Medical imaging-based model training method, device and storage medium

Abstract

The present application discloses a model training method, device and storage medium based on medical images, which relate to the field of medical technology. The method includes: obtaining a medical image, wherein the medical image includes L outlined areas, and L is an integer greater than 1; determining a target outlined area from the L outlined areas according to the size and texture shape of each outlined area; slicing the medical image according to the target outlined area to obtain multiple sliced ​​images corresponding to the medical image; iteratively training the neural network according to the sliced ​​images to obtain a target neural network model. The present application solves the technical problem in the prior art that the amount of data for model training based on medical images on a single prediction task is insufficient, resulting in poor prediction effect of the trained model.

Description

Technical Field

[0001] The present application relates to the field of medical technology, and in particular to a model training method, device and storage medium based on medical imaging. Background Art

[0002] Medical images, such as CT images, MR images, PET images, etc., often exhibit the characteristics of high dimensionality and small samples. When it comes to classification tasks, such as efficacy prediction, the high dimensionality of medical images usually requires complex deep neural network models to process these images. This is because deep neural network models have more parameters and can effectively process higher dimensional input images.

[0003] However, a deep neural network model with more parameters requires more training samples to ensure the model's fitting effect. However, due to the protection of patient privacy data, it is difficult to obtain a large number of medical images. Therefore, if a deep neural network model is trained only with a small number of medical images, the prediction effect of the trained model is often poor.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The present application provides a model training method, device and storage medium based on medical images, so as to at least solve the technical problem in the prior art that insufficient data is used for model training based on medical images for a single prediction task, resulting in poor prediction effect of the trained model.

[0006] According to one aspect of the present application, a model training method based on medical images is provided, comprising: acquiring a medical image, wherein the medical image includes L outlined areas, and L is an integer greater than 1; determining a target outlined area from the L outlined areas according to the size and texture shape of each outlined area; slicing the medical image according to the target outlined area to obtain multiple sliced ​​images corresponding to the medical image; iteratively training a neural network according to the sliced ​​images to obtain a target neural network model.

[0007] Optionally, the number of medical images is M, wherein the M medical images include medical images involving different diseases, different organ sites and different medical applications.

[0008] Optionally, the model training method based on medical images also includes: obtaining a preset size of the slice image; detecting whether the size of each outlined area is smaller than the preset size to obtain a first detection result; detecting whether the texture shape of each outlined area is a meaningful target texture shape to obtain a second detection result; wherein the meaningful target texture shape presents regular texture changes; and determining the target outlined area from the L outlined areas based on the first detection result and the second detection result.

[0009] Optionally, the model training method based on medical images also includes: determining N candidate outline areas from L outline areas based on the first detection result, wherein N is a positive integer less than or equal to L, and the size of the candidate outline area is less than a preset size; detecting whether there are multiple outline areas among the N candidate outline areas whose overlap exceeds a preset threshold, and obtaining a third detection result; and determining a target outline area from the N candidate outline areas based on the second detection result and the third detection result.

[0010] Optionally, the model training method based on medical images also includes: when the third detection result indicates that there are no multiple outline areas with an overlap exceeding a preset threshold among the N alternative outline areas, the alternative outline area including the target texture shape among the N alternative outline areas is used as the target outline area according to the second detection result; when the third detection result indicates that there are alternative outline areas with an overlap exceeding a preset threshold among the N alternative outline areas, a filtering operation is performed on the multiple outline areas with an overlap exceeding the preset threshold among the N alternative outline areas, wherein the filtering operation is used to retain only the smallest size outline area among the multiple outline areas with an overlap exceeding the preset threshold; and the alternative outline area including the target texture shape among all the alternative outline areas remaining after the filtering operation is completed is used as the target outline area.

[0011] Optionally, the model training method based on medical images also includes: when the medical image is a single-modality image, slicing the medical image according to a first slicing strategy to obtain multiple slice images corresponding to the medical image, wherein the first slicing strategy is used to constrain the image overlap between the multiple slice images obtained by the slicing operation to be lower than a first threshold, and each slice image needs to include a partial area of ​​the target outline area.

[0012] Optionally, the model training method based on medical images also includes: when the medical image is a multimodal image, slicing the medical image according to a first slicing strategy and a second slicing strategy to obtain multiple slice images corresponding to the medical image, wherein the second slicing strategy is used to determine K slice images based on the i-th modality image of the medical image, and to determine H slice images based on the j-th modality image of the medical image, wherein the i-th modality image and the j-th modality image are any two different modality images of the medical image, and K and H are both integers greater than or equal to 1.

[0013] Optionally, the model training method based on medical images also includes: performing data enhancement operations on slice images; encoding the slice images after the data enhancement operations through an encoder in a neural network to obtain a coding vector corresponding to the slice image, wherein the coding vector is used to represent the image features extracted by the encoder from the slice image; determining a prediction vector corresponding to the slice image based on the coding vector through a decoder in the neural network, wherein the prediction vector represents prediction information generated by the decoder based on the coding vector; determining a loss function value of the neural network based on the coding vector and prediction vector corresponding to each slice image; and iteratively training the encoder and decoder of the neural network based on the loss function value to obtain a target neural network model.

[0014] Optionally, the model training method based on medical images also includes: performing multiple target operations on multiple slice images corresponding to the medical image according to the encoding vector and the prediction vector corresponding to each slice image, wherein the target operation is used to determine the common feature information and difference information between different slice images; and determining the loss function value of the neural network according to the operation results of the multiple target operations.

[0015] Optionally, the target operation includes the following steps: selecting a slice image from multiple slice images corresponding to the medical image as the first slice image; using each slice image except the first slice image from the multiple slice images corresponding to the medical image as the second slice image; calculating the maximized negative cosine similarity between the prediction vector of the first slice image and the encoding vector corresponding to each second slice image; and characterizing the common feature information and difference information between the first slice image and each second slice image by maximizing the negative cosine similarity.

