Unsupervised lesion segmentation method based on large model and pseudo-label learning

Abstract

The invention discloses an unsupervised lesion segmentation method based on a large model and pseudo label learning, and belongs to the field of image recognition. The method comprises the following steps of: firstly, generating high-quality classification pseudo labels by adopting a visual-language large model (CLIP), and then, supervising and training an image classification network by using the pseudo labels; in the process, a class activation map (CAM) of the classification network is further extracted, and a data enhancement strategy based on dynamic shielding is introduced, so that the performance of the classification network is further improved. Then, the class activation mapping with higher precision is utilized to generate guidance information for a segmentation large model (SAM), so that a high-quality segmentation pseudo label is obtained. Finally, the invention provides a self-training strategy, and the segmentation network is trained and continuously optimized through the pseudo labels, so that higher segmentation precision and efficiency are achieved.

Description

Technical Field

[0001] The present invention relates to an unsupervised lesion segmentation technology in medical images, belonging to the technical field of image recognition. Background Art

[0002] In the field of medical image segmentation, accurate identification and separation of lesion areas are crucial for disease diagnosis, treatment planning, and efficacy evaluation. Traditional methods, such as image segmentation techniques based on thresholds, edge detection, and active contours, are effective in some cases, but they mainly rely on low-level features of images and have difficulty coping with complex image conditions such as low contrast, blurred boundaries, and noise interference, making it difficult to obtain high-precision segmentation results.

[0003] In recent years, the rise of deep learning technology has brought revolutionary progress to medical image segmentation. Segmentation methods based on deep learning, especially the application of convolutional neural networks, have significantly improved the accuracy and efficiency of segmentation due to their powerful feature extraction capabilities. However, these methods rely on a large amount of high-quality pixel-level annotations. This densely annotated data is not only expensive to obtain, but also time-consuming and requires the deep participation of professional medical personnel. This has become the main bottleneck limiting the popularization of deep learning image segmentation technology in practical applications.

[0004] To solve this problem, researchers began to explore weakly supervised learning methods to reduce the reliance on densely labeled data. Common weakly supervised annotations include graffiti annotations, bounding boxes, point annotations, and image-level labels. Although these methods have reduced the annotation cost to a certain extent, they still face challenges such as insufficient supervision information and limited segmentation accuracy. Especially when dealing with multi-category segmentation tasks, it is still very difficult to obtain a high-precision segmentation model with limited annotation information. In addition, these methods still require a certain amount of manual annotation, and the experience and time requirements of the annotator are still high.

[0005] Unlike weakly supervised segmentation, unsupervised lesion segmentation methods in medical images can completely avoid manual annotation of training images to achieve image segmentation. Typical unsupervised lesion segmentation methods include methods based on color thresholds, clustering, edge detection, etc. However, these unsupervised image segmentation techniques often perform poorly when processing medical images with low contrast and unclear boundaries. These methods usually find it difficult to automatically learn enough features to distinguish different categories or regions from the image, resulting in limited accuracy and reliability of the segmentation results.

[0006] In response to these challenges, this paper proposes an unsupervised lesion segmentation method based on a large model and class activation mapping. This method does not rely on any manual annotation. It generates high-quality classification pseudo-labels by using the "visual-language" large model (CLIP), and combines the guidance information generated by the dynamic occlusion data enhancement strategy and the segmentation large model (SAM) to achieve high-quality segmentation model construction. This innovative method not only reduces the dependence on labeled data, but also effectively improves the automation and accuracy of lesion segmentation tasks, providing a new technical path for the field of medical image segmentation. Summary of the invention

[0007] The present invention aims to solve the problem of existing deep learning technology's dependence on pixel-level fine annotation in the field of lesion image segmentation, and proposes an innovative unsupervised segmentation method. This method does not rely on any annotation cost, and can achieve high-quality segmentation model construction. Specifically, the present invention first uses the "visual-language" large model (CLIP) to generate high-quality classification pseudo-labels, which are then used to supervise the training of an image classification network. In this process, we further extract the class activation map (CAM) of the classification network and introduce a data enhancement strategy based on dynamic occlusion to further improve the performance of the classification network. Subsequently, we use these more accurate class activation maps to generate guidance information for the segmentation large model (SAM) to obtain high-quality segmentation pseudo-labels. Finally, the present invention proposes a self-training strategy to train and continuously optimize the segmentation network through these pseudo-labels to achieve higher segmentation accuracy and efficiency.

