A multi-parameter magnetic resonance imaging classification method for liver cancer based on residual network
Through the multi-parameter NMR image classification method based on residual network, the multi-parameter feature extraction module, asymmetric receptive field module and multi-level feature fusion module are used to solve the problem of confusion in image features in IMCC and HCC diagnosis, and the accurate classification and pathological diagnosis assistance of the two images are achieved.
Patent Information
- Application Number
- CN202310330284.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-03-30
- Publication Date
- 2025-05-27
- Estimated Expiration
- 2043-03-30
AI Technical Summary
The prior art is difficult to accurately diagnose mass intrahepatic cholangiocarcinoma (IMCC) and hepatocellular carcinoma (HCC). The imaging characteristics are easily confused and lead to a low diagnosis rate.
The multi-parameter NMR image classification method based on residual network is adopted, and image features are extracted and fused through the multi-parameter feature extraction module (MFE), asymmetric receptive field module (ARF) and multi-level feature fusion module (MLFF) to achieve accurate classification of IMCC and HCC.
The feature extraction and lesion feature recognition capabilities of multi-parameter nuclear magnetic data are improved, and effective image classification of IMCC and HCC on multi-parameter nuclear magnetic images is realized, and doctors are assisted in preoperative pathological diagnosis.
Smart Images
Figure CN116630229B_ABST
Abstract
Description
Technical Field
[0001] The present invention provides a method for classifying multi-parametric nuclear magnetic resonance images of liver cancer based on a residual network, in particular a method for classifying multi-parametric nuclear magnetic resonance images of mass-forming intrahepatic cholangiocarcinoma and hepatocellular carcinoma based on a residual network, belonging to the field of medical image classification by deep learning. Background Art
[0002] Mass-forming intrahepatic cholangiocarcinoma (IMCC) is the second most common primary intrahepatic malignant tumor after hepatocellular carcinoma (HCC). As one of the common subtypes of intrahepatic cholangiocarcinoma (ICC), IMCC accounts for about 78% of ICC. Since patients have no specific symptoms in the early stage of the disease, the clinical diagnosis rate is relatively low. Patients are usually diagnosed at an advanced stage. Since its treatment methods and prognosis are different from those of other liver malignancies, early and accurate diagnosis is very crucial for the treatment of the disease. However, in the key early diagnosis, IMCC is most likely to be misdiagnosed with hepatocellular carcinoma, seriously hindering the development of clinical treatment work.
[0003] Recent studies have pointed out that MRI is the preferred imaging method for the diagnosis, staging, and prognosis evaluation of IMCC / HCC, and the combination of two or more MR parameters can complement each other in terms of time and spatial resolution. The advantages of nuclear magnetic resonance imaging technology are obvious, and it can well judge the size and shape of IMCC lesions, the degree of tumor invasion of adjacent tissues, and whether it invades the biliary system and blood vessels, and then judge whether there is lymph node metastasis. It is of great significance to carry out IMCC / HCC diagnosis research through nuclear magnetic resonance images. However, pre-clinical imaging diagnosis is time-consuming and cumbersome, and is a highly subjective task, usually affected by personal clinical experience, which is disadvantageous for IMCC / HCC images with a high degree of confusion.
[0004] Technically, the residual network is a very effective network for alleviating the vanishing gradient problem, greatly increasing the depth of the network that can be effectively trained. Due to its performance superiority, the residual network is often used in liver cancer imaging-related work. In recent years, significant progress has been made in liver cancer image classification. However, there are relatively few classification methods for IMCC / HCC images, and IMCC lesions often cannot be separated from the environment, making it difficult for the model to accurately extract the morphological feature information of the lesions. We found that the related technologies in the field of camouflage detection have good object recognizability, providing the possibility for the effective recognition of IMCC / HCC lesion features. Camouflage object detection aims to detect target objects that are hidden and camouflaged in the environment and difficult to recognize from the environment, and is also applied to lesion detection and segmentation tasks in the medical field. Therefore, the present invention focuses on multi-parametric nuclear magnetic resonance images and performs IMCC / HCC image classification based on the residual network. Summary of the Invention
[0005] In view of the defects or deficiencies in the prior art and aiming at the above research gap regarding IMCC / HCC, the present invention proposes a method for classifying multi-parametric nuclear magnetic resonance images of liver cancer based on the residual network and its residual network, solving the problem of multi-parametric nuclear magnetic resonance image classification in the early diagnosis of mass-type intrahepatic cholangiocarcinoma and hepatocellular carcinoma, and achieving a good classification effect of distinguishing mass-type intrahepatic cholangiocarcinoma and hepatocellular carcinoma in the field of deep learning.