[0016] Optionally, the model training method based on medical images also includes: when the medical image is a single-modality image, multiple target operations are used to determine the common feature information and difference information between different slice images in the single-modality image; when the medical image is a multi-modality image, multiple target operations are used to determine the common feature information and difference information between at least two slice images corresponding to the same modality image, and the common feature information and difference information between at least two slice images corresponding to different modality images.

[0017] Optionally, the data enhancement operation includes at least one of the following operations: an image flipping operation, which is used to flip the slice image at a random angle; an image rotation operation, which is used to rotate the slice image at a random angle; an image noise operation, which is used to add or filter noise information to the slice image; an image blurring operation, which is used to reduce the clarity of the slice image.

[0018] According to another aspect of the present application, a model training device based on medical images is also provided, which includes: an acquisition unit, used to acquire a medical image, wherein the medical image includes L outlined areas, and L is an integer greater than 1; a determination unit, used to determine a target outlined area from the L outlined areas according to the size and texture shape of each outlined area; a slicing operation unit, used to perform a slicing operation on the medical image according to the target outlined area to obtain multiple slice images corresponding to the medical image; and a training unit, used to iteratively train a neural network according to the slice images to obtain a target neural network model.

[0019] According to another aspect of the present application, an electronic device is also provided, wherein the electronic device includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors execute any one of the above-mentioned medical image-based model training methods.

[0020] According to another aspect of the present application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer-readable storage medium, wherein when the computer program is executed, the device where the computer-readable storage medium is located executes any one of the above-mentioned medical image-based model training methods.

[0021] In the present application, a method of determining multiple slice images based on the outlined areas obtained by screening is adopted. First, a medical image is obtained, wherein the medical image includes L outlined areas, and L is an integer greater than 1. Then, a target outlined area is determined from the L outlined areas according to the size and texture shape of each outlined area, and a slice operation is performed on the medical image according to the target outlined area to obtain multiple slice images corresponding to the medical image. Finally, the neural network is iteratively trained according to the sliced ​​images to obtain a target neural network model.

[0022] From the above content, it can be seen that the present application first screens the ROI (Region of Interest) area (corresponding to the above-mentioned outlined area) according to the size and texture shape of the outlined area to obtain the target outlined area, and then slices the medical image based on the screened target outlined area as a guide, thereby achieving the purpose of obtaining a sufficient number of slice images using a contrast learning algorithm based on the characteristic of the medical image having an ROI area, and then training a neural network model with a better fitting effect at a faster training speed, solving the technical problem in the prior art of insufficient data for model training based on medical images on a single prediction task, resulting in poor prediction effect of the trained model. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0024] Figure 1 is a flowchart of an optional medical image-based model training method according to an embodiment of the present application;

[0025] Figure 2 is a flowchart of an optional target operation according to an embodiment of the present application;

[0026] Figure 3 is a schematic diagram of an optional execution process of a target operation for a single-modality image according to an embodiment of the present application;

[0027] Figure 4 is a schematic diagram of an optional execution process of a target operation for a multimodal image according to an embodiment of the present application;

[0028] Figure 5 It is a schematic diagram of an optional medical image-based model training device according to an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0030] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] It should also be noted that the information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) collected by this application are information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of relevant data are in compliance with the relevant laws, regulations and standards of the relevant regions, necessary confidentiality measures are taken, and public order and good customs are not violated, and corresponding operation portals are provided for users to choose to authorize or refuse. For example, an interface is set up between this system and relevant users or institutions. Before obtaining relevant information, it is necessary to send an acquisition request to the aforementioned user or institution through the interface, and obtain relevant information after receiving the consent information fed back by the aforementioned user or institution.

[0032] According to an embodiment of the present application, an embodiment of a model training method based on medical images is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0033] In an optional embodiment, a medical image processing system can be used as the execution subject of the model training method for medical images in the embodiment of the present application, wherein the medical image processing system can be a software system or an embedded system combining software and hardware. In addition, those skilled in the art can understand that in addition to using the above-mentioned medical image processing system as the execution subject, the execution subject of the model training method based on medical images in the present application can also be other devices or equipment, and the embodiment of the present application does not specifically limit this. For the convenience of description, the medical image processing system is used as the execution subject below.

[0034] Figure 1is a flowchart of an optional medical image-based model training method according to an embodiment of the present application, such as Figure 1 As shown, the method comprises the following steps:

[0035] Step S101, obtaining medical images.

[0036] In step S101, the medical image includes L delineated regions, where L is an integer greater than 1. The medical image may be a single-modality image or a multi-modality image, and the medical image may be various types of medical images such as an MR image, a CT image, and a PET image. In addition, the delineated region may also be referred to as an ROI region, and the delineated region is used to characterize the contour region information of the target area and / or the organ at risk on the medical image.

[0037] Step S102 : determining a target delineation region from the L delineation regions according to the size and texture shape of each delineation region.

[0038] Optionally, there may be many outlined areas on the medical image, but not every outlined area may be suitable for training the model. Therefore, in order to improve the training efficiency of the model, the present application provides two dimensions of the size and texture shape of the outlined area as screening criteria for the outlined area.

[0039] Step S103 , performing a slicing operation on the medical image according to the target outlined area to obtain a plurality of slice images corresponding to the medical image.

[0040] Optionally, after obtaining the target outlined area, the medical image processing system may use the target outlined area as guidance information and adopt a contrastive learning algorithm to perform a slice operation on the medical image, thereby obtaining a plurality of slice images corresponding to the medical image.

[0041] Optionally, the contrastive learning algorithm is an algorithm mainly applied to natural images, which is characterized by randomly cutting multiple smaller image slices on the same natural image, and then training the model to recognize whether the two slices belong to the same image.

