is, the medical image corresponding to any side and the predicted image corresponding to the other side are images under the same perspective. The right breast MLO view image can also be generated according to the left breast MLO view image, or the left breast CC or MLO view image can be generated using the right breast CC or MLO view image.

[0073] In an example, according to at least two medical images corresponding to any side under different perspectives, the image features corresponding to each of the at least two medical images can be obtained. For example, as Figure 2 shown, by inputting the same side breast CC view image and MLO view image into encoder 1 and encoder 2 respectively, the encoded features output by their respective encoders can be obtained. For example, if ResNet50 is used as the encoder, 2048-dimensional features can be obtained, and then the mapper can be used to mutually predict the features under different perspectives.

[0074] Specifically, the mapper can be a multi-layer perceptron (MLP) structure, which can adopt linear (fully connected layer) or non-linear (fully connected layer followed by RELU activation function) mapping, and can convert the 2048-dimensional features output by the encoder into 128-dimensional features. Furthermore, according to the 128-dimensional features of the CC view image, the 128-dimensional features of the MLO view image can be predicted, or according to the 128-dimensional features of the MLO view image, the 128-dimensional features of the CC view image can be predicted.

[0075] In another example, the image consistency loss value and the perspective feature transformation loss value can be used to train the network parameters of the multi-perspective image generation model to be trained, and a pre-trained multi-perspective image generation model can be obtained, such as Figure 2As shown, for the images predicted and generated by a network structure using an encoder and a decoder, a consistency loss function (i.e., using the image consistency loss value) can be used to train network parameters, such as MSE loss, so that the generated predicted images are as close as possible to the corresponding original medical images; for the features in different perspectives predicted by mappers from each other, such as the mutual transformation prediction between MLO-view images and CC-view images, a cross-entropy loss function (i.e., using the perspective feature transformation loss value) can be used to train network parameters, so that the entire network structure can simultaneously learn the mapping relationship from one side of the breast to the other side, as well as the mapping relationship of mutual transformation between the CC view and the MLO view.

[0076] In an alternative embodiment, the encoder in the multi-view image generation model to be trained can be an encoder based on a transformer, an encoder based on a convolutional neural network, or an encoder constructed using a convolutional neural network and a transformer; the decoder in the multi-view image generation model to be trained can be a decoder based on a transformer, a decoder based on a convolutional neural network, or a decoder constructed using a convolutional neural network and a transformer.

[0077] In this embodiment, by obtaining the third training sample, and then according to the medical image corresponding to any side of the preset object, obtaining the predicted image corresponding to the other side of the preset object, for any side of the preset object, according to at least two medical images corresponding to the side in different perspectives, obtaining the image features corresponding to each of the at least two medical images, and then using the image consistency loss value and the perspective feature transformation loss value to train the multi-view image generation model to be trained, obtaining a pre-trained multi-view image generation model, which can generate multi-view images of mammography images from each other, can play a role in expanding the dataset, and provides data support for subsequent model training.

[0078] In an embodiment, the self-supervised training of the feature extraction model to be trained according to the target-side medical image and the target-side predicted image of the first sample object to obtain a pre-trained feature extraction model may include the following steps:

[0079] Generating negative sample data according to the target-side predicted image of the first sample object; generating positive sample data according to the target-side medical image of the first sample object; determining the positive sample features corresponding to the positive sample data and the negative sample features corresponding to the negative sample data; and training the feature extraction model to be trained using the feature consistency loss value and the feature contrast loss value to obtain the pre-trained feature extraction model.

[0080] Among them, there are lesions in the target side of the first sample object, and there are no lesions in the non-target side of the first sample object.

[0081] As an example, the feature consistency loss value can include the loss value obtained based on multiple positive sample features and the loss value obtained based on multiple negative sample features. The feature contrast loss value can be the loss value obtained according to the positive sample features and negative sample features belonging to the same side under the same perspective.