[0008] The purpose of the present invention can be achieved through the following technical solution: an unsupervised lesion segmentation method based on a large model and pseudo-label learning; the specific steps are as follows:

[0009] Step 1: Collect a set of unlabeled images as a training set for the lesion segmentation task, and use the "vision-language" model to generate image classification pseudo labels;

[0010] Step 2: Use the classification pseudo-labels to train a classification model to generate a preliminary class activation map of the lesion on the training set;

[0011] Step 3: Use image occlusion based on the category activation map to perform data augmentation, retrain the classification model, and generate an enhanced category activation map;

[0012] Step 4: Generate segmentation pseudo labels using the category activation map and the segmentation large model SAM model;

[0013] Step 5: Use images with segmentation pseudo-labels as training sets and train an image segmentation network using an iterative training method based on pseudo-label screening.

[0014] Furthermore, the classification pseudo-label generation method in step 1 is:

[0015] Using an image encoder E img and text encoder E text Visual-language model; c is used to represent the image-level category, c = 0 means that the image does not contain the target lesion, and c = 1 means that the image contains the target lesion; T c Represents the prompt sentence corresponding to category c; for each image X i , use g c Indicates prompt statement T c The characteristics of i Represents image X i Characteristics of c and f i The calculation method is as follows:

[0016] f i =E img (X i ) Formula 1

[0017] g c =E text (Φ(T c )) Formula 2

[0018] Where Φ represents a word segmenter, which has the function of transforming the prompt sentence T c Convert to feature expression function; on this basis, use Indicates image X i The probability of belonging to category c is calculated as:

[0019]

[0020] Where Sim(f i , g c ) is f i and g c The cosine similarity of , j is the category index value; X is obtained by argmax function i The pseudo label y i :

[0021] y i = argmax(p i ) Formula 4.

[0022] Furthermore, the preliminary category activation map based on the classification model in step 2 is generated as follows:

[0023] The initial training set with pseudo labels obtained in step 1 is denoted as Where N s is the number of two-dimensional slices in the training set; in the dataset D containing classification pseudo labels1 Using standard supervised learning methods, we train an image classification network θ using the cross entropy loss function. c ;

[0024] Based on the training completed θ c , for category label y i = 1, the image is back-propagated through the gradient to obtain the preliminary category activation map, which is used to identify the area in the image that contributes most to the lesion classification; on the feature map F of a certain layer in the classification network, Represents X i The corresponding preliminary category activation map is calculated as follows:

[0025]

[0026] Among them, F k represents the kth channel of the feature map F, is a gradient operator; P avg represents the global average pooling operation; ReLU is a linear rectification function used to filter out negative values ​​in the gradient.

[0027] Furthermore, the enhanced category activation map in step 3 is generated as follows:

[0028] For each image marked y i =1 image X i , through the preliminary category activation map Q 0 Perform data augmentation based on adaptive occlusion; the specific method is to first pass a threshold τ to Q 0 Convert to a binary mask M i :

[0029]

[0030] Then the input image X i Medium i The area with value 1 is set to 0, and the occluded image is obtained:

[0031]

[0032] For the y i =0 image X i , then by randomly changing X i The occluded image is obtained by setting a sub-region to zero The dataset after occlusion enhancement is represented as Using D 2 , retrain the classification network through supervised training based on the cross entropy loss function; after the training is completed, the label y is retrained through formula 5 i= 1 generates an enhanced category activation map, denoted as and compared to, More marginal areas of the lesion can be excavated with higher quality.

[0033] Furthermore, the segmentation pseudo-labels in step 4 are generated as follows:

[0034] For y i =1 training image X i , from its corresponding enhanced class activation map Generate two kinds of prompt information: one is calculation The bounding box of the area above τ in the figure is used to obtain the bounding box hint. The second is to use the center point of the bounding box as the foreground point and its four vertices as the background points to form a point hint. These two hint information are combined to form a point hint. Represents X i Corresponding prompt information;

[0035] Then and X i Input the segmentation model SAM model to generate segmentation pseudo-annotations; the SAM model contains an image encoder E m , a prompt information encoder E prom and a decoder D m To obtain the segmentation mask of the image; for a given image X i , get the predicted mask M i ′ The formula is as follows:

[0036]

[0037] For each y i =1 image X i , use formula 8 to obtain its segmentation mask as the segmentation pseudo label; for the labeled y i = 0, the segmentation mask is set to empty; finally, the masks of all 2D slices in a 3D medical image are stacked to form a complete 3D segmentation pseudo label S i ; The corresponding 3D image segmentation training dataset is expressed as Where N v Represents the number of 3D images in the training dataset.