[0006] The technical solution of the present invention is: a method for classifying multi-parametric nuclear magnetic resonance images of liver cancer based on the residual network, and the method includes the following steps:
[0007] Step1. First, for the preprocessing process, it is divided into two parts;
[0008] The first part is to preprocess each parameter nuclear magnetic resonance image, including reducing the dimension of the nuclear magnetic resonance images of three 3D sequences respectively and screening clear 2D lesion images, cutting the tumor with the maximum radius and in a square specification; scaling the image proportionally to protect the unified size of the tumor shape features; randomly adjusting the contrast, contrast-limited adaptive histogram equalization, median filtering, and flipping to enhance the image;
[0009] The second part is to splice the processed nuclear magnetic resonance images of the three parameters by channels to form a fused image;
[0010] Step 2. Secondly, input the fused multi-parameter nuclear magnetic resonance imaging into the proposed multi-parameter feature extraction module MFE (Multi-parameter Feature Extraction Module, MFE), which is used to bring a larger receptive field and is used to enhance the feature extraction of the fused image. This module takes into account both the depth of the structure and the large receptive field brought by a slightly larger convolutional kernel; select a 5×5 convolutional kernel to bring a larger receptive field, and at the same time use a 3×3 convolutional kernel to increase the depth of the structure, increasing the non-linearity of the structure, so as to obtain more feature information;
[0011] Step 3. Then, input the features at each level into a newly proposed structured medical lesion camouflage detection module, that is, the asymmetric receptive field ARF module (Asymmetric Receptive Field, ARF) for lesion detection; the idea of the asymmetric receptive field ARF module is: use a set of dilated convolutions with different convolutional kernel sizes to simulate the area near the center of the retina, which is sensitive to small spatial displacements and helps to merge more discriminative features of the lesion in the case where the lesion is difficult to distinguish from the background. During this process, this module reduces the loss of feature information and expands the model's feature extraction of the lesion;
[0012] Step 4. Finally, integrate the feature information at each level processed by the ARF module into the multi-level feature fusion module MLFF (Multi-level Feature Fusion Module, MLFF) to generate fused features; this module better fuses the feature information extracted from the middle and deep layers of the network with the high-resolution features of the shallow layer through upsampling operations, and further increases the local context information through two 3×3 convolutions, increasing the network depth and improving the non-linear expression ability, enabling the model to extract more effective lesion features and making the lesion features more discriminative;
[0013] Step 5. Further extract information from the fused features at each level through two residual blocks, and finally classify through a fully connected layer.
[0014] As a further solution of the present invention, in Step 2, a multi-parameter feature extraction module MFE is defined, which is used to enhance the feature extraction of the fused image. The specific steps are as follows:
[0015] Step 2.1. The input fused image first passes through a 5×5 convolutional kernel to obtain more global feature information with a larger receptive field than a 3×3 convolutional kernel;
[0016] Step 2.2. In subsequent feature extraction, further enhance the depth of the network through two serially connected 3×3 convolutional layers, and effectively expand the receptive field by stacking more layers.
[0017] As a further solution of the present invention, in Step 3, the asymmetric receptive field (ARF) module is improved from the RF module in SINet, and there are two aspects of improvements, specifically:
[0018] Step 3.1. Considering that excessive dimensionality reduction or feature contraction will cause a certain degree of information loss, two 1×1 convolutional kernels are used for dimensionality reduction to reduce parameters and feature information loss in a more stable manner before and after dilated convolution respectively.
[0019] Step 3.2. Considering that large convolutional kernels can increase the receptive field of the network in classification tasks, and dilated convolution with larger convolutional kernels in the shallow layer of the structure can enable the ARF module to obtain more global feature information, thereby expanding the model's extraction of lesion features; therefore, the positions of the convolutional layers of the RF module are adjusted, and 3×3, 5×5, and 7×7 dilated convolutions with larger convolutional kernels are placed in shallower positions.
[0020] As a further solution of the present invention, in Step 4, a new multi-level feature fusion module (MLFF) is defined to better fuse multi-feature matrices; the specific steps of Step 4 are as follows:
[0021] Step 4.1. Drawing on the fusion method of features at different levels in the feature pyramid, bilinear interpolation upsampling of different multiples is performed on the features in the middle and deep layers of the network, and the deep and shallow layer features are spliced through the concept of horizontal connection, so as to combine the high and low layer feature matrices at a higher resolution.
[0022] Step 4.2. The middle and deep layer feature information is processed through 3×3 convolutional layers before and after fusion, which deepens the network and correspondingly enhances the network's expression ability.