[0082] In a specific implementation, mammogram images with lesions on one side can be collected to obtain the first training sample. Then, the obtained sample mammogram images can be subjected to gray-scale normalization. For example, methods such as resampling method, threshold limiting method, and gray-scale histogram normalization method can be used to preprocess the images, and the threshold method can be used to obtain the breast boundary region of the image, such as Figure 3 As shown, the CC-view image and MLO-view image of the breast on the side without lesions (i.e., the non-target side medical images) can be input into a pre-trained multi-view image generation model. For example, using the encoder and decoder network structures trained in the pre-trained multi-view image generation model, the predicted CC-view image and MLO-view image of the breast on the side with lesions (i.e., the target side predicted images) can be obtained.

[0083] In an alternative embodiment, the generated target side predicted images can be segmented into small block images. For example, they can be segmented into 64*64 image blocks, or randomly cropped into 64*64 or 128*128 or 256*256 image blocks. If the server performance is good, the generated target side predicted images can also be directly used without being segmented into small block images, and negative sample data can be obtained through random rotation, flipping, scaling, etc.

[0084] In an example, if the lesion in the target side medical image is large, the target side medical image can be cropped into 64*64 or 128*128 or 256*256 image blocks by using the random cropping method. If the server performance is good, the target side medical image can also be directly used without being segmented into small block images to obtain positive sample data; if the lesion in the target side medical image is small, in order to improve the recognition accuracy for small lesions, the imaging report can also be obtained by using the PACS system (picture archiving and communication system), and the approximate position marked in the report can be used as a soft label to segment the small block images, such as being segmented into 64*64 image blocks, and then positive sample data can be obtained through random rotation, flipping, scaling, etc.

[0085] In another example, the positive sample data and negative sample data can be used as input data to train the feature extraction model to be trained, such as Figure 3The encoder 3 and the mapper 3 in the feature extraction model to be trained can then train the parameters of the encoder 3 to optimize the network. For example, by respectively obtaining the positive sample features and the negative sample features, a contrast loss function (i.e., using the feature contrast loss value), such as InfoNCE loss, can be used between the positive and negative sample features, so that the feature difference between the positive and negative samples belonging to the same side under the same viewing angle can be large; a consistency loss function can be used between the positive sample features, and a consistency loss function (i.e., using the feature consistency loss value), such as L2 loss, can be used between the negative sample features, so that the positive sample features are similar and the negative sample features are similar.

[0086] In this embodiment, negative sample data is generated according to the target side predicted image of the first sample object, positive sample data is generated according to the target side medical image of the first sample object, and then the positive sample features corresponding to the positive sample data are determined, and the negative sample features corresponding to the negative sample data are determined, and then the feature extraction model to be trained is trained using the feature consistency loss value and the feature contrast loss value to obtain a pre-trained feature extraction model, which can perform self-supervised adversarial training on the predicted images generated according to the multi-view images and the real medical images, thereby providing data support for the training of subsequent regional prediction models.

[0087] In one embodiment, the region detection model to be trained may include a target encoder in a pre-trained feature extraction model, and a target decoder to be trained connected to an output position of the target encoder, and the target-side medical image of the second sample object is used as input, and the target-side labeled image of the second sample object is used as supervision information to train the region prediction model to be trained to obtain the pre-trained region prediction model, which may include the following steps:

[0088] The target side medical image of the second sample object is input into the target encoder in the region detection model to be trained, and the output result is input into the target decoder to be trained to obtain a sample prediction image; based on the difference between the target side annotated image of the second sample object and the sample prediction image, the target decoder to be trained is trained to obtain a trained target decoder, and the pre-trained region prediction model is obtained based on the trained target decoder.

[0089] As an example, a breast mammography image with lesions and a gold standard sample (i.e., a target side annotated image) with lesions annotated by a doctor can be collected to obtain a second training sample, wherein the target side annotated image can be obtained by a radiologist using manual outlining, or by automatically, semi-automatically, manually, or other methods to outline the lesion area of the breast mammography image.

