A liver cancer lesion recognition method and system based on machine vision

Abstract

The application discloses a liver cancer lesion identification method and system based on machine vision, and relates to the technical field of image recognition. A set of continuous CT images of a target liver is acquired, an improved U-Net lesion segmentation model is inputted with any target slice image, and a preliminary lesion segmentation result of the slice is acquired; subsequently, the preliminary segmentation result is reviewed and corrected according to the segmentation results of the adjacent slices before and after the target slice, and a final segmentation result is obtained. The application can accurately identify and locate the liver cancer lesion on a single CT image, and can remove the false positive area by using the adjacent slice information, thereby significantly reducing the false detection risk and improving the stability and reliability of the lesion segmentation.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and specifically to a method and system for identifying liver cancer lesions based on machine vision. Background Technology

[0002] Liver cancer is one of the most common malignant tumors, and timely detection and accurate delineation of lesions are crucial for treatment planning, efficacy evaluation, and patient prognosis. In recent years, with the development of artificial intelligence and deep learning technologies, automatic segmentation methods based on medical images have been increasingly applied to liver CT image analysis to assist doctors in improving image interpretation efficiency and reducing the burden of manual annotation. These technologies typically employ deep models such as convolutional neural networks to automatically identify and segment the liver and lesion areas, thus alleviating the workload of doctors to some extent and increasing the speed of initial screening.

[0003] However, existing image segmentation models still face significant false positives and false negatives in practical applications. Due to CT image noise, the diverse morphology of the liver and lesions, and differences in imaging conditions among patients, models often produce false positives (misclassifying normal tissue or image artifacts as lesions) or false negatives (missing or minimizing the extent of actual lesions). False positives increase the workload of physicians in verification; false negatives may cause lesions to be overlooked, delaying diagnosis and treatment, and affecting prognosis. With the ever-increasing scale of medical imaging data and the continuous growth of clinical needs, how to further reduce the false positive rate of models and improve the stability and consistency of segmentation results on the basis of automatic segmentation has become an important problem that urgently needs to be solved in the field of medical image processing. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of false detection mentioned in the background art, and to propose a machine vision-based method and system for identifying liver cancer lesions.

[0005] A first aspect of this invention provides a machine vision-based method for identifying liver cancer lesions, the method comprising:

[0006] Acquire a series of consecutive CT images of the target liver;

[0007] The target slice image is used as input to a pre-trained lesion segmentation model to obtain the lesion segmentation result of the target slice image, which is denoted as the first segmentation result; the target slice image is any one of the set of continuous CT images; the lesion segmentation model is an improved U-Net model;

[0008] Based on the lesion segmentation results of neighboring slice images of the target slice image, the first segmentation result is reviewed and corrected to obtain the second segmentation result.

[0009] Optionally, the lesion segmentation model, compared to the original U-Net model, includes the following specific improvements:

[0010] Add a downsampling branch to perform multi-step downsampling on the input image and input the scaled versions at each scale into the corresponding coding layer;

[0011] The original U-Net model's encoding and decoding layers are replaced with a residual convolutional module based on residual blocks and attention mechanisms.

[0012] Optionally, the operational expression of the residual convolution module is:

[0013] ;

[0014] Where X0 and X4 are the input and output of the residual convolution module; X2 and X3 are the features generated during the operation. and These represent convolutional layers with kernel sizes of 3×3 and 1×1, respectively; concat represents channel concatenation; GAP represents global average pooling; and FC represents a fully connected layer.

[0015] Optionally, the lesion segmentation model employs an improved loss function during training, the expression of which is:

[0016] ;

[0017] Where Loss is the improved loss function; L CE It is cross-entropy loss; L ED It is an improved Dice coefficient loss; , These are weighting coefficients; It is the set of all pixels in the batch sample; N is the number of pixels in the batch sample; y i It is the true label of pixel i; p i It is the model's predicted probability for pixel i; It is a preset constant; L D It is the Dice coefficient loss for batch samples; L HD It is the Dice coefficient loss for the majority class samples; It is the set of pixels of the majority class in the sample; L MD It is the Dice coefficient loss for minority class samples; It is the set of pixels of the minority class samples in the sample.

