A training method for a medical image segmentation model, a segmentation method, and related products

Abstract

This application discloses a training method, segmentation method, and related products for a medical image segmentation model. In this application, global features are first learned through an image training model with a higher degree of blur, and then more detailed local features are learned through an image training model with a lower degree of blur. The model training is performed step by step from simple to complex, so as to better identify shadows, noise, false textures, etc. in the image, and avoid the interference of the above details on the accuracy of image segmentation. Compared with medical staff manually performing image segmentation and marking abnormal areas, the use of the trained medical image segmentation model for image segmentation in this solution can improve the efficiency and accuracy of medical image segmentation.

Description

Technical Field

[0001] The present application relates to the field of medical image processing technology, and in particular to a training method, a segmentation method, and related products of a medical image segmentation model. Background Art

[0002] Currently, in the field of medical image processing, medical staff typically manually identify and segment medical images to mark abnormal areas within them. However, the presence of shadows, noise, or false textures from adjacent organs in medical images can necessitate significant effort on the part of medical staff to identify the images, resulting in low image segmentation efficiency and high recognition costs. Furthermore, the subjective nature of medical staff's identification of abnormal areas in images also affects the accuracy of image segmentation. Improving the efficiency and accuracy of medical image segmentation is a key concern for those skilled in the art. Summary of the Invention

[0003] Based on the above problems, the present application provides a training method, segmentation method and related products of a medical image segmentation model to improve the efficiency and accuracy of medical image recognition.

[0004] The embodiments of this application disclose the following technical solutions:

[0005] In a first aspect, the present application provides a method for training a medical image segmentation model. The training method comprises:

[0006] Acquiring a sample medical image and lesion annotation data corresponding to the sample medical image;

[0007] Performing blur processing on the sample medical image to obtain a plurality of blurred images corresponding to the sample medical image; the blurring degrees of the plurality of blurred images are different;

[0008] Dividing the multiple blurred images into multiple training stages according to blur levels, wherein blurred images with relatively high blur levels are divided into earlier training stages among the multiple training stages, and blurred images with relatively low blur levels are divided into later training stages among the multiple training stages;

[0009] When the target model is trained in stages, the parameters of the target model are adjusted in each training stage based on the lesion annotation data and the blurred image corresponding to the training stage until the last training stage of the multiple training stages is completed, thereby obtaining a medical image segmentation model.

[0010] Optionally, adjusting the parameters of the target model based on the lesion annotation data and the blurred image corresponding to the training stage in each training stage includes:

[0011] In each training stage, the corresponding blurred image is used as the input of the target model, and the target model is used to perform image segmentation based on the input blurred image to predict the lesion area in the input blurred image to obtain lesion area prediction data;

[0012] Based on the gap between the lesion area prediction data and the lesion annotation data, the parameters of the target model are adjusted.

[0013] Optionally, the target model has an encoder-decoder structure of a UNet network; a multi-scale fusion module is provided in the encoder-decoder structure, and the multi-scale fusion module is used to split the feature map output by the convolutional layer in the encoder-decoder structure and perform multi-scale fusion to form a new feature map.

[0014] Optionally, the encoder-decoder structure includes a feature extraction structure as an encoder and an upsampling structure as a decoder;

[0015] The feature extraction structure includes five feature extraction modules, specifically a first feature extraction module, a second feature extraction module, a third feature extraction module, a fourth feature extraction module and a fifth feature extraction module connected in sequence;

[0016] The upsampling structure includes four upsampling modules, specifically a first upsampling module, a second upsampling module, a third upsampling module and a fourth upsampling module connected in sequence;

[0017] The first feature extraction module, the second feature extraction module, the third feature extraction module, and the fourth feature extraction module each include a first convolution layer, a first multi-scale fusion module, a second convolution layer, a second multi-scale fusion module, and a maximum pooling layer connected in sequence; the fifth feature extraction module includes a convolution layer and a multi-scale fusion module; among the five feature extraction modules, the maximum pooling layer of the previous feature extraction module is connected to the first convolution layer of the next feature extraction module;

[0018] The first upsampling module, the second upsampling module and the third upsampling module each include a deconvolution layer, a feature splicing layer, a third convolution layer, a third multi-scale fusion module, a fourth convolution layer and a fourth multi-scale fusion module; the fourth upsampling module includes a deconvolution layer, a feature splicing layer, a third convolution layer, a third multi-scale fusion module, a fourth convolution layer and a fourth multi-scale fusion module and a fifth convolution layer; among the four upsampling modules, the fourth multi-scale fusion module of the previous upsampling module is connected to the deconvolution layer of the subsequent upsampling module;

[0019] The feature splicing layer of the first upsampling module is also connected to the second multi-scale fusion module of the fourth feature extraction module, the feature splicing layer of the second upsampling module is connected to the second multi-scale fusion module of the third feature extraction module, the feature splicing layer of the third upsampling module is connected to the second multi-scale fusion module of the second feature extraction module, and the feature splicing layer of the fourth upsampling module is connected to the second multi-scale fusion module of the first feature extraction module.

[0020] Optionally, the multi-scale fusion module works as follows:

[0021] Dividing the feature map received by the multi-scale fusion module into N subsets;

[0022] Directly generate corresponding subsets to be spliced from one or more subsets of the N subsets;

[0023] Directly perform a convolution operation on one or more subsets of the N subsets, and use the obtained convolution results as the corresponding subsets to be spliced;

[0024] After concatenating one or more subsets of the N subsets with the convolution results of other subsets, further performing a convolution operation, and then using the obtained further convolution results as the corresponding subset to be concatenated;

[0025] The corresponding subsets to be spliced of the N subsets are spliced to obtain a new feature map; each subset in the N subsets has the same spatial size as the corresponding subset to be spliced; the multi-scale fusion module uses convolution kernels of different sizes for convolution operations performed on different subsets.

[0026] Optionally, the acquiring of a sample medical image and lesion annotation data corresponding to the sample medical image includes:

[0027] Obtaining raw ultrasound images from a medical system;

[0028] Processing the original ultrasound image using an image enhancement technique to obtain an enhanced image of the original ultrasound image;

[0029] Acquiring lesion annotation data of the original ultrasound image, and using the lesion annotation data of the original ultrasound image as the lesion annotation data of the enhanced image;

[0030] A model training sample data set is constructed using the original ultrasound image and the enhanced image, and the model sample data set includes a plurality of sample medical images.