[0090] In practical applications, based on the target encoder in the pre-trained feature extraction model, such as Figure 4 the pre-trained encoder 3 in

[0091] At the output position of the target encoder, a target decoder to be trained is connected. The target decoder can be a network structure for the segmentation task. Furthermore, a mammogram image with lesions can be used as the input image (i.e., the target-side medical image of the second sample object) and input into the region detection model to be trained, and an output sample prediction image can be obtained, which can be used to represent the predicted lesion region or the region of interest in the mammogram image.

[0092] In an example, the network parameters of the trained target encoder can be fixed to train the target decoder to be trained, and the network can be fine-tuned to make the segmentation prediction result of the network closest to the labeled gold standard sample (i.e., the target-side labeled image). Furthermore, in the model application stage, the trained region prediction model can be used to perform the segmentation / target detection task.

[0093] In this embodiment, by inputting the target-side medical image of the second sample object into the target encoder in the region detection model to be trained, and inputting the output result into the target decoder to be trained, a sample prediction image is obtained. Furthermore, based on the difference between the target-side labeled image of the second sample object and the sample prediction image, the target decoder to be trained is trained to obtain a trained target decoder, and a pre-trained region prediction model is obtained according to the trained target decoder, which can accurately achieve breast lesion segmentation and avoid false positives.

[0094] In an embodiment, the region detection model to be trained can include the target encoder in the pre-trained feature extraction model and a target classifier to be trained connected to the output position of the target encoder. Using the target-side medical image of the second sample object as the input and the target-side labeled image of the second sample object as the supervision information, the region prediction model to be trained is trained to obtain a pre-trained region prediction model, which can include the following steps:

[0095] Input the target-side medical image of the second sample object into the target encoder in the region detection model to be trained, and input the output result into the target classifier to be trained to obtain a sample classification result; based on the difference between the reference classification result corresponding to the target-side annotation image of the second sample object and the sample classification result, train the target classifier to be trained to obtain a trained target classifier, and obtain the pre-trained region prediction model according to the trained target classifier.

[0096] In one example, it is possible to use the target encoder in the pre-trained feature extraction model, such as Figure 4 the pre-trained encoder 3 therein. Connect the target classifier to be trained at the output position of this target encoder. This target classifier can be used to classify medical tasks for the identified region of interest in the image. Furthermore, a mammogram image with lesions can be used as the input image (i.e., the target-side medical image of the second sample object) and input into the region detection model to be trained, and the output sample classification result can be obtained.

[0097] In another example, it is possible to fix the network parameters of the trained target encoder to train the target classifier to be trained, and by fine-tuning the network, the network can obtain a classification prediction result that is closest to the reference classification result corresponding to the target-side annotation image. Furthermore, in the model application stage, the trained region prediction model can be used to perform classification tasks.

[0098] In this embodiment, by inputting the target-side medical image of the second sample object into the target encoder in the region detection model to be trained, and inputting the output result into the target classifier to be trained to obtain a sample classification result, and then based on the difference between the reference classification result corresponding to the target-side annotation image of the second sample object and the sample classification result, training the target classifier to be trained to obtain a trained target classifier, and obtaining the pre-trained region prediction model according to the trained target classifier, the classification accuracy of breast lesions is improved.

[0099] In one embodiment, after the step of obtaining the pre-trained region prediction model, the following steps may be included:

[0100] Obtain an image to be detected; the image to be detected is a medical image of any side of the object to be detected at any perspective; input the image to be detected into the pre-trained region prediction model to obtain a segmentation prediction result; the segmentation prediction result includes a region prediction image, and the region prediction image is used to represent the prediction probabilities corresponding to different lesion regions.

[0101] Among them, the image to be detected can be a medical image of the side containing the lesion from any perspective, or a medical image of the side without the lesion from any perspective.

[0102] In practical applications, during the model application stage, a medical image of the side containing the lesion of the object to be detected from any perspective, such as the CC view image or MLO view image of any breast side with a lesion, can be used as the image to be detected. Then, this image to be detected can be input into a pre-trained region prediction model to obtain a segmentation prediction result based on the lesion.