[0038] Furthermore, the iterative training method of the segmentation model based on pseudo-label screening in step 5 is:

[0039] Based on the segmentation pseudo-labels, a segmentation model is trained using an iterative training method with two stages. In the first stage, a 3D image segmentation dataset D with pseudo-labels is trained. 3 As a training set, a 3D segmentation network θ is trained using the standard supervised image segmentation model training strategy. s ; This process uses Dice loss function and cross entropy loss function to update the model parameters;

[0040] Training θ s Then, use its model parameters to make predictions in all training set images to get V i The updated pseudo-labels for For slices with non-zero areas in the , the bounding box is calculated for the non-zero area in each slice, and the SAM prompt is generated using the bounding box area and its center point and corner point in the same way as in step 4. The slice and the corresponding prompt are input into SAM to obtain the correction result of SAM; the results in each slice are stacked, and the resulting 3D segmentation correction result is recorded as

[0041] To evaluate The mass of and The higher the similarity, the The more reliable the similarity index is, the more reliable the similarity index is; the similarity index is defined as the Dice value between the two:

[0042]

[0043] in The value of is between 0 and 1. The closer the value is to 1, the and The higher the consistency between them; define a similarity threshold It is used to select pseudo labels with high reliability; the selected three-dimensional image and its pseudo label are recorded as in Threshold The number of selected 3D image samples;

[0044] In the second stage, the selected pseudo-label dataset D is used 4 The segmentation model is trained again. The training process also adopts the standard supervised training method, and the Dice loss function and the cross entropy loss function are used to update the model parameters to obtain the final lesion automatic segmentation model.

[0045] Compared with the prior art, the present invention has the following advantages:

[0046] (1) In today's medical image segmentation field, deep learning methods are highly dependent on large-scale, high-quality pixel-level annotated images. The acquisition of these annotated data is not only time-consuming and labor-intensive, but also extremely costly and often requires a large number of professional medical personnel to participate. The unsupervised lesion segmentation method proposed in the present invention does not require any annotation and can achieve a higher lesion segmentation accuracy without relying on manual annotation.

[0047] (2) Existing unsupervised image segmentation techniques often perform poorly when processing images with low contrast and unclear boundaries. In contrast, the present invention fully utilizes the powerful feature extraction capabilities of large models and combines a series of innovative strategies such as image occlusion based on category activation maps for data augmentation and pseudo-label similarity filtering, significantly improving the performance of the model at all stages.

[0048] (3) Existing pseudo-label-based image segmentation methods ignore the noise in pseudo-labels and lack effective strategies to screen low-quality pseudo-labels. The present invention uses the consistency between the pseudo-labels predicted by the neural network and the pseudo-labels generated by the SAM model to evaluate the quality of pseudo-labels, which can filter out low-quality pseudo-labels, thereby improving the performance of the final segmentation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the overall flow chart of the present invention.

[0050] Figure 2 Example diagram of the category activation map and segmentation pseudo-label of the present invention.

[0051] Figure 3 Comparison of the brain tumor segmentation effect in MRI images between the present invention and the existing method; the black line represents the gold standard, and the white area represents the segmentation result of the algorithm. DETAILED DESCRIPTION

[0052] In conjunction with the content of the present invention, the following embodiment of tumor lesion segmentation in MRI images is provided. This embodiment is implemented in a computer with an Intel(R) Core(TM) i7-6850K CPU@3.60GHZ, a GeForce GTX1080Ti GPU, and 32.0GB of memory, and the programming language is Python.

[0053] Step 1. Dataset establishment and classification pseudo-label generation. A dataset was established. 369 three-dimensional brain MRI images were used from the BraTS public dataset, of which 258 were used as training sets, 37 were used as validation sets, and 74 were used as test sets. For each image, mean-standard deviation normalization (Z-score Normalization) was applied to convert the pixel values ​​of the brain region into a normal distribution with a mean of 0 and a standard deviation of 1, and the segmentation labels of the image training images were discarded.