[0031] Optionally, constructing a model training sample dataset from the original ultrasound image and the enhanced image includes:

[0032] For the images in the original ultrasound image and the enhanced image that do not meet the target size, the image size is adjusted by cropping and / or scaling so that the adjusted image size meets the target size; the target size is the image input size of the target model.

[0033] Optionally, the performing blur processing on the sample medical image to obtain a plurality of blurred images corresponding to the sample medical image includes:

[0034] Setting Gaussian kernels of various sizes, wherein the lateral size of the Gaussian kernel corresponds to the number of pixels scanned horizontally, and the longitudinal size of the Gaussian kernel corresponds to the number of pixels scanned vertically;

[0035] The sample medical images are scanned respectively using the set Gaussian kernel to perform Gaussian blur processing on the sample medical images to different degrees, thereby obtaining a plurality of blurred images corresponding to the sample medical images; wherein, the larger the Gaussian kernel size, the higher the blur degree of the blurred image obtained by the Gaussian blur processing; and the smaller the Gaussian kernel size, the lower the blur degree of the blurred image obtained by the Gaussian blur processing.

[0036] A second aspect of the present application provides a medical image segmentation method. The medical image segmentation method includes:

[0037] Acquire a medical image to be segmented;

[0038] The medical image to be segmented is used as the input of the medical image segmentation model, and the medical image to be segmented is processed by the medical image segmentation model to obtain the image segmentation result output by the medical image segmentation model; the image segmentation result includes the lesion area predicted by the medical image segmentation model; the medical image segmentation model is a model trained according to the training method of the medical image segmentation model described in the first aspect. The third aspect of the present application provides a training device for a medical image segmentation model. The training device includes:

[0039] A sample data acquisition unit, configured to acquire a sample medical image and lesion annotation data corresponding to the sample medical image;

[0040] an image blurring processing unit, configured to perform blurring on the sample medical image to obtain a plurality of blurred images corresponding to the sample medical image; the blurring degrees of the plurality of blurred images being different;

[0041] a training stage division unit, configured to divide the plurality of blurred images into a plurality of training stages according to blurring degrees, wherein blurred images with relatively high blurring degrees are divided into earlier training stages among the plurality of training stages, and blurred images with relatively low blurring degrees are divided into later training stages among the plurality of training stages;

[0042] A model parameter adjustment unit is used to adjust the parameters of the target model based on the lesion annotation data and the blurred image corresponding to the training stage in each training stage when the target model is trained in stages, until the last training stage of the multiple training stages is completed, thereby obtaining a medical image segmentation model.

[0043] A fourth aspect of the present application provides a medical image segmentation device. The medical image segmentation device includes:

[0044] An image acquisition unit for acquiring a medical image to be segmented, configured to acquire the medical image to be segmented;

[0045] A segmentation result obtaining unit is used to take the medical image to be segmented as the input of a medical image segmentation model, process the medical image to be segmented through the medical image segmentation model, and obtain the image segmentation result output by the medical image segmentation model; the image segmentation result includes the lesion area predicted by the medical image segmentation model; the medical image segmentation model is a model trained according to the training method of the medical image segmentation model described in the first aspect.

[0046] A fifth aspect of the present application provides a computer device. The identification device includes:

[0047] a memory having a computer program stored thereon;

[0048] A processor is used to execute the computer program in the memory to implement the steps of the training method of the medical image segmentation model provided in the first aspect, or the steps of the medical image segmentation method provided in the second aspect.

[0049] A sixth aspect of the present application provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the program implements the steps of the training method for the medical image segmentation model provided in the first aspect, or implements the steps of the medical image segmentation method provided in the second aspect.

[0050] Compared with the existing technology, this application has the following beneficial effects:

[0051] In the present application, firstly, a sample medical image and lesion annotation data corresponding to the sample medical image are obtained; at this stage, the sample medical image can be blurred to obtain a plurality of blurred images corresponding to the sample medical image; then the plurality of blurred images are divided into a plurality of training stages according to the degree of blur, specifically, the blurred images with a relatively high degree of blur are divided into the earlier training stages among the plurality of training stages, and the blurred images with a relatively low degree of blur are divided into the later training stages among the plurality of training stages; finally, when the target model is trained in stages, the parameters of the target model are adjusted in each training stage based on the lesion annotation data and the blurred images corresponding to the training stage, until the last training stage in the plurality of training stages is completed, and a medical image segmentation model is obtained. In the present application, global features are first learned through an image training model with a higher degree of blur, and then more detailed local features are learned through an image training model with a lower degree of blur, and the model training is performed step by step from simple to complex, so as to better identify shadows, noise, false textures, etc. in the image, and avoid the interference of the above details on the accuracy of image segmentation. Compared with medical staff manually performing image segmentation and marking abnormal areas, this solution uses the trained medical image segmentation model to perform image segmentation, which can improve the efficiency and accuracy of medical image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0053] Figure 1 A flowchart of a method for training a medical image segmentation model provided in an embodiment of the present application;

[0054] Figure 2 A schematic diagram of image blurring processing for a training method of a medical image segmentation model provided in an embodiment of the present application;

[0055] Figure 3 A workflow diagram of a multi-scale fusion module of a training method for a medical image segmentation model provided in an embodiment of the present application;

[0056] Figure 4 A schematic diagram of the workflow of a multi-scale fusion module of a training method for a medical image segmentation model provided in an embodiment of the present application;

[0057] Figure 5 A schematic diagram of the structure of an encoder in a training method for a medical image segmentation model provided in an embodiment of the present application;

[0058] Figure 6 A schematic diagram of the structure of a decoder in a training method for a medical image segmentation model provided in an embodiment of the present application;

[0059] Figure 7 A schematic diagram of the encoder-decoder structure in a training method for a medical image segmentation model provided in an embodiment of the present application;

[0060] Figure 8 A flowchart of a medical image segmentation method provided in an embodiment of the present application;

[0061] Figure 9 A schematic diagram of the structure of a training device for a medical image segmentation model provided in an embodiment of the present application;

[0062] Figure 10 A schematic structural diagram of a medical image segmentation device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] As previously described, in the current field of medical image processing, medical staff typically manually identify and segment medical images to mark abnormal areas within them. However, due to the presence of shadows, noise, or false textures from adjacent organs, medical staff expend considerable effort on image recognition, resulting in low image segmentation efficiency and high recognition costs. Furthermore, medical staff's identification of abnormal areas within images is subjective, which, to a certain extent, affects the accuracy of image segmentation. Improving the efficiency and accuracy of medical image segmentation is a key concern for those skilled in the art.