[0103] For example, the prediction probabilities corresponding to different lesion regions can be obtained based on the region prediction image in the segmentation prediction result. For example, the output of the segmentation task can include the segmentation probability of breast cancer, the segmentation probability of breast masses, the segmentation probability of breast calcifications, the segmentation probability of breast vascular calcifications, etc., and can also include the segmentation probability of breast nipples, the segmentation probability of breast lymph nodes, etc.

[0104] In one example, a medical image of the side without the lesion from any perspective, such as the CC view image or MLO view image of any breast side without the lesion, can also be used as the image to be detected. Then, this image to be detected can be input into a pre-trained region prediction model to obtain a segmentation prediction result based on the normal breast. For example, the prediction probability corresponding to the lesion region can be 0 or a relatively small value to indicate the absence of a lesion.

[0105] In this embodiment, by obtaining the image to be detected and then inputting the image to be detected into a pre-trained region prediction model to obtain a segmentation prediction result, an accurate breast lesion segmentation effect can be achieved.

[0106] In one embodiment, as Figure 5 shown, a schematic flowchart of another method for generating an image prediction model is provided. In this embodiment, the method includes the following steps:

[0107] In step 501, obtain a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a predicted image corresponding to the other side according to the medical image corresponding to any side of a preset object. In step 502, input the medical image of the non-target side of the first sample object into the pre-trained multi-view image generation model to obtain the predicted image of the target side of the first sample object. In step 503, generate negative sample data according to the predicted image of the target side of the first sample object, and generate positive sample data according to the medical image of the target side of the first sample object. In step 504, determine the positive sample features corresponding to the positive sample data and determine the negative sample features corresponding to the negative sample data. In step 505, train the feature extraction model to be trained using the feature consistency loss value and the feature contrast loss value to obtain a pre-trained feature extraction model. In step 506, obtain a second training sample, use the medical image of the target side of the second sample object as the input, and use the labeled image of the target side of the second sample object as the supervision information to train the region prediction model to be trained to obtain a pre-trained region prediction model. In step 507, obtain the image to be detected; the image to be detected is the medical image of any side of the object to be detected from any perspective. In step 508, input the image to be detected into the pre-trained region prediction model to obtain a segmentation prediction result; the segmentation prediction result includes a region prediction image, and the region prediction image is used to represent the predicted probabilities corresponding to different lesion regions. It should be noted that the specific limitations of the above steps can be referred to the specific limitations of a method for generating an image prediction model described above, and will not be elaborated here.

[0108] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or alternately with at least a part of other steps or steps or stages in other steps.

[0109] Based on the same inventive concept, the embodiments of the present application also provide an image prediction model generation device for implementing the above-mentioned image prediction model generation method. The solution provided by this device to solve the problem is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the image prediction model generation device provided below can be referred to the limitations of the image prediction model generation method described above, and will not be elaborated here.

[0110] In one embodiment, as Figure 6 shown, an image prediction model generation device is provided, including:

[0111] An image generation model acquisition module 601, configured to acquire a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a predicted image corresponding to the other side according to a medical image corresponding to any side of a preset object;

[0112] An image generation module 602, configured to acquire a first training sample, input a medical image of the non-target side of the first sample object into the pre-trained multi-view image generation model, and obtain a predicted image of the target side of the first sample object;

[0113] A feature extraction model acquisition module 603, configured to perform self-supervised training on a feature extraction model to be trained according to the medical image of the target side and the predicted image of the target side of the first sample object, and obtain a pre-trained feature extraction model;

[0114] A region prediction model training module 604, configured to acquire a second training sample, use the medical image of the target side of the second sample object as an input, and use the labeled image of the target side of the second sample object as supervision information to train a region prediction model to be trained, and obtain a pre-trained region prediction model; the region prediction model to be trained is constructed based on the pre-trained feature extraction model.