[0042] However, there is a flaw in the contrastive learning algorithm when applied to natural images. For example, there is an object in a natural image that the model is expected to recognize (such as a kitten), but the randomly cut image slices may all cut into the background area and not the area where the kitten is located. In this case, using image slices containing only the background area to train the model will be very inefficient. A more effective method is to ensure that each image slice contains the kitten or part of the kitten.

[0043] Medical imaging has advantages in overcoming the above-mentioned defects, because there are many ROI areas of target areas and organs at risk on medical images, so some meaningful ROI areas can be screened out to guide the cutting of slices. Target areas and ROI areas of organs at risk exist in large quantities in medical images and can be easily obtained, and when screening ROI areas, there is no need to have high requirements on the pixels of the ROI area, only the target area and organs at risk need to be roughly identified. This is because the purpose of screening the ROI area is to guide the generation of slice images containing meaningful ROI areas during the slice operation. Therefore, the present application can even use a segmentation model to automatically generate a delineated area without the need for manual high-precision delineation. Among them, the segmentation model is a pre-trained neural network model that can realize the delineation of target areas and organs at risk in medical images.

[0044] Step S104, iteratively training the neural network according to the slice images to obtain a target neural network model.

[0045] In step S104, the slice image can be used as model training data. In addition, the model training label can be set according to the actual purpose of the target neural network model, wherein the target neural network model can be a model for predicting and generating a patient's radiotherapy plan, or a model for predicting the patient's survival rate after receiving the radiotherapy plan, or a model for predicting the patient's side effect information after receiving the radiotherapy plan. In short, after obtaining the slice image, various neural network models for performing downstream prediction tasks can be obtained based on the slice image training.

[0046] Based on the contents of the above steps S101 to S104, it can be known that in the present application, a method of determining multiple slice images based on the outlined areas obtained by screening is adopted, and a medical image is first obtained, wherein the medical image includes L outlined areas, and L is an integer greater than 1. Then, a target outlined area is determined from the L outlined areas according to the size and texture shape of each outlined area, and a slice operation is performed on the medical image according to the target outlined area to obtain multiple slice images corresponding to the medical image. Finally, the neural network is iteratively trained according to the sliced ​​images to obtain a target neural network model.

[0047] From the above content, it can be seen that the present application first screens the ROI area (corresponding to the above-mentioned outlined area) according to the size and texture shape of the outlined area to obtain the target outlined area, and then slices the medical image based on the screened target outlined area as a guide, thereby achieving the purpose of obtaining a sufficient number of slice images using a contrast learning algorithm based on the characteristic of the medical image having an ROI area, and then training a neural network model with a better fitting effect at a faster training speed, solving the technical problem in the prior art of insufficient data for model training based on medical images on a single prediction task, resulting in poor prediction effect of the trained model.

[0048] In an optional embodiment, the number of medical images is M, wherein the M medical images include medical images involving different diseases, different organ sites, and different medical applications.

[0049] Optionally, the M medical images may include both single-modality images and multi-modality images. For multi-modality images, rigid registration is also required. It should be noted that one advantage of the present application is that very precise registration images are not required, so rigid registration can be used. Of course, the present application also supports flexible registration of multi-modality images.

[0050] In addition, the present application also supports the integration of a large number of medical images of different diseases, different parts, and different tasks to form a training data set (that is, M medical images include medical images involving different diseases, different organ parts, and different medical applications), and then uses the training data set to train the target neural network model, thereby solving the problem that the training data used in the prior art involves a single disease type and too few samples.

[0051] In an optional embodiment, in order to determine the target delineated area from the L delineated areas according to the size and texture shape of each delineated area, the medical image processing system can obtain a preset size of the slice image, and then detect whether the size of each delineated area is smaller than the preset size to obtain a first detection result; detect whether the texture shape of each delineated area is a meaningful target texture shape to obtain a second detection result; wherein the meaningful target texture shape presents regular texture changes. Finally, the medical image processing system determines the target delineated area from the L delineated areas according to the first detection result and the second detection result.

[0052] Optionally, the user can set a preset size of the slice image they want to obtain in the medical image processing system. For example, for 3D slice images, the preset size can be set to "128 x 128 x 16"; for 2D slice images, the preset size can be set to "128 x 128".

[0053] In addition, the user can also set texture shape positive samples and texture shape negative samples in the medical image processing system, where the texture shape positive samples are texture shapes with clear meanings that the user has determined, and the texture shape negative samples are texture shapes without clear meanings that the user has determined.

[0054] Optionally, a first method of determining whether the texture shape of the outlined area is a meaningful target texture shape may be as follows:

[0055] The medical image processing system can extract the texture shape of each outlined area. If the similarity between the texture shape of the outlined area and the texture shape of the texture shape positive sample is greater than a first similarity, the medical image processing system can determine that the texture shape of the outlined area is a meaningful target texture shape. If the similarity between the texture shape of the outlined area and the texture shape of the texture shape positive sample is less than or equal to the first similarity, the medical image processing system can determine that the texture shape of the outlined area is not a meaningful target texture shape.

[0056] Optionally, a second method for determining whether the texture shape of the outlined area is a meaningful target texture shape may be as follows:

[0057] The medical image processing system can compare the similarity between the texture shape of the outlined area and the texture shape of the texture shape negative sample. If the similarity between the texture shape of the outlined area and the texture shape of the texture shape negative sample is greater than a second similarity, the medical image processing system can determine that the texture shape of the outlined area is not a meaningful target texture shape; if the similarity between the texture shape of the outlined area and the texture shape of the texture shape negative sample is less than or equal to the second similarity, the medical image processing can determine that the texture shape of the outlined area is a meaningful target texture shape.

[0058] In an optional embodiment, the medical image processing system may determine N candidate delineation regions from L delineation regions based on the first detection result, wherein N is a positive integer less than or equal to L, and the size of the candidate delineation region is less than a preset size. Then, the medical image processing system may detect whether there are multiple delineation regions with an overlap exceeding a preset threshold among the N candidate delineation regions, and obtain a third detection result. Finally, the medical image processing system determines a target delineation region from the N candidate delineation regions based on the second detection result and the third detection result.