[0064] In view of the above problems, a solution is provided in the embodiment of the present application, and a training method, segmentation method and related products of a medical image segmentation model are proposed, aiming to improve the efficiency and accuracy of identifying medical images. In the technical solution of the present application, a sample medical image and lesion annotation data corresponding to the sample medical image are first obtained; at this stage, the sample medical image can be blurred to obtain a plurality of blurred images corresponding to the sample medical image; then the plurality of blurred images are divided into a plurality of training stages according to the degree of blur, specifically, blurred images with a relatively high degree of blur are divided into the earlier training stages among the plurality of training stages, and blurred images with a relatively low degree of blur are divided into the later training stages among the plurality of training stages; finally, when the target model is trained in stages, the parameters of the target model are adjusted in each training stage based on the lesion annotation data and the blurred images corresponding to the training stage, until the last training stage among the plurality of training stages is completed, and a medical image segmentation model is obtained.

[0065] As can be seen, in this application, the global features are first learned by training the model with images with a higher degree of blur, and then the more detailed local features are learned by training the model with images with a lower degree of blur. The model training is carried out step by step from simple to complex, so as to better identify shadows, noise, false textures, etc. in the image, and avoid the interference of these details on the accuracy of image segmentation. Compared with medical staff manually performing image segmentation and marking abnormal areas, the image segmentation performed by the trained medical image segmentation model in this solution can improve the efficiency and accuracy of medical image segmentation.

[0066] In order to help those skilled in the art better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0067] See also Figure 1 , which is a flow chart of a training method for a medical image segmentation model provided in an embodiment of the present application. Figure 1 The training method for the medical image segmentation model shown includes:

[0068] S101: Acquire a sample medical image and lesion annotation data corresponding to the sample medical image.

[0069] In this step, the sample medical image is used as the input of the target model, and the lesion annotation data corresponding to the sample medical image is used to compare with the training image obtained in the target model training phase to adjust the parameters. Specifically, in this application, the original ultrasound image can be obtained from the medical system, wherein the original ultrasound image is the ultrasound image of the patient to be identified. At this stage, the original ultrasound image can be processed by image enhancement technology to obtain an enhanced image of the original ultrasound image, wherein the image enhancement technology may include rotation, flipping, contrast change and noise addition, which are not specifically limited here and can be selected and adopted in actual applications. The image enhancement technology may also include other technologies to achieve image enhancement effects.

[0070] Then, lesion annotation data from the original ultrasound image can be obtained and used as lesion annotation data for the enhanced image. The lesion annotation data for the original ultrasound image is the result of professional lesion annotation performed by medical professionals on the original ultrasound image. Finally, the original ultrasound images and the enhanced images can be used to construct a model training sample dataset, which includes multiple sample medical images. In this way, multiple enhanced images of the original ultrasound image are obtained through image enhancement technology, increasing the number of samples.

[0071] In addition, because different imaging devices produce ultrasound images of varying sizes, and the target model typically requires input images of the same size, images in the original ultrasound image and enhanced image that do not conform to the target size can be resized by cropping and / or scaling to conform to the target size; the target size is the image input size of the target model.

[0072] S102: Performing fuzzy processing on the sample medical image to obtain a plurality of fuzzy processed images corresponding to the sample medical image.

[0073] In this step, the blurring degrees of the multiple blurred images are different. That is, in order to enable the target model to learn image features from simple to complex during training, the sample medical image can be blurred to different degrees, so that multiple blurred images of different degrees corresponding to the sample medical image are obtained.

[0074] Specifically, in the present application, Gaussian kernels of various sizes can be set to perform Gaussian blur processing on sample medical images, wherein the lateral size of the Gaussian kernel corresponds to the number of pixels scanned horizontally, and the longitudinal size of the Gaussian kernel corresponds to the number of pixels scanned vertically. It should be noted that, in general, (N×N) is used to represent the Gaussian kernel size, and the lateral N value is the same as the longitudinal N value. The N value of the Gaussian kernel can be set according to the Gaussian blur parameter σ. The relationship between the Gaussian blur parameter σ and the Gaussian kernel N value is specifically expressed as: N=2×σ+1. As an example, the Gaussian blur parameter σ represents the standard deviation of the Gaussian kernel.

[0075] Furthermore, the present application can use the set Gaussian kernel to scan the sample medical images separately to perform Gaussian blur processing on the sample medical images to different degrees, and obtain multiple blurred images corresponding to the sample medical images. Among them, the larger the Gaussian kernel size, the higher the blur degree of the blurred image obtained by the Gaussian blur processing; the smaller the Gaussian kernel size, the lower the blur degree of the blurred image obtained by the Gaussian blur processing. In this way, by performing Gaussian blur processing on the sample medical images to different degrees, the image resolution can be gradually improved during the training of the target model, so that the target model learns the image features from simple to complex, thereby improving the efficiency and accuracy of model prediction and recognition. That is, the target model is trained to learn the global features of the image when the image is at a lower resolution, while ignoring shadows, noise and irrelevant textures, and the target model is trained to learn the local features of the image when the image is at a higher resolution, so as to make more accurate predictions and recognition of the image.