[0018] Optionally, the segmentation result is a binary mask image; the step of correcting the first segmentation result based on the segmentation results of neighboring slice images of the target slice image to obtain the second segmentation result includes:

[0019] Perform connected component analysis on the first segmentation result and record each connected region as a candidate lesion region.

[0020] Select a preset number of slice images before and after the target slice image, and record them as neighboring slice images;

[0021] Based on the candidate lesion regions in the target slice image, the corresponding regions in adjacent slice images are determined as the verification regions;

[0022] The overlapping area of ​​the target candidate lesion region and its corresponding verification region is statistically analyzed, and the ratio of the overlapping area to the area of ​​the target candidate lesion region is calculated; the target candidate lesion region can be any candidate lesion region.

[0023] If the percentage of overlapping area is greater than a preset threshold, the target candidate lesion area is determined to be a true positive; otherwise, the target candidate lesion area is determined to be a false positive.

[0024] The false positive candidate lesion areas are removed from the first segmentation result to obtain the second segmentation result.

[0025] A second aspect of this invention provides a machine vision-based liver cancer lesion identification system, the system comprising:

[0026] The data acquisition module is used to acquire a set of consecutive CT images of the target liver;

[0027] The lesion identification and localization module is used to take the target slice image as input to a pre-trained lesion segmentation model to obtain the lesion segmentation result of the target slice image, which is denoted as the first segmentation result; the target slice image is any one of the set of continuous CT images; the lesion segmentation model is an improved U-Net model;

[0028] The false positive removal module is used to review and correct the first segmentation result based on the lesion segmentation results of neighboring slice images of the target slice image, so as to obtain the second segmentation result.

[0029] Optionally, the lesion segmentation model, compared to the original U-Net model, includes the following specific improvements:

[0030] Add a downsampling branch to perform multi-step downsampling on the input image and input the scaled versions at each scale into the corresponding coding layer;

[0031] The original U-Net model's encoding and decoding layers are replaced with a residual convolutional module based on residual blocks and attention mechanisms.

[0032] Optionally, the operational expression of the residual convolution module is:

[0033] ;

[0034] Where X0 and X4 are the input and output of the residual convolution module; X2 and X3 are the features generated during the operation. and These represent convolutional layers with kernel sizes of 3×3 and 1×1, respectively; concat represents channel concatenation; GAP represents global average pooling; and FC represents a fully connected layer.

[0035] Optionally, the system further includes a model training module for training the lesion segmentation model using an improved loss function, the expression of which is:

[0036] ;

[0037] Where Loss is the improved loss function; L CE It is cross-entropy loss; L ED It is an improved Dice coefficient loss; , These are weighting coefficients; It is the set of all pixels in the batch sample; N is the number of pixels in the batch sample; y i It is the true label of pixel i; p i It is the model's predicted probability for pixel i; It is a preset constant; L D It is the Dice coefficient loss for batch samples; L HD It is the Dice coefficient loss for the majority class samples; It is the set of pixels of the majority class in the sample; L MD It is the Dice coefficient loss for minority class samples; It is the set of pixels of the minority class samples in the sample.

[0038] Optionally, the segmentation result is a binary mask image; the false positive removal module includes:

[0039] The connectivity analysis module is used to perform connectivity analysis on the first segmentation result and record each connected region as a candidate lesion region.

[0040] The comparison and selection module is used to select a preset number of slice images before and after the target slice image, which are denoted as neighboring slice images.

[0041] The region determination module is used to determine the corresponding region of the adjacent slice image as the verification region based on the candidate lesion region of the target slice image;

[0042] The overlap calculation module is used to count the overlap area between the target candidate lesion region and the lesion pixels in its corresponding verification region, and to calculate the ratio of the overlap area to the area of ​​the target candidate lesion region; the target candidate lesion region can be any candidate lesion region.

[0043] The determination module is used to determine the target candidate lesion region as a true positive if the percentage of overlapping area is greater than a preset threshold; otherwise, it determines the target candidate lesion region as a false positive.

[0044] The false positive correction module is used to remove the candidate lesion regions that produce false positives from the first segmentation result, thereby obtaining the second segmentation result.