[0059] Optionally, a principle for screening the target delineated area is as follows:

[0060] The size of the target delineation area cannot exceed the preset size of the slice image, so as to ensure that the slice image can include a sufficiently large target delineation area.

[0061] In an optional embodiment, in the process of determining the target delineation area from N candidate delineation areas according to the second detection result and the third detection result, if the third detection result indicates that there are no multiple delineation areas with an overlap exceeding a preset threshold value among the N candidate delineation areas, the medical image processing system may use the candidate delineation area including the target texture shape among the N candidate delineation areas according to the second detection result as the target delineation area; if the third detection result indicates that there are candidate delineation areas with an overlap exceeding a preset threshold value among the N candidate delineation areas, the medical image processing system may perform a filtering operation on the multiple delineation areas with an overlap exceeding the preset threshold value among the N candidate delineation areas, wherein the filtering operation is used to retain only the smallest delineation area among the multiple delineation areas with an overlap exceeding the preset threshold value. Finally, the medical image processing system uses the candidate delineation area including the target texture shape among all the candidate delineation areas remaining after the filtering operation is completed as the target delineation area.

[0062] Optionally, two additional principles for screening the target delineated area are as follows:

[0063] Principle 1: The target outline area needs to have a meaningful texture shape. If a meaningful texture shape cannot be found in an outline area on the image, the outline area will not be used as the target outline area.

[0064] Principle 2: If the overlap between any two outlined regions exceeds a preset threshold, only the outlined region with the smallest size is retained.

[0065] It should be noted that, through the above process, a small number of target delineated areas with clear meanings can be determined, so that when the slice image is subsequently determined based on the target delineated areas, the quality of the slice image can be guaranteed.

[0066] In an optional embodiment, when the medical image is a single-modality image, the medical image processing system can perform a slice operation on the medical image according to a first slicing strategy to obtain multiple slice images corresponding to the medical image, wherein the first slicing strategy is used to constrain the image overlap between the multiple slice images obtained by the slicing operation to be lower than a first threshold, and each slice image needs to include a partial area of ​​the target outline area.

[0067] Optionally, the slicing operation can be performed multiple times. When the medical image is a single-modality image, each slicing operation can randomly select a target outline area from at least one target outline area corresponding to the medical image, and then randomly crop based on the selected target outline area to obtain two or more slice images, wherein each slice image needs to include most of the selected target outline area, and there should not be too many overlapping parts between the multiple slice images obtained by cutting (that is, the image overlap between the multiple slice images obtained by the slicing operation is lower than a first threshold, and the first threshold can be custom set), thereby ensuring that the slice images are all related to the target outline area and maximizing the diversity of the slice images.

[0068] It should be noted that, in addition to including part of the target outline area, each slice image may also include a part of the background area (i.e., the area other than the outline area in the medical image), so that when using such slice images to train the model, the fitting effect of the model can be improved.

[0069] In an optional embodiment, when the medical image is a multi-modal image, the medical image processing system can perform a slice operation on the medical image according to a first slicing strategy and a second slicing strategy to obtain multiple slice images corresponding to the medical image, wherein the second slicing strategy is used to determine K slice images based on the i-th modality image of the medical image, and to determine H slice images based on the j-th modality image of the medical image, wherein the i-th modality image and the j-th modality image are any two different modality images of the medical image, and K and H are both integers greater than or equal to 1.

[0070] Optionally, the slicing operation can be performed multiple times. When the medical image is a multi-modal image, each time the slicing operation is performed, two modal images (corresponding to the ith modal image and the jth modal image of the medical image) can be randomly selected, and then a target delineated area is randomly selected from at least one target delineated area of ​​the medical image, and a plurality of slice images are randomly cropped according to the selected target delineated area, wherein the number of slice images cropped for each modality image can be the same or different (corresponding to the above-mentioned determination of K slice images based on the ith modality image of the medical image and the determination of H slice images based on the jth modality image of the medical image, K and H can be the same or different), for example, one slice image is cropped for the ith modality image, and two slice images are cropped for the jth modality image.

[0071] In the case where the medical image is a multimodal image, the slicing operation also needs to be performed in accordance with the first slicing strategy, that is, each slice image needs to include most of the selected target outline area, and there should not be too many overlapping parts between the multiple slice images obtained by cutting, so as to ensure that the slice images are all related to the target outline area and to increase the diversity of the slice images as much as possible. In addition, in addition to including part of the target outline area, each slice image can also include a part of the background area (that is, the area outside the outline area in the medical image), so that when using such slice images to train the model, the model fitting effect can be improved.

[0072] In an optional embodiment, the medical image processing system can perform data enhancement operations on the slice image. In addition, the medical image processing system can also encode the slice image after the data enhancement operation through the encoder in the neural network to obtain the encoding vector corresponding to the slice image, wherein the encoding vector is used to represent the image features extracted by the encoder from the slice image; the decoder in the neural network determines the prediction vector corresponding to the slice image based on the encoding vector, wherein the prediction vector represents the prediction information generated by the decoder based on the encoding vector. Finally, the medical image processing system can determine the loss function value of the neural network based on the encoding vector and prediction vector corresponding to each slice image, and iteratively train the encoder and decoder of the neural network based on the loss function value to obtain the target neural network model.

[0073] Optionally, the data enhancement operation includes at least one of the following operations: an image flipping operation, which is used to flip the slice image at a random angle; an image rotation operation, which is used to rotate the slice image at a random angle; an image noise operation, which is used to add or filter noise information to the slice image; an image blurring operation, which is used to reduce the clarity of the slice image.