[0076] Figure 2 A schematic diagram of image fuzzy processing for a training method of a medical image segmentation model provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, Figure 2The figure shows a plurality of blurred images with different blur degrees corresponding to the sample medical images obtained after Gaussian blurring of the sample medical images. Gaussian blurring of different degrees is performed on the sample medical images to obtain blurred images with a Gaussian kernel size of 3×3, blurred images with a Gaussian kernel size of 5×5, blurred images with a Gaussian kernel size of 7×7, blurred images with a Gaussian kernel size of 9×9, blurred images with a Gaussian kernel size of 11×11, blurred images with a Gaussian kernel size of 13×13, blurred images with a Gaussian kernel size of 15×15, blurred images with a Gaussian kernel size of 17×17, and blurred images with a Gaussian kernel size of 19×19, where σ represents the Gaussian blur parameter. Figure 2 As can be seen in the figure, the blurred image with a Gaussian kernel size of 3×3 has a lower degree of blur. As the Gaussian kernel size increases, the blurred image becomes more blurred. For example, the blurred image with a Gaussian kernel size of 19×19 has a higher degree of blur than the other blurred images. Furthermore, the Gaussian blur in this application uses a Gaussian kernel to scan each pixel in the image, uses a convolution operation to determine the weighted average grayscale value of the pixels in the neighborhood, and replaces this value with the center pixel of the template.

[0077] S103: Divide the multiple blurred images into multiple training stages according to blurring degrees.

[0078] In this step, to allow the target model to better learn image features from simple to complex, blurred images with relatively high blur levels can be assigned to the earlier training stages, while blurred images with relatively low blur levels can be assigned to the later training stages. This means that in the initial stages of model training, using blurred images with relatively high blur levels for model training, and in the more mature stages of model training, using blurred images with relatively low blur levels for model training, can further enable the target model to better learn the global and local features of the image during the model training process, thereby improving the efficiency and accuracy of model prediction and recognition.

[0079] As an example, the total number of rounds of pseudo-model training is t, the number of Gaussian blur parameters σ is γ, and the order from large to small is {σ1, σ2, ..., σ γ}, the training round is divided into γ stages, and each stage trains floor(t / γ) rounds, where floor represents the floor round function. In the first training segment, i.e., the first floor(t / γ) rounds, the blurred image obtained by Gaussian blur parameter σ1 is input; in the second training segment, i.e., (floor(t / γ), 2floor(t / γ)] rounds, the blurred image obtained by Gaussian blur parameter σ2 is input; in the third training segment, i.e., (2floor(t / γ), 3floor(t / γ)] rounds, the blurred image obtained by Gaussian blur parameter σ3 is input; in the γth training segment, i.e., ((γ-1)floor(t / γ), γfloor(t / γ)], the blurred image obtained by Gaussian blur parameter σ γ The blurred image is obtained after Gaussian blurring; this process is repeated until the entire training process is completed.

[0080] For example, if the total number of target model training rounds is 3000 and the number of Gaussian blur parameters σ is 10, then the order from large to small is {σ1, σ2, ..., σ 10}, and the corresponding blurred images after Gaussian blur are sorted from high to low according to the degree of blur. The training rounds are divided into 10 stages, and each training stage will train 300 rounds. In the first training stage, that is, rounds 1 to 300, the blurred image obtained by Gaussian blur parameter σ1 is input (the blurred image with a relatively high degree of blur); in the second training segment, that is, rounds 301 to 600, the blurred image obtained by Gaussian blur parameter σ2 is input; execute in sequence; in the tenth training segment, that is, rounds 2701 to 3000, the blurred image obtained by Gaussian blur parameter σ 10 The blurred image obtained after Gaussian blurring (a blurred image with a relatively low degree of blur).

[0081] S104: When the target model is trained in stages, the parameters of the target model are adjusted in each training stage based on the lesion annotation data and the blurred image corresponding to the training stage, until the last training stage of the multiple training stages is completed, thereby obtaining a medical image segmentation model.

[0082] It is understood that during the staged training of the target model, the parameters of the target model can be adjusted in each training stage based on the lesion annotation data and the blurred images obtained during the training in that training stage, until the training of the last training stage is completed, and a medical image segmentation model is obtained. Specifically, blurred images with different blur levels can be input in each training stage, and the target model can perform image segmentation based on the input blurred images to predict the lesion area in the input blurred images and obtain lesion area prediction data, wherein the lesion area is the area of the body in the predicted blurred image where the lesion occurs. The parameters of the target model are then adjusted based on the difference between the lesion area prediction data and the lesion annotation data.

[0083] It should be further explained that the target model in this application has an encoder-decoder structure of a UNet network. It can be understood that the encoder is used for feature extraction and the decoder is used for downsampling. A multi-scale fusion module is provided in the encoder-decoder structure. The multi-scale fusion module is used to segment the feature map output by the convolutional layer in the encoder-decoder structure and perform multi-scale fusion to form a new feature map. In this way, by embedding the multi-scale fusion module in the target model, the ability of the target model to learn local features of the image can be further improved. Next, the working method of the multi-scale fusion module will be introduced in conjunction with steps S301-S305. Figure 3 The workflow diagram of the multi-scale fusion module of the training method of a medical image segmentation model provided in the embodiment of the present application is as follows: Figure 3 As shown:

[0084] S301: Divide the feature map received by the multi-scale fusion module into N subsets.

[0085] N is an integer greater than 1. For example, N=4, and in this step, the feature map can be divided into 4 subsets, {x1, x2, x3, x4}.

[0086] S302: directly generate corresponding subsets to be spliced from one or more subsets of the N subsets.

[0087] S303: directly performing a convolution operation on one or more subsets of the N subsets, and directly using the obtained convolution result as the corresponding subset to be spliced.

[0088] S304: After concatenating one or more subsets of the N subsets with the convolution results of other subsets, further convolution operation is performed, and the obtained further convolution results are used as the corresponding subsets to be concatenated.

[0089] It can be understood that there are three situations in which subsets to be spliced are obtained in the present application. The first situation (corresponding to the explanation of step S302): If N subsets are 4 subsets, the first subset x1 in the 4 subsets can be directly used to generate its corresponding subset to be spliced; the second situation (corresponding to the explanation of step S303): If N subsets are 4 subsets, the second subset x2 in the 4 subsets can be convolved to obtain a convolution result and the convolution result can be used as the subset to be spliced corresponding to the second subset x2; the third situation (corresponding to the explanation of step S304): If N subsets are 4 subsets, the third subset x3 in the 4 subsets can be spliced with the convolution result of the second subset x2, and convolution can be performed to obtain a convolution result of the third subset x3 and the convolution result can be used as the subset to be spliced corresponding to the third subset x3. It should be noted that in the present application, N subsets being 4 subsets is only an example and is not specifically limited here.

[0090] S305: Splicing the subsets to be spliced corresponding to the N subsets to obtain a new feature map.