[0045] The beneficial effects of this invention are:

[0046] By improving the U-Net model to identify and locate lesions in a single liver CT image, and combining the segmentation information of adjacent slices to verify and correct the results, false positive areas are eliminated, reducing the risk of misdiagnosis and improving the stability and reliability of liver cancer CT image analysis. Attached Figure Description

[0047] Figure 1 A flowchart illustrating a machine vision-based liver cancer lesion identification method provided in an embodiment of the present invention;

[0048] Figure 2 A network architecture diagram of the original U-Net model provided in an embodiment of the present invention;

[0049] Figure 3 A network architecture diagram of a lesion segmentation model provided in an embodiment of the present invention;

[0050] Figure 4 This is a schematic diagram of the structure of a residual convolution module provided in an embodiment of the present invention;

[0051] Figure 5 This is an architecture diagram of a machine vision-based liver cancer lesion recognition system provided in an embodiment of the present invention. Detailed Implementation

[0052] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0053] This invention provides a machine vision-based method for identifying liver cancer lesions. See also... Figure 1 , Figure 1 A flowchart illustrating a machine vision-based method for identifying liver cancer lesions, provided in an embodiment of the present invention. The method includes the following steps:

[0054] S101, acquire a set of consecutive CT images of the target liver.

[0055] S102, the target slice image is used as input to the pre-trained lesion segmentation model to obtain the lesion segmentation result of the target slice image, which is denoted as the first segmentation result.

[0056] S103, Based on the lesion segmentation results of the neighboring slice images of the target slice image, the first segmentation result is reviewed and corrected to obtain the second segmentation result.

[0057] In this context, consecutive CT images mean that the liver slices are spatially continuous and adjacent. The target slice image is any one of a set of consecutive CT images. The lesion segmentation model is an improved U-Net model, and the segmentation result is a binary mask image.

[0058] The present invention provides a machine vision-based method for identifying liver cancer lesions. By improving the U-Net model, lesions are identified and located in a single liver CT image. The results are then verified and corrected by combining the segmentation information of adjacent slices, eliminating false positive areas, reducing the risk of misdetection, and improving the stability and reliability of liver cancer CT image analysis.

[0059] In one embodiment, see Figure 2 and Figure 3 , Figure 2 This is the network architecture diagram of the original U-Net model. Figure 3 This is a network architecture diagram of a lesion segmentation model provided in an embodiment of the present invention. Compared with the original U-Net model, the specific improvements include:

[0060] 1. A downsampling branch is added to the encoding network to perform multi-step downsampling of the input image. A feature integration module is proposed to input the scaled versions at each scale into the corresponding encoding layer to enhance the multi-scale feature representation capability. Specifically:

[0061] The newly added downsampling branch uses bilinear interpolation to halve the image layer by layer, outputting 1 / 2 scaled, 1 / 4 scaled, and 1 / 8 scaled versions sequentially. The feature integration module's operation process includes: first, using a 3×3 convolutional layer to expand the channels of the scaled image; then concatenating it with the output of the corresponding coding layer; finally, adjusting the number of channels through a 1×1 convolutional layer before inputting it into subsequent coding layers.

[0062] Traditional U-Net relies solely on single-scale image downsampling, and max pooling downsampling inevitably discards fine-grained features (such as tumor edges and local textures). In contrast, image scaling provides each layer of the encoder with original image information of a matching size, allowing each layer to retain tumor details at that resolution without relying solely on the feature map downsampled from the previous layer. This reduces feature loss at the input, improves the semantic integrity of deep features, and enhances the accuracy of identifying and segmenting small lesions and regions with blurred boundaries.

[0063] 2. A residual convolutional module is proposed, replacing the encoder and decoder layers of the original U-Net model. This residual convolutional module enhances feature extraction capabilities by incorporating an attention mechanism into the residual structure. For details, see [link to relevant documentation]. Figure 4 , Figure 4 This is a schematic diagram of a residual convolution module provided in an embodiment of the present invention. The operational expression of the residual convolution module is:

[0064] ;

[0065] Where X0 and X4 are the input and output of the residual convolution module; X2 and X3 are the feature maps generated during the operation. and represents convolutional layers with kernel sizes of 3×3 and 1×1, respectively. Each convolutional layer uses batch normalization and ReLU activation function; concat represents channel concatenation; GAP represents global average pooling; FC is a fully connected layer that uses Sigmoid activation function.