[0054] Generation of classification pseudo labels. In this example, the industry-trained "visual-language" model CLIP is used to generate image classification pseudo labels. CLIP contains an image encoder E img and text encoder E text Visual-language model. Let c represent the image-level category, c = 0 means that the image does not contain the target lesion, and c = 1 means that the image contains the target lesion. c The prompt statement corresponding to category c: T 0 = "This is a magnetic resonance image that does not contain a brain tumor, and there is no significant bright lesion area"; T 1 = "This is a magnetic resonance image of a brain tumor, in which there is a significantly bright lesion area." For each image X i , use g c Indicates prompt statement T c The characteristics of i Represents image X i Features. Features c and f i The calculation method is as follows:

[0055] f i =E img (X i ) Formula 1

[0056] g c =E text (Φ(T c )) Formula 2

[0057] Where Φ represents a word segmenter, which has the function of transforming the prompt sentence T c On this basis, we use Indicates image X i The probability of belonging to category c is calculated as:

[0058]

[0059] Where Sim(f i , g c ) is f i and gc The cosine similarity of , j is the category index value. X is obtained by argmax function i The pseudo label y i :

[0060] y i = argmax(p i ) Formula 4

[0061] Step 2: Train a classification network and obtain a preliminary category activation map:

[0062] The initial training set with pseudo labels obtained in step 1 is denoted as Where N s is the number of two-dimensional slices in the training set. 1 Using standard supervised learning methods, we train an image classification network θ using the cross entropy loss function. c The classification network uses the pre-trained ResNet50 structure. The training process uses the Adam optimizer, the learning rate is 0.001, and the iteration round (epoch) is 100 rounds.

[0063] Based on the training completed θ c , for category label y i = 1, the initial category activation map is obtained by gradient back propagation, which is used to identify the area in the image that contributes most to the lesion classification. Represents X i The corresponding preliminary category activation map is calculated as follows:

[0064]

[0065] Among them, F k represents the kth channel of the feature map F, is a gradient operator. P avg represents the global average pooling operation. ReLU is a linear rectification function used to filter out negative values ​​in the gradient.

[0066] Step 3: After enhancing the training set based on the preliminary category activation map, retrain the classification network and obtain the enhanced category activation map:

[0067] For each image marked y i =1 image X i , through the preliminary category activation map Q 0 Perform data augmentation based on adaptive occlusion. The specific method is to first pass a threshold τ = 0.5 to set Q 0 Convert to a binary mask M i :

[0068]

[0069] Then the input image X i Medium i The area with value 1 is set to 0, and the occluded image is obtained:

[0070]

[0071] For the y i =0 image X i , then by randomly changing X i The occluded image is obtained by setting a sub-region to zero The dataset after occlusion enhancement is represented as Using D 2 , retrain the classification network through supervised training based on the cross entropy loss function. After the training is completed, the label y is retrained through formula 5 i = 1 generates an enhanced category activation map, denoted as and compared to, More marginal areas of the lesion can be excavated with higher quality.

[0072] Step 4: Generate segmentation pseudo labels through enhanced category activation maps and SAM:

[0073] For y i =1 training image X i , from its corresponding enhanced class activation map Generate two kinds of prompt information: one is calculation The bounding box of the area above τ is used to obtain the bounding box hint. The second is to use the center point of the bounding box as the foreground point and its four vertices as the background points to form a point hint. These two hint information are combined and used Represents X i The corresponding prompt information.

[0074] Then and X i Input the segmentation model SAM to generate segmentation pseudo-annotations. The SAM model contains an image encoder E m , a prompt information encoder E prom and a decoder D m To obtain the segmentation mask of the image. In this example, the "ViT-B" pre-trained model is used for the image encoder of SAM. For a given image X i , get the predicted mask M i ′ The formula is as follows:

[0075]

[0076] For each y i =1 image X i , use Formula 8 to obtain its segmentation mask as the segmentation pseudo label. i = 0, the segmentation mask is set to empty. Finally, the masks of all 2D slices in a 3D medical image are stacked to form a complete 3D segmentation pseudo label S i The corresponding 3D image segmentation training dataset is represented as Where N v Represents the number of 3D images in the training dataset.