[0091] It should be noted that each subset in the N subsets has the same spatial size as the corresponding subset to be spliced, that is, the same feature map size and number of channels, and the multi-scale fusion module uses convolution kernels of different sizes for convolution operations performed on different subsets.

[0092] The following combination Figure 4 Let's introduce the workflow diagram of the multi-scale fusion module in detail. Figure 4 A schematic diagram of the workflow of a multi-scale fusion module of a training method for a medical image segmentation model provided in an embodiment of the present application is shown in FIG. Figure 4 As shown. Figure 4 In

[15] , the feature maps received by the multi-scale fusion module are divided into four subsets: {x1, x2, x3, x4}. Subset xi generates the corresponding subset yi to be spliced, where i represents 1, 2, 3, or 4. Subset x1 can be directly used to generate subset y1 to be spliced without any further operations. Subset x2 is convolved with a 3×3 convolution kernel to generate subset y2 to be spliced, with the same spatial size as subset x2. Subset x3 is concatenated with subset y2 and then convolved with a 5×5 convolution kernel to generate subset y3 to be spliced, with the same spatial size as subset x3. Subset x4 is concatenated with subset y3 and then convolved with a 7×7 convolution kernel to generate subset y4 to be spliced, with the same spatial size as subset x4. The resulting subsets {y1, y2, y3, y4} are then concatenated to produce a new feature map.

[0093] It can be seen that in the present application, the multi-scale fusion module can divide the feature map into multiple subsets, and directly generate subsets to be spliced from some subsets in the multiple subsets, or after convolving some subsets in the multiple subsets, use the convolution result as the subset to be spliced, or after splicing some subsets in the multiple subsets with the convolution result of some subsets after convolution, perform convolution again to obtain the convolution result as the subset to be spliced. In this way, by segmenting the feature map and convolving the obtained subset after segmenting the feature map to achieve feature extraction and fusion, it is possible to capture features in the image from a variety of different scales, further improve the model's ability to learn more detailed local features, and facilitate the final trained medical image segmentation model to have both the ability to extract coarse-grained features and fine-grained features in the image and the ability to segment regions. Furthermore, the encoder-decoder structure includes a feature extraction structure as an encoder and an upsampling structure as a decoder. Next, the feature extraction structure as an encoder and the upsampling structure as a decoder are explained in detail.

[0094] See also Figure 5 , Figure 5 A schematic diagram of the structure of the encoder in the training method of a medical image segmentation model provided in an embodiment of the present application is shown as follows: Figure 5 As shown. Figure 5 The feature extraction structure of the encoder is shown in the figure, wherein the feature extraction structure of the encoder includes five feature extraction modules, specifically a first feature extraction module, a second feature extraction module, a third feature extraction module, a fourth feature extraction module and a fifth feature extraction module connected in sequence.

[0095] Specifically, the first feature extraction module, the second feature extraction module, the third feature extraction module and the fourth feature extraction module all include a first convolution layer, a first multi-scale fusion module, a second convolution layer, a second multi-scale fusion module and a maximum pooling layer connected in sequence; the fifth feature extraction module includes a convolution layer and a multi-scale fusion module. Among the five feature extraction modules, the maximum pooling layer of the previous feature extraction module is connected to the first convolution layer of the next feature extraction module. Figure 5 In the figure, large arrows are used to represent convolution processing, and large dotted arrows are used to represent maximum pooling processing.

[0096] It can be understood that in the first feature extraction module, the blurred image is input into the decoder, and then 64 3×3 convolution kernels are used to convolve the blurred image in the first convolution layer to generate a feature map of 64 channels. At this stage, the feature map is processed using the first multi-scale fusion module to continue to generate a feature map of 64 channels. Next, 64 3×3 convolution kernels are used to convolve the feature map in the second convolution layer to continue to generate a feature map of 64 channels. Finally, the feature map is processed using the second multi-scale fusion module to finally generate a feature map of 64 channels. After this stage, the second feature extraction module performs 2×2 maximum pooling on the feature map finally output by the first feature extraction module. Among them, the activation function of the convolution layer is ReLU, and the step size can be 1.

[0097] It should be noted that the operations of the second, third, and fourth feature extraction modules are similar to those of the first feature extraction module. The second feature extraction module ultimately generates a feature map of 128 channels; the third feature extraction module ultimately generates a feature map of 256 channels; and the fourth feature extraction module ultimately generates a feature map of 512 channels. It should also be noted that although the fifth feature extraction module only includes one convolutional layer and one multi-scale fusion module, the operations of the fifth feature extraction module are still similar to those of the first feature extraction module. The fifth feature extraction module ultimately generates a feature map of 1024 channels.

[0098] See also Figure 6 , Figure 6 A schematic diagram of the structure of a decoder in a training method for a medical image segmentation model provided in an embodiment of the present application is shown in FIG. Figure 6 As shown. Figure 6 , the upsampling structure of the decoder is shown in FIG, wherein the upsampling structure of the decoder includes four upsampling modules, specifically a first upsampling module, a second upsampling module, a third upsampling module and a fourth upsampling module connected in sequence.

[0099] Specifically, the first upsampling module, the second upsampling module and the third upsampling module each include a deconvolution layer, a feature splicing layer, a third convolution layer, a third multi-scale fusion module, a fourth convolution layer and a fourth multi-scale fusion module; the fourth upsampling module includes a deconvolution layer, a feature splicing layer, a third convolution layer, a third multi-scale fusion module, a fourth convolution layer and a fourth multi-scale fusion module and a fifth convolution layer. Among the four upsampling modules, the fourth multi-scale fusion module of the previous upsampling module is connected to the deconvolution layer of the next upsampling module. Figure 6 In the figure, small arrows are used to represent convolution processing, and small dotted arrows are used to represent deconvolution processing.

[0100] See also Figure 7 , Figure 7 A schematic diagram of the structure of a decoder in a training method for a medical image segmentation model provided in an embodiment of the present application is shown in FIG. Figure 7 As shown. Figure 7 The feature extraction structure of the encoder and the upsampling structure of the decoder are shown in . Specifically, the feature splicing layer of the first upsampling module is also connected to the second multi-scale fusion module of the fourth feature extraction module, the feature splicing layer of the second upsampling module is connected to the second multi-scale fusion module of the third feature extraction module, the feature splicing layer of the third upsampling module is connected to the second multi-scale fusion module of the second feature extraction module, and the feature splicing layer of the fourth upsampling module is connected to the second multi-scale fusion module of the first feature extraction module. Figure 7 In FIG, the largest arrow is used to represent the splicing process.