[0074] Optionally, in the process of determining the loss function value of the neural network according to the encoding vector and the prediction vector corresponding to each slice image, the medical image processing system may perform multiple target operations on multiple slice images corresponding to the medical image according to the encoding vector and the prediction vector corresponding to each slice image, wherein the target operation is used to determine the common feature information and difference information between different slice images. Then, the medical image processing system determines the loss function value of the neural network according to the operation results of the multiple target operations.

[0075] in, Figure 2 is a flowchart of an optional target operation according to an embodiment of the present application, such as Figure 2 As shown, the following steps are included:

[0076] Step S201 : selecting a slice image from a plurality of slice images corresponding to the medical image as a first slice image.

[0077] Step S202: taking each slice image except the first slice image among the multiple slice images corresponding to the medical image as a second slice image.

[0078] Step S203 , calculating the maximum negative cosine similarity between the prediction vector of the first slice image and the encoding vector corresponding to each second slice image.

[0079] Step S204 , characterizing the common feature information and difference information between the first slice image and each second slice image by maximizing the negative cosine similarity.

[0080] Optionally, the following describes various steps of the target operation for single-modality images and multi-modality images respectively.

[0081] Optionally, Figure 3 is a schematic diagram of an optional execution process of a target operation for a single-modality image according to an embodiment of the present application, such as Figure 3 As shown, a target outline region is randomly selected from at least one outline region of a single modality image, two slice images are randomly cropped according to the selected target outline region, and data enhancement is performed on the two slice images respectively. Assume that the two slice images after data enhancement are: x1 and x2, the image features of x1 are encoded by the encoder to obtain the encoding vector y1, and the image features of x2 are also encoded by the encoder to obtain the encoding vector y2.

[0082] Alternatively, if Figure 3 As shown, in the first target operation, x2 is selected as the first slice image, x1 is selected as the second slice image, and the maximum negative cosine similarity D(y1, z2) is calculated, where z2 is the prediction vector determined by the decoder based on y2. The stop-gradient (a property used to control gradient propagation in a deep learning framework) operation is applied on the path that does not pass through the predictor (decoder).

[0083] Optionally, in the second target operation, x1 can be selected as the first slice image, x2 can be selected as the second slice image, and the maximum negative cosine similarity (y2, z1) can be calculated, where z1 is the prediction vector determined by the decoder based on y1. Finally, D(y1, z2) and D(y2, z1) can be combined to perform back propagation, thereby updating the parameters of the encoder and decoder, and completing the iterative training of the encoder and decoder.

[0084] It should be noted that formula (1) shows the calculation process of D(y1, z2).

[0085]

[0086] It should also be noted that formula (2) shows the calculation process of D(y2, z1).

[0087]

[0088] Optionally, Figure 4 is a schematic diagram of an optional execution process of a target operation for a multimodal image according to an embodiment of the present application, such as Figure 4 As shown, two modal images are randomly selected from the multimodal images, a target outline area is randomly selected from at least one outline area of ​​the multimodal images, and three slice images are randomly cropped according to the selected target outline area, wherein two slice images are randomly cropped for the first modal image, and one slice image is randomly cropped for the second modal image, and data enhancement is performed on all three slice images. Assume that the three slice images after data enhancement are: a1, a2, and a3, and the image features of a1 are encoded by the encoder to obtain the encoding vector b1; the image features of a2 are encoded by the encoder to obtain the encoding vector b2; the image features of a3 are encoded by the encoder to obtain the encoding vector b3.

[0089] Alternatively, if Figure 4 As shown, during the execution of the first target operation, a2 can be selected as the first slice image, and a1 and a3 can be selected as the second slice image, wherein the decoder applies a stop-gradient operation on a path that does not pass through the decoder based on the prediction vector c2 determined by the encoding vector b2.

[0090] Optionally, the maximum negative cosine similarity D(b1, c2) and D(b3, c2) are calculated.

[0091] Similarly, during the execution of the second target operation, a1 can be selected as the first slice image, and a2 and a3 can be selected as the second slice image, so that the maximum negative cosine similarity D(b2, c1) and D(b3, c1) can be calculated, where c1 is the prediction vector determined by the decoder based on the encoding vector b1. During the execution of the third target operation, a3 can be selected as the first slice image, and a1 and a3 can be selected as the second slice image, so that the maximum negative cosine similarity D(b2, c3) and D(b1, c3) can be calculated, where c3 is the prediction vector determined by the decoder based on the encoding vector b3.

[0092] It should be noted that, through the above-mentioned training method, when the medical image is a single-modality image, multiple target operations can be used to determine the common feature information and difference information between different slice images in the single-modality image; when the medical image is a multi-modality image, multiple target operations can be used to determine the common feature information and difference information between at least two slice images corresponding to the same modality image, as well as the common feature information and difference information between at least two slice images corresponding to different modality images.

[0093] It should also be noted that the target neural network model trained using the technical solution of the embodiment of the present application has a powerful medical image processing capability and can be used for various downstream tasks. If multimodal medical images are used for training, the trained target neural network model has the processing capability of multimodal medical images, and multimodal medical images can continue to be used for training and prediction in downstream tasks. For example, assuming that the pre-trained model is f(F), and the current three modal images are F1, F2, and F3, using the model f(F), three prediction results f(F1), f(F2), and f(F3) can be obtained. Then, by fusing the three prediction results through an integrated learning method, a more accurate downstream task prediction result can be obtained than a single modality image.

[0094] It should be noted that due to the protection of patient privacy data, it is difficult to obtain a large number of medical images. Therefore, the prediction effect of the trained model is often poor when only a small number of samples of medical images are used to train the deep neural network model. In view of this problem, the deep neural network algorithm has not yet been effectively applied to the classification task of medical images in the medical field. At present, there are three common methods to deal with medical image classification tasks:

[0095] Method 1: By applying radiomics technology, fixed statistical formulas are used to calculate radiomics features from the image ROI area. These radiomics features are then input into a common traditional machine learning model for classification tasks.