[0077] Step 5: Iterative training of the segmentation model based on pseudo-label screening Based on the segmentation pseudo-labels, a segmentation model is trained using an iterative training method including two stages. The three-dimensional image segmentation model in this embodiment adopts a 3DUNet structure.

[0078] In the first stage, the 3D image segmentation dataset D 3 As a training set, a 3D segmentation network θ is trained using the standard supervised image segmentation model training strategy. s In this process, the Dice loss function and the cross entropy loss function are used to update the model parameters. The optimizer is Adam, the iteration round is 100 rounds, and the learning rate is set to 0.001.

[0079] Training θ s Then, use its model parameters to make predictions in all training set images to get V i The updated pseudo-labels for For slices with non-zero areas in the , calculate the bounding box of the non-zero area in each slice, and generate SAM hints using the bounding box area and its center point and corner points in the same way as in step 4. Input the slice and the corresponding hint into SAM to obtain the SAM correction result. The results in each slice are stacked, and the resulting 3D segmentation correction result is recorded as

[0080] To evaluate The mass of and The higher the similarity, the The more reliable the similarity index is. The similarity index is defined as the Dice value between the two:

[0081]

[0082] in The value of is between 0 and 1. The closer the value is to 1, the and The higher the consistency between them, the higher the similarity threshold is. It is used to select pseudo labels with high reliability. The selected three-dimensional image and its pseudo label are recorded as in Threshold The number of selected 3D image samples.

[0083] In the second stage, the selected pseudo-label dataset D is used 4 The segmentation model is trained again. The training process also adopts the standard supervised training method, and the Dice loss function and the cross entropy loss function are used to update the model parameters to obtain the final lesion automatic segmentation model 3D UNet.

[0084] After training is completed, the final 3D UNet model is used to infer the samples in the test set to obtain the prediction results.

[0085] Figure 2 This is an example of the class activation map and pseudo-label based on the method of the present invention on the training set. It can be seen that the preliminary class activation map locates the lesion area to a certain extent, and the enhanced class activation map can more fully explore the lesion area. Based on the hints generated by the class activation map, the segmentation pseudo-label obtained using SAM has a good quality.

[0086] Figure 3 This is a comparison between the segmentation model trained by the unsupervised learning method of the present invention and other existing unsupervised learning methods in brain tumor lesion segmentation. It can be seen that compared with other methods, the segmentation results obtained are more accurate.

Claims

1. An unsupervised lesion segmentation method based on large model and pseudo-label learning; the specific steps are as follows: Step 1: Collect a set of unlabeled images as a training set for the lesion segmentation task, and use the "vision-language" model to generate image classification pseudo labels; Step 2: Use the classification pseudo-labels to train a classification model to generate a preliminary class activation map of the lesion on the training set; Step 3: Use image occlusion based on the category activation map to perform data augmentation, retrain the classification model, and generate an enhanced category activation map; Step 4: Generate segmentation pseudo labels using the category activation map and the segmentation large model SAM model; Step 5: Use images with segmentation pseudo-labels as training sets and train an image segmentation network using an iterative training method based on pseudo-label screening.

2. The unsupervised lesion segmentation method based on large model and pseudo-label learning as claimed in claim 1, characterized in that: The classification pseudo label generation method in step 1 is: Using an image encoder E img and text encoder E text Visual-language model; c is used to represent the image-level category, c = 0 means that the image does not contain the target lesion, and c = 1 means that the image contains the target lesion; T c Represents the prompt sentence corresponding to category c; for each image X i , use g c Indicates prompt statement T c The characteristics of f i Represents image X i Characteristics of c and f i The calculation method is as follows: f i =E img (X i ) Formula 1 g c = E text (Φ(T c )) Equation 2 Where Φ represents a word segmenter, which has the function of transforming the prompt sentence T c Convert to feature expression function; on this basis, use p i c Indicates image X i The probability of belonging to category c is calculated as: Where Sim(f i , g c ) is f i and g c The cosine similarity of , j is the category index value; X is obtained by argmax function i The pseudo label y i : y i = argmax(p i )Formula 4.