[0101] It can be understood that in the first upsampling module, the feature map of 1024 channels ultimately generated by the fifth feature extraction module is subjected to deconvolution or interpolation processing, that is, the width and height of the 1024-channel feature map are doubled to generate a 512-channel feature map. Next, the feature splicing layer of the first upsampling module is connected to the second multi-scale fusion module of the fourth feature extraction module to generate a spliced feature map. Then, in the third convolutional layer, 512 3×3 convolution kernels are used to convolve the spliced feature map to generate a 512-channel feature map. At this stage, the feature map is processed by the third scale fusion module to also generate a 512-channel feature map. After this, the feature map is further convolved with 512 3×3 convolution kernels in the fourth convolutional layer to generate a 512-channel feature map. The feature map is then processed by the fourth scale fusion module to finally generate a 512-channel feature map.

[0102] It should be noted that the operations of the second upsampling module and the third upsampling module are similar to those of the first upsampling module. The second upsampling module ultimately generates a feature map of 256 channels; the third upsampling module ultimately generates a feature map of 128 channels. It should also be noted that although the fourth upsampling module also includes a fifth convolutional layer, the operation of the fourth upsampling module is still similar to that of the first upsampling module. The fourth upsampling module ultimately generates a feature map of 64 channels, which is then processed by the convolution of the fifth convolutional layer to ultimately generate a feature map of 1 channel, where the feature map of 1 channel is the predicted map of the blurred image.

[0103] The training process of the medical image segmentation model has been introduced in detail in the above process. Next, the application process of the medical image segmentation model will be introduced. Figure 8 A flowchart of a medical image segmentation method provided in an embodiment of the present application is shown in FIG. Figure 8As shown. The medical image segmentation method includes:

[0104] S801: Acquire a medical image to be segmented.

[0105] In this step, the original ultrasound image of the breast can be obtained from the medical system, and the image can be processed to obtain the medical image to be segmented which is input into the medical image segmentation model. The processing process is similar to the above-mentioned model training process and will not be repeated here.

[0106] S802: Using the medical image to be segmented as input of a medical image segmentation model, processing the medical image to be segmented by the medical image segmentation model to obtain an image segmentation result output by the medical image segmentation model.

[0107] In this step, since the model has learned the global and local features of the image during the aforementioned model training process, accordingly, during the application of the model, the medical image segmentation model can perform segmentation and recognition processing on the medical image to be segmented to obtain an image segmentation result. Based on the image segmentation result, it is possible to distinguish whether the nodules in the breast ultrasound image are tough in texture, smooth in surface, and have no obvious adhesion to the surrounding skin, etc. In this way, using the trained medical image segmentation model to perform segmentation and recognition processing on the medical image to be segmented can not only realize automatic labeling of abnormal areas in the medical image, but also further improve the efficiency and accuracy of medical image segmentation. The image segmentation result includes the lesion area predicted by the medical image segmentation model; the medical image segmentation model is a model trained according to the training method of the medical image segmentation model in the above process.

[0108] It should be noted that different medical image segmentation models can be trained for image segmentation and recognition processing of different organisms. It only requires inputting the ultrasound image corresponding to the organism in the initial stage of model training. After the model is trained, the medical image segmentation model corresponding to the organism can be generated.

[0109] Based on the training method of the medical image segmentation model provided in the aforementioned embodiment, the present application also provides a training device for a medical image segmentation model. Figure 9 This is a schematic diagram of the structure of a training device for a medical image segmentation model provided in an embodiment of the present application. Figure 9 The training device includes:

[0110] The sample data acquisition unit 901 is used to acquire a sample medical image and lesion annotation data corresponding to the sample medical image;

[0111] An image blurring processing unit 902 is configured to perform blurring on the sample medical image to obtain a plurality of blurred images corresponding to the sample medical image; the blurring degrees of the plurality of blurred images are different;

[0112] a training stage division unit 903 configured to divide the plurality of blurred images into a plurality of training stages according to blurring degrees, wherein blurred images with a relatively high blurring degree are divided into an earlier training stage among the plurality of training stages, and blurred images with a relatively low blurring degree are divided into a later training stage among the plurality of training stages;

[0113] The model parameter adjustment unit 904 is used to adjust the parameters of the target model based on the lesion annotation data and the blurred image corresponding to the training stage in each training stage when the target model is trained in stages, until the last training stage of the multiple training stages is completed, thereby obtaining a medical image segmentation model.

[0114] Optionally, the model parameter adjustment unit 904 includes:

[0115] a prediction data obtaining unit, configured to use the corresponding blurred image as input to the target model in each training stage, perform image segmentation based on the input blurred image by the target model, predict the lesion area in the input blurred image, and obtain lesion area prediction data;

[0116] A gap adjustment parameter unit is used to adjust the parameters of the target model based on the gap between the lesion area prediction data and the lesion annotation data.

[0117] Optionally, the target model is specifically used to:

[0118] The target model has an encoder-decoder structure of a UNet network; a multi-scale fusion module is provided in the encoder-decoder structure, and the multi-scale fusion module is used to split the feature map output by the convolutional layer in the encoder-decoder structure and perform multi-scale fusion to form a new feature map.