[0096] Method 2,The second method uses a deep neural network model pre-trained on a natural image dataset as a feature extractor. The feature extractor is specifically used for prediction of medical images and extracts features of the model output layer. Subsequently, these deep neural network features are input into a general traditional machine learning model for classification tasks.

[0097] Method 3: This method directly trains deep neural networks on small sample medical image data.

[0098] However, among the three methods mentioned above, the third method usually leads to a lower prediction accuracy of the model due to the small number of training samples. Specifically, an important component of a deep neural network is the operator for extracting features (such as the convolution kernel in a convolutional neural network). A large number of convolution kernels act on the input image to extract different features (texture, shape, etc. in the image), and the features extracted by these convolution kernels of the convolutional neural network before model training are generally random and invalid. It relies on a large number of samples for training and optimizing the convolution kernel parameters to finally extract effective features. When there are enough training samples, the deep neural network extracts features through adaptive training iterations, which have a higher upper limit than the features calculated by the imaging omics technology used in the above method 1 using a fixed statistical formula. However, the insufficient number of training samples has prevented the advantages of deep neural networks from being reflected.

[0099] In addition, the first and second methods are only stopgap measures and cannot fully tap the potential of deep neural network algorithms in medical image classification tasks. Instead, they use traditional machine learning models that require fewer samples. However, it is precisely because of the use of these general machine learning models that the upper limit of these methods is limited to a low level and there is limited room for improvement. Specifically, both method 1 and method 2 ultimately use general traditional machine learning models for modeling, and the elasticity and scalability of the methods are poor, and they fail to take advantage of deep neural network algorithms. Secondly, method 2 uses a deep neural network model pre-trained on a natural image dataset as a feature extractor. However, natural images and medical images have essential differences in data characteristics. For example, natural images are RGB three-channel images, while medical images are gray and white. Natural images are 2D, while medical images are 3D. Therefore, the effectiveness of the features extracted in the first step of method 2 is questionable and the feature redundancy is also very large.

[0100] However, the technical solution of the present application can overcome the technical defects existing in the above three methods. Specifically, compared with the first method, the present application uses a large number of medical images to pre-train the neural network through a comparative learning paradigm. The neural network trained by this method has excellent medical image feature extraction capabilities and can be used for various downstream tasks, such as image classification, image segmentation, etc. Compared with the second method, the present application directly trains the deep neural network on medical image data, and the model can be further fine-tuned on downstream tasks, avoiding the direct use of the model trained on the natural image data set, thereby avoiding the degradation of model performance due to data modality differences. Compared with the third method, the present application pre-trains the neural network by fusing medical images of multiple modalities, different tasks and different organ parts, thereby effectively solving the problem of lack of samples for single medical image classification tasks.

[0101] According to another aspect of the present application, a model training device based on medical images is also provided, wherein: Figure 5 is a schematic diagram of an optional medical image-based model training device according to an embodiment of the present application, such as Figure 5 As shown, it includes: an acquisition unit 501, a determination unit 502, a slicing operation unit 503 and a training unit 504.

[0102] Optionally, an acquisition unit is used to acquire a medical image, wherein the medical image includes L outlined areas, and L is an integer greater than 1; a determination unit is used to determine a target outlined area from the L outlined areas according to the size and texture shape of each outlined area; a slicing operation unit is used to perform a slicing operation on the medical image according to the target outlined area to obtain multiple slice images corresponding to the medical image; and a training unit is used to iteratively train a neural network according to the slice images to obtain a target neural network model.

[0103] Optionally, the number of medical images is M, wherein the M medical images include medical images involving different diseases, different organ sites and different medical applications.

[0104] Optionally, the determination unit 502 includes: a first acquisition subunit, a first detection subunit, a second detection subunit, and a first determination subunit. The first acquisition subunit is used to acquire a preset size of the slice image; the first detection subunit is used to detect whether the size of each delineated area is smaller than the preset size to obtain a first detection result; the second detection subunit is used to detect whether the texture shape of each delineated area is a meaningful target texture shape to obtain a second detection result; wherein the meaningful target texture shape presents a regular texture change; the first determination subunit is used to determine the target delineated area from the L delineated areas according to the first detection result and the second detection result.

[0105] Optionally, the first determination subunit includes: a first determination module, a first detection module, and a second determination module. The first determination module is used to determine N candidate delineation areas from L delineation areas according to the first detection result, wherein N is a positive integer less than or equal to L, and the size of the candidate delineation area is less than a preset size; the first detection module is used to detect whether there are multiple delineation areas with an overlap exceeding a preset threshold in the N candidate delineation areas, and obtain a third detection result; the second determination module is used to determine the target delineation area from the N candidate delineation areas according to the second detection result and the third detection result.

[0106] Optionally, the second determination module includes: a first processing submodule, a second processing submodule and a third processing submodule. The first processing submodule is used to, when the third detection result indicates that there are no multiple delineating areas with an overlap exceeding a preset threshold among the N candidate delineating areas, select the candidate delineating area including the target texture shape among the N candidate delineating areas according to the second detection result as the target delineating area; the second processing submodule is used to, when the third detection result indicates that there are multiple delineating areas with an overlap exceeding a preset threshold among the N candidate delineating areas, perform a filtering operation on the multiple delineating areas with an overlap exceeding the preset threshold among the N candidate delineating areas, wherein the filtering operation is used to retain only the smallest delineating area among the multiple delineating areas with an overlap exceeding the preset threshold; the third processing submodule is used to select the candidate delineating area including the target texture shape among all the candidate delineating areas remaining after the filtering operation is completed as the target delineating area.