[0066] Traditional U-Net convolutional layers are prone to gradient vanishing or feature degradation in deep networks. This embodiment employs a residual structure, which enhances gradient flow while maintaining the expressive power of deep networks, facilitating stable training and deeper feature extraction. Furthermore, an attention mechanism is introduced into the module, adaptively highlighting key features of lesion regions and suppressing irrelevant or background features, allowing the model to focus on more diagnostically valuable information.

[0067] 3. An improved loss function is proposed, the expression of which is:

[0068] ;

[0069] Where Loss is the improved loss function; L CE It is cross-entropy loss; L ED It is an improved Dice coefficient loss; , These are weighting coefficients, which can be used to... Set to 1, and dynamically adjust during training. The value; It is the set of all pixels in the batch sample; N is the number of pixels in the batch sample; y i It is the true label of pixel i; p i It is the model's predicted probability for pixel i; It is a preset decimal, such as 10. -5 L D This is the Dice coefficient loss for this batch of samples; L HD It is the Dice coefficient loss of the majority class samples in this batch of samples; It is the set of pixels of the majority class samples in this batch of samples; L MD It is the Dice coefficient loss of the minority class samples in this batch of samples; It is the set of pixels of the minority class samples in this batch of samples.

[0070] Different types of liver cancer have significantly different incidence rates in clinical practice, which directly leads to a huge disparity in the number of liver cancer samples collected. For example, there are many samples of primary liver cancer such as hepatocellular carcinoma and intrahepatic cholangiocarcinoma, while there are few samples of malignant tumors formed by metastasis of other cancers to the liver.

[0071] The CT imaging features of different types of liver cancer differ significantly (primary liver cancer is a single large lesion, while secondary liver cancer is multiple small lesions). Traditional loss functions (such as CEDL Loss) treat all lesions as the same class and cannot distinguish these differences. As a result, the model can only learn general features and can only handle the single imbalance between foreground (all lesions) and background (normal liver tissue), but cannot solve the class imbalance between different types of liver cancer. This causes the model to prioritize learning features of the majority class (primary liver cancer) and have poor segmentation performance for the minority class (secondary liver cancer).

[0072] The improved loss function is achieved through L D Continuing the advantages of Dice Loss, this method measures the overlap between all lesions and the background, resolving foreground-background imbalance and ensuring the model does not ignore any lesions; through L... HD and L MD The loss is calculated separately for primary liver cancer and secondary liver cancer, and the model is forced to assign independent learning weights to the minority class (MET). In this way, even if the number of secondary liver cancer samples is small, its segmentation error will be quantified separately and will not be masked by the large sample size of primary liver cancer. This significantly improves the underfitting problem of the minority class.

[0073] In one embodiment, step S103, which involves reviewing and correcting the first segmentation result based on the lesion segmentation results of neighboring slice images of the target slice image to obtain the second segmentation result, includes:

[0074] Step 1: Perform connected component analysis on the first segmentation result and record each connected region as a candidate lesion region.

[0075] Step two: Select a preset number of slice images before and after the target slice image, and call them neighboring slice images. For example, take one image before and one image after the target slice image, for a total of two images as neighboring slice images.

[0076] Step 3: Based on the candidate lesion regions in the target slice image, determine the corresponding regions in neighboring slice images as the verification regions. Specifically, the corresponding regions are obtained by mapping the pixel positions of the candidate lesion regions.

[0077] Step four: For any candidate lesion region, designate it as the target candidate lesion region. Calculate the overlap area between the target candidate lesion region and the lesion pixels in its corresponding verification region, and then calculate the ratio of the overlap area to the area of ​​the target candidate lesion region. Specifically, the area is represented by the number of pixels. Pixels predicted as lesions in all three images are considered overlapping pixels, and the number of overlapping pixels is recorded as the overlap area.

[0078] Step 5: If the percentage of overlapping area is greater than a preset threshold, the target candidate lesion area is determined to be a true positive; otherwise, the target candidate lesion area is determined to be a false positive. The preset threshold can be set to 30%.

[0079] Step six: Remove the false positive candidate lesion regions from the first segmentation result to obtain the second segmentation result.

[0080] The U-Net model relies solely on single-slice information analysis, making it prone to misclassifying isolated noise in CT images (such as low-density shadows in the gallbladder or scanning artifacts) as lesions (false positives). Real liver cancer lesions are three-dimensional entities and inevitably exist across multiple adjacent slices. By verifying the continuity of lesions across adjacent slices, we can accurately screen for real lesions and reduce false positives.