[0119] Optionally, the target model is specifically used to:

[0120] The encoder-decoder structure includes a feature extraction structure as an encoder and an upsampling structure as a decoder;

[0121] The feature extraction structure includes five feature extraction modules, specifically a first feature extraction module, a second feature extraction module, a third feature extraction module, a fourth feature extraction module and a fifth feature extraction module connected in sequence;

[0122] The upsampling structure includes four upsampling modules, specifically a first upsampling module, a second upsampling module, a third upsampling module and a fourth upsampling module connected in sequence;

[0123] The first feature extraction module, the second feature extraction module, the third feature extraction module, and the fourth feature extraction module each include a first convolution layer, a first multi-scale fusion module, a second convolution layer, a second multi-scale fusion module, and a maximum pooling layer connected in sequence; the fifth feature extraction module includes a convolution layer and a multi-scale fusion module; among the five feature extraction modules, the maximum pooling layer of the previous feature extraction module is connected to the first convolution layer of the next feature extraction module;

[0124] The first upsampling module, the second upsampling module and the third upsampling module each include a deconvolution layer, a feature splicing layer, a third convolution layer, a third multi-scale fusion module, a fourth convolution layer and a fourth multi-scale fusion module; the fourth upsampling module includes a deconvolution layer, a feature splicing layer, a third convolution layer, a third multi-scale fusion module, a fourth convolution layer and a fourth multi-scale fusion module and a fifth convolution layer; among the four upsampling modules, the fourth multi-scale fusion module of the previous upsampling module is connected to the deconvolution layer of the subsequent upsampling module;

[0125] The feature splicing layer of the first upsampling module is also connected to the second multi-scale fusion module of the fourth feature extraction module, the feature splicing layer of the second upsampling module is connected to the second multi-scale fusion module of the third feature extraction module, the feature splicing layer of the third upsampling module is connected to the second multi-scale fusion module of the second feature extraction module, and the feature splicing layer of the fourth upsampling module is connected to the second multi-scale fusion module of the first feature extraction module.

[0126] Optionally, the multi-scale fusion module is specifically used to:

[0127] Dividing the feature map received by the multi-scale fusion module into N subsets;

[0128] Directly generate corresponding subsets to be spliced from one or more subsets of the N subsets;

[0129] Directly perform a convolution operation on one or more subsets of the N subsets, and use the obtained convolution results as the corresponding subsets to be spliced;

[0130] After concatenating one or more subsets of the N subsets with the convolution results of other subsets, further performing a convolution operation, and then using the obtained further convolution results as the corresponding subset to be concatenated;

[0131] The corresponding subsets to be spliced of the N subsets are spliced to obtain a new feature map; each subset in the N subsets has the same spatial size as the corresponding subset to be spliced; the multi-scale fusion module uses convolution kernels of different sizes for convolution operations performed on different subsets.

[0132] Optionally, the sample data acquisition unit 901 includes:

[0133] An original image acquisition unit, configured to acquire an original ultrasound image from a medical system;

[0134] an enhanced image obtaining unit, configured to process the original ultrasonic image using an image enhancement technique to obtain an enhanced image of the original ultrasonic image;

[0135] a labeling data acquiring unit, configured to acquire the lesion labeling data of the original ultrasound image, and use the lesion labeling data of the original ultrasound image as the lesion labeling data of the enhanced image;

[0136] The sample set construction unit is used to construct a model training sample data set using the original ultrasound image and the enhanced image, wherein the model sample data set includes a plurality of sample medical images.

[0137] Optionally, the sample set construction unit is specifically configured to:

[0138] For the images in the original ultrasound image and the enhanced image that do not meet the target size, the image size is adjusted by cropping and / or scaling so that the adjusted image size meets the target size; the target size is the image input size of the target model.

[0139] Optionally, the image blur processing unit 902 is specifically configured to:

[0140] Setting Gaussian kernels of various sizes, wherein the lateral size of the Gaussian kernel corresponds to the number of pixels scanned horizontally, and the longitudinal size of the Gaussian kernel corresponds to the number of pixels scanned vertically;

[0141] The sample medical images are scanned respectively using the set Gaussian kernel to perform Gaussian blur processing on the sample medical images to different degrees, thereby obtaining a plurality of blurred images corresponding to the sample medical images; wherein, the larger the Gaussian kernel size, the higher the blur degree of the blurred image obtained by the Gaussian blur processing; and the smaller the Gaussian kernel size, the lower the blur degree of the blurred image obtained by the Gaussian blur processing.

[0142] Based on the medical image segmentation method provided in the aforementioned embodiment, the present application also provides a medical image segmentation device accordingly. Figure 10 This is a schematic diagram of the structure of a medical image segmentation device. Figure 10 The training device includes:

[0143] The image to be segmented acquiring unit 1001 is used to acquire the medical image to be segmented;

[0144] The segmentation result obtaining unit 1002 is used to take the medical image to be segmented as the input of the medical image segmentation model, process the medical image to be segmented through the medical image segmentation model, and obtain the image segmentation result output by the medical image segmentation model; the image segmentation result includes the lesion area predicted by the medical image segmentation model.

[0145] The present application also provides a device for identifying a high-density sign of the middle cerebral artery. The device comprises:

[0146] A memory having a computer program stored thereon.

[0147] A processor is used to execute a computer program in a memory to implement some or all of the steps in the training method of the medical image segmentation model introduced in the aforementioned embodiment, or to implement some or all of the steps in the medical image segmentation method introduced in the aforementioned embodiment.

[0148] The present application also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the program implements some or all of the steps of the medical image segmentation model training method described in the aforementioned embodiment, or implements some or all of the steps of the medical image segmentation method described in the aforementioned embodiment.

[0149] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments. The device and equipment embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components indicated as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0150] The above is merely one specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A training method for a medical image segmentation model, characterized in that: include: Acquiring a sample medical image and lesion annotation data corresponding to the sample medical image; Performing blur processing on the sample medical image to obtain a plurality of blurred images corresponding to the sample medical image; the blurring degrees of the plurality of blurred images are different; Dividing the multiple blurred images into multiple training stages according to blur levels, wherein blurred images with relatively high blur levels are divided into earlier training stages among the multiple training stages, and blurred images with relatively low blur levels are divided into later training stages among the multiple training stages; When the target model is trained in stages, the parameters of the target model are adjusted in each training stage based on the lesion annotation data and the blurred image corresponding to the training stage until the last training stage of the multiple training stages is completed, thereby obtaining a medical image segmentation model.

2. The training method according to claim 1, characterized in that The adjusting the parameters of the target model based on the lesion annotation data and the blurred image corresponding to the training stage in each training stage includes: In each training stage, the corresponding blurred image is used as the input of the target model, and the target model is used to perform image segmentation based on the input blurred image to predict the lesion area in the input blurred image to obtain lesion area prediction data; Based on the gap between the lesion area prediction data and the lesion annotation data, the parameters of the target model are adjusted.