[0107] Optionally, the slicing operation unit 503 includes: a first processing subunit, used to perform a slicing operation on the medical image according to a first slicing strategy when the medical image is a single-modality image, to obtain multiple slice images corresponding to the medical image, wherein the first slicing strategy is used to constrain the image overlap between the multiple slice images obtained by the slicing operation to be lower than a first threshold, and each slice image needs to include a partial area of ​​the target outline area.

[0108] Optionally, the slicing operation unit 503 includes: a second processing subunit, used to perform a slicing operation on the medical image according to a first slicing strategy and a second slicing strategy when the medical image is a multi-modal image, to obtain multiple slice images corresponding to the medical image, wherein the second slicing strategy is used to determine K slice images based on the i-th modality image of the medical image, and to determine H slice images based on the j-th modality image of the medical image, wherein the i-th modality image and the j-th modality image are any two different modality images of the medical image, and K and H are both integers greater than or equal to 1.

[0109] Optionally, the training unit 504 includes: a data enhancement subunit, a third processing subunit, a fourth processing subunit, a fifth processing subunit and a training subunit. The data enhancement subunit is used to perform data enhancement operation on the slice image; the third processing subunit is used to encode the slice image after the data enhancement operation through the encoder in the neural network to obtain the encoding vector corresponding to the slice image, wherein the encoding vector is used to represent the image features extracted by the encoder from the slice image; the fourth processing subunit is used to determine the prediction vector corresponding to the slice image according to the encoding vector through the decoder in the neural network, wherein the prediction vector represents the prediction information generated by the decoder according to the encoding vector; the fifth processing subunit is used to determine the loss function value of the neural network according to the encoding vector and the prediction vector corresponding to each slice image; the training subunit is used to iteratively train the encoder and decoder of the neural network according to the loss function value to obtain the target neural network model.

[0110] Optionally, the fifth processing subunit includes: a target operation module and a third determination module. The target operation module is used to perform multiple target operations on multiple slice images corresponding to the medical image according to the encoding vector and the prediction vector corresponding to each slice image, wherein the target operation is used to determine the common feature information and difference information between different slice images; the third determination module is used to determine the loss function value of the neural network according to the operation results of the multiple target operations.

[0111] Optionally, the target operation includes the following steps: selecting a slice image from multiple slice images corresponding to the medical image as the first slice image; using each slice image except the first slice image from the multiple slice images corresponding to the medical image as the second slice image; calculating the maximized negative cosine similarity between the prediction vector of the first slice image and the encoding vector corresponding to each second slice image; and characterizing the common feature information and difference information between the first slice image and each second slice image by maximizing the negative cosine similarity.

[0112] Optionally, when the medical image is a single-modality image, multiple target operations are used to determine the common feature information and difference information between different slice images in the single-modality image; when the medical image is a multi-modality image, multiple target operations are used to determine the common feature information and difference information between at least two slice images corresponding to the same modality image, as well as the common feature information and difference information between at least two slice images corresponding to different modality images.

[0113] Optionally, the data enhancement operation includes at least one of the following operations: an image flipping operation, which is used to flip the slice image at a random angle; an image rotation operation, which is used to rotate the slice image at a random angle; an image noise operation, which is used to add or filter noise information to the slice image; an image blurring operation, which is used to reduce the clarity of the slice image.

[0114] According to another aspect of the present application, an electronic device is also provided, wherein the electronic device includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by one or more processors, the one or more processors execute any one of the above-mentioned medical image-based model training methods.

[0115] According to another aspect of the present application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer-readable storage medium, wherein when the computer program is executed, the device where the computer-readable storage medium is located executes any one of the medical imaging-based model training methods described above.

[0116] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0117] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0118] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0119] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0120] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0121] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc., which can store program code.

[0122] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A model training method based on medical images, characterized in that: include: Acquire a medical image, wherein the medical image includes L delineated areas, where L is an integer greater than 1; determine a target delineated area from the L delineated areas according to a size and a texture shape of each delineated area; Slicing the medical image according to the target delineated area to obtain a plurality of slice images corresponding to the medical image; iteratively training the neural network according to the slice images to obtain a target neural network model; Wherein, determining the target delineation area from the L delineation areas according to the size and texture shape of each delineation area includes: obtaining a preset size of the slice image; detecting whether the size of each delineation area is smaller than the preset size, and obtaining a first detection result; detecting whether the texture shape of each delineation area is a meaningful target texture shape, and obtaining a second detection result; wherein the meaningful target texture shape presents a regular texture change; determining the target delineation area from the L delineation areas according to the first detection result and the second detection result; According to the first detection result and the second detection result, the target outlining area is determined from the L outlining areas, including: according to the first detection result, N candidate outlining areas are determined from the L outlining areas, wherein N is a positive integer less than or equal to L, and the size of the candidate outlining area is less than the preset size; detecting whether there are multiple outlining areas with an overlap exceeding a preset threshold value in the N candidate outlining areas, and obtaining a third detection result; if the third detection result indicates that there are no multiple outlining areas with an overlap exceeding the preset threshold value in the N candidate outlining areas, according to the second detection result, The alternative outlining area that includes the target texture shape among the N alternative outlining areas is taken as the target outlining area; when the third detection result indicates that there are alternative outlining areas whose overlap exceeds a preset threshold among the N alternative outlining areas, a filtering operation is performed on multiple outlining areas whose overlap exceeds the preset threshold among the N alternative outlining areas, wherein the filtering operation is used to retain only the smallest outlining area among the multiple outlining areas whose overlap exceeds the preset threshold; and the alternative outlining area that includes the target texture shape among all the alternative outlining areas remaining after completing the filtering operation is taken as the target outlining area.

2. The model training method based on medical images according to claim 1, characterized in that: The number of the medical images is M, wherein the M medical images include medical images involving different diseases, different organ parts and different medical applications.