[0081] This invention provides a machine vision-based system for identifying liver cancer lesions. See also... Figure 5 , Figure 5 This is an architecture diagram of a machine vision-based liver cancer lesion recognition system provided in an embodiment of the present invention. The system includes:

[0082] The data acquisition module is used to acquire a set of consecutive CT images of the target liver.

[0083] The lesion identification and localization module is used to take the target slice image as input to the pre-trained lesion segmentation model and obtain the lesion segmentation result of the target slice image, which is denoted as the first segmentation result.

[0084] The false positive removal module is used to review and correct the first segmentation result based on the lesion segmentation results of neighboring slice images of the target slice image, and obtain the second segmentation result.

[0085] In this context, consecutive CT images mean that the liver slices are spatially continuous and adjacent. The target slice image is any one of a set of consecutive CT images. The lesion segmentation model is an improved U-Net model, and the segmentation result is a binary mask image.

[0086] The present invention provides a machine vision-based liver cancer lesion identification system. By improving the U-Net model, the system identifies and locates lesions in a single liver CT image. The system also combines the segmentation information of adjacent slices to verify and correct the results, eliminates false positive areas, reduces the risk of misdetection, and improves the stability and reliability of liver cancer CT image analysis.

[0087] In one implementation, the system further includes a model training module for training the lesion segmentation model using an improved loss function.

[0088] In one implementation, the false positive removal module includes:

[0089] The connectivity analysis module is used to perform connectivity analysis on the first segmentation result and record each connected region as a candidate lesion region.

[0090] The comparison selection module is used to select a preset number of slice images before and after the target slice image, which are denoted as neighboring slice images.

[0091] The region determination module is used to determine the corresponding region of a neighboring slice image as the verification region based on the candidate lesion region of the target slice image.

[0092] The overlap calculation module is used to count the overlapping area of ​​the target candidate lesion region and the lesion pixels in its corresponding verification region, and to calculate the ratio of the overlapping area to the area of ​​the target candidate lesion region.

[0093] The determination module is used to determine the target candidate lesion area as a true positive if the proportion of overlapping area is greater than a preset threshold; otherwise, it determines the target candidate lesion area as a false positive.

[0094] The false positive correction module is used to remove the candidate lesion regions that are false positives from the first segmentation result and obtain the second segmentation result.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall fall within the scope of the present invention.

Claims

1. A machine vision-based method for identifying liver cancer lesions, characterized in that, The method includes: Acquire a series of consecutive CT images of the target liver; The target slice image is used as input to a pre-trained lesion segmentation model to obtain the lesion segmentation result of the target slice image, which is denoted as the first segmentation result; the target slice image is any one of the set of continuous CT images; the lesion segmentation model is an improved U-Net model; Based on the lesion segmentation results of neighboring slice images of the target slice image, the first segmentation result is reviewed and corrected to obtain the second segmentation result; Compared to the original U-Net model, the lesion segmentation model has the following improvements: replacing the encoding and decoding layers of the original U-Net model with a residual convolutional module based on residual blocks and attention mechanisms; the operational expression of the residual convolutional module is as follows: ； Where X0 and X4 are the input and output of the residual convolution module; X2 and X3 are the features generated during the operation. and These represent convolutional layers with kernel sizes of 3×3 and 1×1, respectively; concat represents channel concatenation; GAP represents global average pooling; FC is a fully connected layer. The segmentation result is a binary mask image; the step of correcting the first segmentation result based on the segmentation results of neighboring slice images of the target slice image to obtain the second segmentation result includes: Perform connected component analysis on the first segmentation result and record each connected region as a candidate lesion region. Select a preset number of slice images before and after the target slice image, and record them as neighboring slice images; Based on the candidate lesion regions in the target slice image, the corresponding regions in adjacent slice images are determined as the verification regions; The overlapping area of ​​the target candidate lesion region and its corresponding verification region is statistically analyzed, and the ratio of the overlapping area to the area of ​​the target candidate lesion region is calculated; the target candidate lesion region can be any candidate lesion region. If the percentage of overlapping area is greater than a preset threshold, the target candidate lesion area is determined to be a true positive; otherwise, the target candidate lesion area is determined to be a false positive. The false positive candidate lesion areas are removed from the first segmentation result to obtain the second segmentation result.