3. The training method according to claim 1, characterized in that The target model has an encoder-decoder structure of a UNet network; a multi-scale fusion module is provided in the encoder-decoder structure, and the multi-scale fusion module is used to split the feature map output by the convolutional layer in the encoder-decoder structure and perform multi-scale fusion to form a new feature map.

4. The training method according to claim 3, characterized in that The encoder-decoder structure includes a feature extraction structure as an encoder and an upsampling structure as a decoder; The feature extraction structure includes five feature extraction modules, specifically a first feature extraction module, a second feature extraction module, a third feature extraction module, a fourth feature extraction module and a fifth feature extraction module connected in sequence; The upsampling structure includes four upsampling modules, specifically a first upsampling module, a second upsampling module, a third upsampling module and a fourth upsampling module connected in sequence; The first feature extraction module, the second feature extraction module, the third feature extraction module, and the fourth feature extraction module each include a first convolution layer, a first multi-scale fusion module, a second convolution layer, a second multi-scale fusion module, and a maximum pooling layer connected in sequence; the fifth feature extraction module includes a convolution layer and a multi-scale fusion module; among the five feature extraction modules, the maximum pooling layer of the previous feature extraction module is connected to the first convolution layer of the next feature extraction module; The first upsampling module, the second upsampling module and the third upsampling module each include a deconvolution layer, a feature splicing layer, a third convolution layer, a third multi-scale fusion module, a fourth convolution layer and a fourth multi-scale fusion module; the fourth upsampling module includes a deconvolution layer, a feature splicing layer, a third convolution layer, a third multi-scale fusion module, a fourth convolution layer and a fourth multi-scale fusion module and a fifth convolution layer; among the four upsampling modules, the fourth multi-scale fusion module of the previous upsampling module is connected to the deconvolution layer of the subsequent upsampling module; The feature splicing layer of the first upsampling module is also connected to the second multi-scale fusion module of the fourth feature extraction module, the feature splicing layer of the second upsampling module is connected to the second multi-scale fusion module of the third feature extraction module, the feature splicing layer of the third upsampling module is connected to the second multi-scale fusion module of the second feature extraction module, and the feature splicing layer of the fourth upsampling module is connected to the second multi-scale fusion module of the first feature extraction module.

5. The training method according to claim 3 or 4, characterized in that: The working mode of the multi-scale fusion module is as follows: Dividing the feature map received by the multi-scale fusion module into N subsets; Directly generate corresponding subsets to be spliced from one or more subsets of the N subsets; Directly perform a convolution operation on one or more subsets of the N subsets, and use the obtained convolution results as the corresponding subsets to be spliced; After concatenating one or more subsets of the N subsets with the convolution results of other subsets, further performing a convolution operation, and then using the obtained further convolution results as the corresponding subset to be concatenated; Splicing the subsets to be spliced corresponding to the N subsets to obtain a new feature map; Each subset in the N subsets has the same spatial size as the corresponding subset to be spliced; the multi-scale fusion module uses convolution kernels of different sizes for the convolution operations performed on different subsets.

6. The training method according to any one of claims 1 to 4, characterized in that: The acquiring of the sample medical image and the lesion annotation data corresponding to the sample medical image includes: Obtaining raw ultrasound images from a medical system; Processing the original ultrasound image using an image enhancement technique to obtain an enhanced image of the original ultrasound image; Acquiring lesion annotation data of the original ultrasound image, and using the lesion annotation data of the original ultrasound image as the lesion annotation data of the enhanced image; A model training sample data set is constructed by combining the original ultrasound image and the enhanced image, wherein the model training sample data set includes a plurality of sample medical images.

7. The training method according to claim 6, characterized in that The constructing a model training sample data set from the original ultrasound image and the enhanced image includes: For the images in the original ultrasound image and the enhanced image that do not meet the target size, the image size is adjusted by cropping and / or scaling so that the adjusted image size meets the target size; the target size is the image input size of the target model.

8. The training method according to any one of claims 1 to 4, characterized in that: The fuzzy processing of the sample medical image to obtain a plurality of fuzzy processed images corresponding to the sample medical image includes: Setting Gaussian kernels of various sizes, wherein the lateral size of the Gaussian kernel corresponds to the number of pixels scanned horizontally, and the longitudinal size of the Gaussian kernel corresponds to the number of pixels scanned vertically; The sample medical images are scanned respectively using the set Gaussian kernel to perform Gaussian blur processing on the sample medical images to different degrees, thereby obtaining a plurality of blurred images corresponding to the sample medical images; wherein, the larger the Gaussian kernel size, the higher the blur degree of the blurred image obtained by the Gaussian blur processing; and the smaller the Gaussian kernel size, the lower the blur degree of the blurred image obtained by the Gaussian blur processing.

9. A medical image segmentation method, characterized in that: include: Acquire a medical image to be segmented; The medical image to be segmented is used as the input of a medical image segmentation model, and the medical image to be segmented is processed by the medical image segmentation model to obtain an image segmentation result output by the medical image segmentation model; the image segmentation result includes the lesion area predicted by the medical image segmentation model; the medical image segmentation model is a model trained according to the training method of the medical image segmentation model according to any one of claims 1-8.

10. A training device for a medical image segmentation model, characterized in that: include: A sample data acquisition unit, configured to acquire a sample medical image and lesion annotation data corresponding to the sample medical image; an image blurring processing unit, configured to perform blurring on the sample medical image to obtain a plurality of blurred images corresponding to the sample medical image; the blurring degrees of the plurality of blurred images being different; a training stage division unit, configured to divide the plurality of blurred images into a plurality of training stages according to blurring degrees, wherein blurred images with relatively high blurring degrees are divided into earlier training stages among the plurality of training stages, and blurred images with relatively low blurring degrees are divided into later training stages among the plurality of training stages; A model parameter adjustment unit is used to adjust the parameters of the target model based on the lesion annotation data and the blurred image corresponding to the training stage in each training stage when the target model is trained in stages, until the last training stage of the multiple training stages is completed, thereby obtaining a medical image segmentation model.

11. A computer device, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the training method for the medical image segmentation model described in any one of claims 1 to 8, or to implement the steps of the medical image segmentation method described in claim 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, it implements the steps of the training method of the medical image segmentation model described in any one of claims 1 to 8, or implements the steps of the medical image segmentation method described in claim 9.

Images