3. The model training method based on medical images according to claim 1, characterized in that: Slicing the medical image according to the target delineated area to obtain a plurality of slice images corresponding to the medical image includes: In the case where the medical image is a single-modality image, a slice operation is performed on the medical image according to a first slicing strategy to obtain a plurality of slice images corresponding to the medical image, wherein the first slicing strategy is used to constrain the image overlap between the plurality of slice images obtained by the slicing operation to be lower than a first threshold, and each slice image needs to include a partial area of ​​the target outline area.

4. The medical imaging-based model training method according to claim 3, characterized in that: Slicing the medical image according to the target delineated area to obtain a plurality of slice images corresponding to the medical image includes: In the case where the medical image is a multimodal image, a slicing operation is performed on the medical image according to the first slicing strategy and the second slicing strategy to obtain multiple slice images corresponding to the medical image, wherein the second slicing strategy is used to determine K slice images based on the i-th modality image of the medical image, and to determine H slice images based on the j-th modality image of the medical image, wherein the i-th modality image and the j-th modality image are any two different modality images of the medical image, and K and H are both integers greater than or equal to 1.

5. The medical imaging-based model training method according to claim 1, characterized in that: Iteratively training the neural network according to the slice image to obtain a target neural network model, including: Performing data enhancement operation on the slice image; Encoding the slice image after the data enhancement operation is completed by the encoder in the neural network to obtain a coding vector corresponding to the slice image, wherein the coding vector is used to represent the image features extracted by the encoder from the slice image; Determining, by a decoder in the neural network, a prediction vector corresponding to the slice image according to the encoding vector, wherein the prediction vector represents prediction information generated by the decoder according to the encoding vector; Determining the loss function value of the neural network according to the encoding vector and the prediction vector corresponding to each of the slice images; The encoder and decoder of the neural network are iteratively trained according to the loss function value to obtain the target neural network model.

6. The medical imaging-based model training method according to claim 5, characterized in that: Determining the loss function value of the neural network according to the encoding vector and the prediction vector corresponding to each of the slice images includes: Performing multiple target operations on the multiple slice images corresponding to the medical image according to the encoding vector and the prediction vector corresponding to each slice image, wherein the target operation is used to determine the common feature information and difference information between different slice images; The loss function value of the neural network is determined based on the operation results of multiple target operations.

7. The medical imaging-based model training method according to claim 6, characterized in that: The target operation includes the following steps: Selecting one of the slice images from the plurality of slice images corresponding to the medical image as a first slice image; taking each of the slice images except the first slice image among the plurality of slice images corresponding to the medical image as a second slice image; Calculating the maximum negative cosine similarity between the prediction vector of the first slice image and the encoding vector corresponding to each of the second slice images; The common feature information and difference information between the first slice image and each of the second slice images are characterized by maximizing the negative cosine similarity.

8. The medical imaging-based model training method according to claim 6, characterized in that: Medical imaging-based model training methods also include: In the case where the medical image is a single-modality image, the multiple target operations are used to determine the common feature information and difference information between different slice images in the single-modality image; In the case where the medical image is a multimodal image, multiple target operations are used to determine common feature information and difference information between at least two slice images corresponding to the same modality image, and common feature information and difference information between at least two slice images corresponding to different modality images.

9. The medical imaging-based model training method according to claim 5, characterized in that: The data enhancement operation includes at least one of the following operations: An image flipping operation is used to perform a flipping transformation of the slice image at a random angle; An image rotation operation is used to perform a random angle rotation transformation on the slice image; Image noise operation, used to add or filter noise information to the slice image; The image blurring operation is used to reduce the clarity of the slice image.

10. A model training device based on medical images, characterized in that: include: An acquisition unit, used for acquiring a medical image, wherein the medical image includes L delineated areas, where L is an integer greater than 1; a determination unit, used for determining a target delineated area from the L delineated areas according to a size and a texture shape of each delineated area; a slicing operation unit, configured to perform a slicing operation on the medical image according to the target delineated area to obtain a plurality of slice images corresponding to the medical image; and a training unit, configured to iteratively train the neural network according to the slice images to obtain a target neural network model; The slicing operation unit includes: a first acquisition subunit, used to acquire a preset size of the slice image; a first detection subunit, used to detect whether the size of each delineated area is smaller than the preset size, and obtain a first detection result; a second detection subunit, used to detect whether the texture shape of each delineated area is a meaningful target texture shape, and obtain a second detection result; wherein the meaningful target texture shape presents a regular texture change; a first determination subunit, used to determine the target delineated area from the L delineated areas according to the first detection result and the second detection result; The first determination subunit includes: a first determination module, used to determine N candidate delineation areas from the L delineation areas according to the first detection result, wherein N is a positive integer less than or equal to L, and the size of the candidate delineation area is less than the preset size; a first detection module, used to detect whether there are multiple delineation areas with an overlap exceeding a preset threshold among the N candidate delineation areas, and obtain a third detection result; a second determination module, used to determine the target delineation area from the N candidate delineation areas according to the second detection result and the third detection result; The second determination module includes: a first processing submodule, which is used to, when the third detection result indicates that there are no multiple outlining areas with an overlap degree exceeding a preset threshold among the N alternative outlining areas, take the alternative outlining area including the target texture shape among the N alternative outlining areas as the target outlining area according to the second detection result; a second processing submodule, which is used to, when the third detection result indicates that there are alternative outlining areas with an overlap degree exceeding a preset threshold among the N alternative outlining areas, perform a filtering operation on the multiple outlining areas with an overlap degree exceeding a preset threshold among the N alternative outlining areas, wherein the filtering operation is used to retain only the smallest outlining area among the multiple outlining areas with an overlap degree exceeding the preset threshold; and a third processing submodule, which is used to take the alternative outlining area including the target texture shape among all the alternative outlining areas remaining after completing the filtering operation as the target outlining area.

11. An electronic device, characterized in that: It includes one or more processors and a memory, wherein the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the medical image-based model training method described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located executes the medical image-based model training method according to any one of claims 1 to 9.

Images