3. The unsupervised lesion segmentation method based on large model and pseudo-label learning as claimed in claim 1, characterized in that: The method for generating the preliminary category activation map based on the classification model in step 2 is: The initial training set with pseudo labels obtained in step 1 is denoted as Where N s is the number of two-dimensional slices in the training set; on the dataset D1 containing classification pseudo labels, a standard supervised learning method is used to train an image classification network θ using the cross entropy loss function c ; Based on the training completed θ c , for category label y i = 1, the initial category activation map is obtained by gradient back propagation, which is used to identify the area in the image that contributes most to the lesion classification; on the feature map F of a certain layer in the classification network, Represents X i The corresponding preliminary category activation map is calculated as follows: where F k represents the kth channel of the feature map F, is a gradient operator; P avg represents the global average pooling operation; ReLU is a linear rectification function used to filter out negative values ​​in the gradient.

4. The unsupervised lesion segmentation method based on large model and pseudo-label learning as claimed in claim 1, characterized in that: The enhanced category activation map in step 3 is generated as follows: For each image marked y i =1 image X i , through the preliminary category activation map Q 0 Perform adaptive occlusion-based data augmentation; The specific method is to first pass a threshold τ to Q 0 Convert to a binary mask M i : Then the input image X i Medium i The area with value 1 is set to 0, and the occluded image is obtained: For the y i =0 image X i , then by randomly changing X i The occluded image is obtained by setting a sub-region to zero The dataset after occlusion enhancement is represented as Use D2 to retrain the classification network through supervised training based on the cross entropy loss function; After training is completed, the label is re-set to y using Formula 5 i = 1 generates an enhanced category activation map, denoted as and compared to, More marginal areas of the lesion can be excavated with higher quality.

5. The unsupervised lesion segmentation method based on large model and pseudo-label learning as claimed in claim 1, characterized in that: The segmentation pseudo-labels in step 4 are generated as follows: For y i =1 training image X i , from its corresponding enhanced class activation map Generate two kinds of prompt information: one is calculation The bounding box of the area above τ in the figure is used to obtain the bounding box hint. The second is to use the center point of the bounding box as the foreground point and its four vertices as the background points to form a point hint. These two hint information are combined to form a point hint. Represents X i Corresponding prompt information; Then and X i Input the segmentation model SAM model to generate segmentation pseudo-annotations; the SAM model contains an image encoder E m , a prompt information encoder E prom and a decoder D m To obtain the segmentation mask of the image; for a given image X i , get the predicted mask M i ′ The formula is as follows: For each y i =1 image X i , use formula 8 to obtain its segmentation mask as the segmentation pseudo label; for the labeled y i = 0, the segmentation mask is set to empty; Finally, the masks of all 2D slices in a 3D medical image are stacked to form a complete 3D segmentation pseudo label S i ; The corresponding 3D image segmentation training dataset is expressed as Where N v Represents the number of 3D images in the training dataset.

6. The unsupervised lesion segmentation method based on large model and pseudo-label learning as claimed in claim 1, characterized in that: The iterative training method for the segmentation model based on pseudo-label screening in step 5 is: Based on the segmentation pseudo-labels, a segmentation model is trained using an iterative training method with two stages. In the first stage, the 3D image segmentation dataset D3 with pseudo-label annotations is used as the training set, and a 3D segmentation network θ is trained using the standard supervised image segmentation model training strategy. s ; This process uses Dice loss function and cross entropy loss function to update the model parameters; Training θ s Then, use its model parameters to make predictions in all training set images to get V i The updated pseudo-labels for For slices with non-zero areas in the , calculate the bounding box of the non-zero area in each slice, generate SAM hints using the bounding box area and its center point and corner points in the same way as in step 4, input the slice and the corresponding hint into SAM, and obtain the correction result of SAM; The results in each slice are stacked, and the resulting three-dimensional segmentation correction result is recorded as To evaluate The mass of and The higher the similarity, the The more reliable the similarity index is, the more reliable the similarity index is; the similarity index is defined as the Dice value between the two: in The value of is between 0 and 1. The closer the value is to 1, the and The higher the consistency between them; define a similarity threshold It is used to select pseudo labels with high reliability; the selected three-dimensional image and its pseudo label are recorded as in Indicates the threshold The number of selected 3D image samples; In the second stage, the segmentation model is trained again using the selected pseudo-label dataset D4. The training process also uses the standard supervised training method, using the Dice loss function and the cross entropy loss function to update the model parameters, thereby obtaining the final lesion automatic segmentation model.

Images