[0115] In one embodiment, the image generation model acquisition module 601 includes:

[0116] A third training sample acquisition sub-module, configured to acquire a third training sample; the third training sample includes medical images corresponding to each of the two sides of the preset object;

[0117] A predicted image acquisition sub-module, configured to obtain a predicted image corresponding to the other side of the preset object according to a medical image corresponding to any side of the preset object; the medical image corresponding to any side and the predicted image corresponding to the other side are images under the same view;

[0118] An image feature acquisition sub-module, configured to, for any side of the preset object, obtain image features corresponding to at least two medical images corresponding to the side under different views;

[0119] An image generation model training sub-module is used to train a multi-view image generation model to be trained by using an image consistency loss value and a perspective feature transformation loss value, so as to obtain the pre-trained multi-view image generation model; the image consistency loss value is a loss value obtained based on a medical image and a predicted image corresponding to the same side under the same perspective, and the perspective feature transformation loss value is a loss value obtained based on image features corresponding to at least two medical images corresponding to the same side under different perspectives.

[0120] In one embodiment, the feature extraction model obtaining module 603 includes:

[0121] A negative sample data obtaining sub-module is used to generate negative sample data according to the target-side predicted image of the first sample object; there are lesions in the target side of the first sample object, and there are no lesions in the non-target side of the first sample object;

[0122] A positive sample data obtaining sub-module is used to generate positive sample data according to the target-side medical image of the first sample object;

[0123] A sample feature determination sub-module is used to determine the positive sample features corresponding to the positive sample data and the negative sample features corresponding to the negative sample data;

[0124] A feature extraction model training sub-module is used to train a feature extraction model to be trained by using a feature consistency loss value and a feature contrast loss value, so as to obtain the pre-trained feature extraction model; the feature consistency loss value includes a loss value obtained based on multiple positive sample features and a loss value obtained based on multiple negative sample features, and the feature contrast loss value is a loss value obtained according to the positive sample features and the negative sample features belonging to the same side under the same perspective.

[0125] In one embodiment, the region detection model to be trained includes the target encoder in the pre-trained feature extraction model and a target decoder to be trained connected to the output position of the target encoder. The region prediction model training module 603 includes:

[0126] A sample predicted image obtaining sub-module is used to input the target-side medical image of the second sample object into the target encoder in the region detection model to be trained, and input the output result into the target decoder to be trained to obtain a sample predicted image;

[0127] A decoder training sub-module is used to train the target decoder to be trained based on the difference between the target-side labeled image of the second sample object and the sample predicted image, so as to obtain a trained target decoder, and obtain the pre-trained region prediction model according to the trained target decoder.

[0128] In one embodiment, the region detection model to be trained includes a target encoder in the pre-trained feature extraction model and a target classifier to be trained connected to the output position of the target encoder. The region prediction model training module 603 includes:

[0129] A sample classification result obtaining sub-module, configured to input the target-side medical image of the second sample object into the target encoder in the region detection model to be trained, and input the output result into the target classifier to be trained to obtain a sample classification result;

[0130] A classifier training sub-module, configured to train the target classifier to be trained based on the difference between the reference classification result corresponding to the target-side labeled image of the second sample object and the sample classification result, obtain a trained target classifier, and obtain the pre-trained region prediction model according to the trained target classifier.

[0131] In one embodiment, the device further includes:

[0132] A to-be-detected image acquisition module, configured to acquire a to-be-detected image; the to-be-detected image is a medical image of any side of a to-be-detected object at any perspective;

[0133] A segmentation prediction module, configured to input the to-be-detected image into the pre-trained region prediction model to obtain a segmentation prediction result; the segmentation prediction result includes a region prediction image, and the region prediction image is used to represent the prediction probabilities corresponding to different lesion regions.

[0134] Each module in the above image prediction model generation device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in or independent of a processor in a computer device in the form of hardware, or stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to the above modules.

[0135] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 7As shown. The computer device includes a processor, a memory, and a network interface connected by a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store image prediction model generation data. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements an image prediction model generation method.

[0136] Those skilled in the art can understand that Figure 7 the structure shown in is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0137] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory. When the processor executes the computer program, the following steps are implemented:

[0138] Obtain a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a predicted image corresponding to the other side according to the medical image corresponding to any side of the preset object;

[0139] Obtain a first training sample, input the medical image of the non-target side of the first sample object into the pre-trained multi-view image generation model, and obtain the predicted image of the target side of the first sample object;

[0140] According to the medical image of the target side of the first sample object and the predicted image of the target side, perform self-supervised training on the feature extraction model to be trained, and obtain a pre-trained feature extraction model;

[0141] Obtain a second training sample, use the medical image of the target side of the second sample object as the input, and use the labeled image of the target side of the second sample object as the supervision information to train the region prediction model to be trained, and obtain a pre-trained region prediction model; the region prediction model to be trained is constructed based on the pre-trained feature extraction model.