2. The method for identifying liver cancer lesions based on machine vision according to claim 1, characterized in that, Compared to the original U-Net model, the lesion segmentation model includes the following improvements: Add a downsampling branch to perform multi-step downsampling on the input image and input the scaled versions at each scale into the corresponding coding layer.

3. The method for identifying liver cancer lesions based on machine vision according to claim 1, characterized in that, The lesion segmentation model uses an improved loss function during training, the expression of which is: ； Where Loss is the improved loss function; L CE It is cross-entropy loss; L ED It is an improved Dice coefficient loss; , These are weighting coefficients; It is the set of all pixels in the batch sample; N is the number of pixels in the batch sample; y i It is the true label of pixel i; p i It is the model's predicted probability for pixel i; It is a preset constant; L D It is the Dice coefficient loss for batch samples; L HD It is the Dice coefficient loss for the majority class samples; It is the set of pixels of the majority class in the sample; L MD It is the Dice coefficient loss for minority class samples; It is the set of pixels of the minority class samples in the sample.

4. A machine vision-based liver cancer lesion recognition system, characterized in that, The system includes: The data acquisition module is used to acquire a set of consecutive CT images of the target liver; The lesion identification and localization module is used to take the target slice image as input to a pre-trained lesion segmentation model to obtain the lesion segmentation result of the target slice image, which is denoted as the first segmentation result; the target slice image is any one of the set of continuous CT images; the lesion segmentation model is an improved U-Net model; The false positive removal module is used to review and correct the first segmentation result based on the lesion segmentation results of neighboring slice images of the target slice image, and obtain a second segmentation result; Compared to the original U-Net model, the lesion segmentation model has the following improvements: replacing the encoding and decoding layers of the original U-Net model with a residual convolutional module based on residual blocks and attention mechanisms; the operational expression of the residual convolutional module is as follows: ； Where X0 and X4 are the input and output of the residual convolution module; X2 and X3 are the features generated during the operation. and These represent convolutional layers with kernel sizes of 3×3 and 1×1, respectively; concat represents channel concatenation; GAP represents global average pooling; and FC represents a fully connected layer. The segmentation result is a binary mask image; the false positive removal module includes: The connectivity analysis module is used to perform connectivity analysis on the first segmentation result and record each connected region as a candidate lesion region. The comparison and selection module is used to select a preset number of slice images before and after the target slice image, which are denoted as neighboring slice images. The region determination module is used to determine the corresponding region of the adjacent slice image as the verification region based on the candidate lesion region of the target slice image; The overlap calculation module is used to count the overlap area between the target candidate lesion region and the lesion pixels in its corresponding verification region, and to calculate the ratio of the overlap area to the area of ​​the target candidate lesion region; the target candidate lesion region can be any candidate lesion region. The determination module is used to determine the target candidate lesion region as a true positive if the percentage of overlapping area is greater than a preset threshold; otherwise, it determines the target candidate lesion region as a false positive. The false positive correction module is used to remove the candidate lesion regions that produce false positives from the first segmentation result, thereby obtaining the second segmentation result.

5. A machine vision-based liver cancer lesion identification system according to claim 4, characterized in that, Compared to the original U-Net model, the lesion segmentation model includes the following improvements: Add a downsampling branch to perform multi-step downsampling on the input image and input the scaled versions at each scale into the corresponding coding layer.

6. A machine vision-based liver cancer lesion identification system according to claim 4, characterized in that, The system also includes a model training module for training the lesion segmentation model using an improved loss function, the expression of which is: ； Where Loss is the improved loss function; L CE It is cross-entropy loss; L ED It is an improved Dice coefficient loss; , These are weighting coefficients; It is the set of all pixels in the batch sample; N is the number of pixels in the batch sample; y i It is the true label of pixel i; p i It is the model's predicted probability for pixel i; It is a preset constant; L D It is the Dice coefficient loss for batch samples; L HD It is the Dice coefficient loss for the majority class samples; It is the set of pixels of the majority class in the sample; L MD It is the Dice coefficient loss for minority class samples; It is the set of pixels of the minority class samples in the sample.

Images