[0142] In one embodiment, when the processor executes the computer program, it also implements the steps of the image prediction model generation method in the above other embodiments.

[0143] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0144] Obtain a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a predicted image corresponding to the other side according to the medical image corresponding to any one side of a preset object;

[0145] Obtain a first training sample, input the medical image of the non-target side of the first sample object into the pre-trained multi-view image generation model, and obtain the predicted image of the target side of the first sample object;

[0146] According to the medical image of the target side of the first sample object and the predicted image of the target side, perform self-supervised training on the feature extraction model to be trained, and obtain a pre-trained feature extraction model;

[0147] Obtain a second training sample, use the medical image of the target side of the second sample object as input, and use the labeled image of the target side of the second sample object as supervision information to train the region prediction model to be trained, and obtain a pre-trained region prediction model; the region prediction model to be trained is constructed based on the pre-trained feature extraction model.

[0148] In one embodiment, when the computer program is executed by a processor, the steps of the image prediction model generation method in the above other embodiments are also implemented.

[0149] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the following steps are implemented:

[0150] Obtain a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a predicted image corresponding to the other side according to the medical image corresponding to any one side of a preset object;

[0151] Obtain a first training sample, input the medical image of the non-target side of the first sample object into the pre-trained multi-view image generation model, and obtain the predicted image of the target side of the first sample object;

[0152] According to the medical image of the target side of the first sample object and the predicted image of the target side, perform self-supervised training on the feature extraction model to be trained, and obtain a pre-trained feature extraction model;

[0153] Obtain a second training sample, use the target-side medical image of the second sample object as input, and the target-side annotation image of the second sample object as supervision information to train the region prediction model to be trained, and obtain a pre-trained region prediction model; the region prediction model to be trained is constructed based on the pre-trained feature extraction model.

[0154] In one embodiment, when the computer program is executed by the processor, it also implements the steps of the image prediction model generation method in the above other embodiments.

[0155] Those of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the method embodiments as described above. Among them, any reference to a memory, database, or other medium provided in the various embodiments of the present application can include at least one of non-volatile and volatile memories. Non-volatile memories can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memories can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the various embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., and are not limited thereto. The processors involved in the various embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., and are not limited thereto.

[0156] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0157] The above-described embodiments merely represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the appended claims.

Claims

1. A method for generating an image prediction model, characterized in that The method includes: Obtaining a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a predicted image corresponding to the other side according to the medical image corresponding to any side of the preset object; Obtaining a first training sample, and inputting the medical image of the non-target side of the first sample object into the pre-trained multi-view image generation model to obtain the predicted image of the target side of the first sample object; Performing self-supervised training on the feature extraction model to be trained according to the medical image of the target side and the predicted image of the target side of the first sample object to obtain a pre-trained feature extraction model; Obtaining a second training sample, using the medical image of the target side of the second sample object as the input, and using the labeled image of the target side of the second sample object as the supervision information to train the region prediction model to be trained to obtain a pre-trained region prediction model; the region prediction model to be trained is constructed based on the pre-trained feature extraction model.

2. The method according to claim 1, characterized in that, The obtaining of the pre-trained multi-view image generation model includes: Obtaining a third training sample; the third training sample includes the medical images corresponding to each of the two sides of the preset object; According to the medical image corresponding to any side of the preset object, obtaining the predicted image corresponding to the other side of the preset object; the medical image corresponding to any side and the predicted image corresponding to the other side are images under the same view; For any side of the preset object, according to at least two medical images corresponding to the side under different views, obtaining the image features corresponding to each of the at least two medical images; Training the multi-view image generation model to be trained by using the image consistency loss value and the view feature transformation loss value to obtain the pre-trained multi-view image generation model; the image consistency loss value is the loss value obtained based on the medical image and the predicted image corresponding to the same side under the same view, and the view feature transformation loss value is the loss value obtained based on the image features corresponding to each of the at least two medical images corresponding to the same side under different views.

3. The method according to claim 1, wherein The performing self-supervised training on the feature extraction model to be trained according to the medical image of the target side and the predicted image of the target side of the first sample object to obtain a pre-trained feature extraction model includes: Generating negative sample data according to the predicted image of the target side of the first sample object; there is a lesion in the target side of the first sample object, and there is no lesion in the non-target side of the first sample object; Generating positive sample data according to the medical image of the target side of the first sample object; Determining the positive sample features corresponding to the positive sample data and determining the negative sample features corresponding to the negative sample data; Training the feature extraction model to be trained by using the feature consistency loss value and the feature contrast loss value to obtain the pre-trained feature extraction model; the feature consistency loss value includes the loss value obtained based on multiple positive sample features and the loss value obtained based on multiple negative sample features, and the feature contrast loss value is the loss value obtained according to the positive sample features and the negative sample features belonging to the same side under the same view.

4. The method according to claim 1, characterized in that, The to-be-trained region detection model includes the target encoder in the pre-trained feature extraction model and a to-be-trained target decoder connected to the output position of the target encoder. Using the target-side medical image of the second sample object as the input and the target-side annotation image of the second sample object as the supervision information, training the to-be-trained region prediction model to obtain a pre-trained region prediction model, including: Input the target-side medical image of the second sample object into the target encoder in the to-be-trained region detection model, and input the output result into the to-be-trained target decoder to obtain a sample prediction image; Based on the difference between the target-side annotation image of the second sample object and the sample prediction image, train the to-be-trained target decoder to obtain a trained target decoder, and obtain the pre-trained region prediction model according to the trained target decoder.

5. The method according to claim 1, wherein The to-be-trained region detection model includes the target encoder in the pre-trained feature extraction model and a to-be-trained target classifier connected to the output position of the target encoder. Using the target-side medical image of the second sample object as the input and the target-side annotation image of the second sample object as the supervision information, training the to-be-trained region prediction model to obtain a pre-trained region prediction model, including: Input the target-side medical image of the second sample object into the target encoder in the to-be-trained region detection model, and input the output result into the to-be-trained target classifier to obtain a sample classification result; Based on the difference between the reference classification result corresponding to the target-side annotation image of the second sample object and the sample classification result, train the to-be-trained target classifier to obtain a trained target classifier, and obtain the pre-trained region prediction model according to the trained target classifier.

6. The method according to any one of claims 1 to 5, characterized in that After the step of obtaining the pre-trained region prediction model, the method further includes: Obtain a to-be-detected image; the to-be-detected image is a medical image of any side of a to-be-detected object at any perspective; Input the to-be-detected image into the pre-trained region prediction model to obtain a segmentation prediction result; the segmentation prediction result includes a region prediction image, and the region prediction image is used to represent the prediction probabilities corresponding to different lesion regions.

7. An image prediction model generation device, characterized in that, The device includes: An image generation model acquisition module, configured to acquire a pre-trained multi-view image generation model; the pre-trained multi-view image generation model is used to generate a prediction image corresponding to the other side according to the medical image corresponding to any side of a preset object; An image generation module, configured to acquire a first training sample, and input the non-target-side medical image of the first sample object into the pre-trained multi-view image generation model to obtain the target-side prediction image of the first sample object; A feature extraction model acquisition module, configured to perform self-supervised training on the to-be-trained feature extraction model according to the target-side medical image and the target-side prediction image of the first sample object to obtain a pre-trained feature extraction model; The region prediction model training module is used to obtain a second training sample, take the target-side medical image of the second sample object as input, and the target-side annotation image of the second sample object as supervision information, and train the region prediction model to be trained to obtain a pre-trained region prediction model; the region prediction model to be trained is constructed based on the pre-trained feature extraction model.

8. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 6.

Images