[0023] On the other hand, the present invention provides a residual network in a multi-parameter nuclear magnetic resonance image classification method for liver cancer based on a residual network. The residual network (Camouflage Classification Residual Network, CCRNet) includes the following modules:
[0024] Multi-parameter feature extraction module (MFE module): used to bring a larger receptive field and increase the feature extraction of the fused images; a 5×5 convolutional kernel is selected to bring a larger receptive field, and at the same time, a 3×3 convolutional kernel is used to increase the depth of the structure, increasing the non-linearity of the structure, so as to obtain more feature information.
[0025] Asymmetric receptive field ARF module: It is used to simulate the area near the center of the retina using dilated convolutions with different convolutional kernel sizes. This area is sensitive to tiny spatial displacements, which helps to combine more discriminative features of lesions in cases where it is difficult to distinguish between lesions and the background. During this process, the module reduces the loss of feature information and expands the model's extraction of lesion features;
[0026] Multi-level feature fusion module MLFF module: It is used to generate fused features; it is used to better fuse the feature information extracted from the middle and deep layers of the network with the high-resolution features of the shallow layer through upsampling operations, and further increase the local context information through two 3×3 convolutions, increasing the network depth and improving the non-linear expression ability, enabling the model to extract more effective lesion features and making the lesion features more discriminative.
[0027] The beneficial effects of the present invention are:
[0028] 1. The present invention improves the extraction of complementary information of multi-parameter nuclear magnetic data through the multi-parameter feature extraction module MFE of the residual network;
[0029] 2. The present invention enhances the discriminability of features through the improved asymmetric receptive field ARF proposed by the residual network, and further accurately identifies lesion feature information;
[0030] 3. The present invention forms more judgmental fused features by combining the feature information from each level through the multi-level feature fusion module MLFF proposed by the residual network;
[0031] 4. The present invention achieves the purpose of effectively classifying and discriminating images of mass-type intrahepatic cholangiocarcinoma and hepatocellular carcinoma on multi-parameter nuclear magnetic data such as arterial phase, delayed phase, and T2WI three-phase fused nuclear magnetic resonance images, and can be applied to assist doctors in preoperative pathological diagnosis of patients. Brief Description of the Drawings
[0032] Figure 1 It is the nuclear magnetic resonance images of the arterial phase, delayed phase, and T2WI phase of the present invention's IMCC;
[0033] Figure 2 It is the network structure diagram of the present invention's CCTNet;
[0034] Figure 3 It is the structure diagram of the multi-parameter feature extraction module MFE of the present invention;
[0035] Figure 4 It is the structure diagram of the asymmetric receptive field ARF module of the present invention;
[0036] Figure 5 It is the structure diagram of the multi-level feature fusion module MLFF of the present invention;
[0037] Figure 6 The experimental ROC curve of the present invention;
[0038] Figure 7 The method flow chart of the present invention. Detailed implementation manners
[0039] Example 1: As Figures 1 - 7 shown, a multi-parameter magnetic resonance imaging classification method for liver cancer based on a residual network, the method includes the following steps:
[0040] Step1. First, for the preprocessing process, it is divided into two parts;
[0041] The first part is to preprocess each parameter nuclear magnetic resonance image (such as Figure 1 the nuclear magnetic resonance data of the arterial phase, delayed phase, and T2WI shown), including reducing the dimension of the three 3D sequence nuclear magnetic resonance images respectively and screening clear 2D lesion images, using the maximum radius and cutting the tumor in a square specification; scaling the image proportionally to protect the unified size of the tumor shape features; randomly adjusting the contrast, contrast-limited adaptive histogram equalization, median filtering, and flipping to enhance the image;
[0042] The second part is to splice the processed nuclear magnetic resonance images of the three parameters by channels to form a fused image;
[0043] Step2. Secondly, input the fused multi-parameter magnetic resonance imaging (MRI) into the proposed multi-parameter feature extraction module MFE, which is used to bring a larger receptive field and is used to enhance the feature extraction of the fused image. This module takes into account both the depth of the structure and the large receptive field brought by a slightly larger convolutional kernel; select a 5×5 convolutional kernel to bring a larger receptive field, and at the same time use a 3×3 convolutional kernel to increase the depth of the structure, increasing the non-linearity of the structure, so as to obtain more feature information; refer to Figure 3 ;
[0044] In the above Step2, a multi-parameter feature extraction module MFE is defined, which is used to enhance the feature extraction of the fused image. The specific steps are as follows:
[0045] Step2.1. The input fused image first passes through a 5×5 convolutional kernel to obtain more global feature information with a larger receptive field than a 3×3 convolutional kernel;
[0046] Step2.2. In the subsequent feature extraction, the depth of the network is further enhanced through two serially connected 3×3 convolutional layers, and the receptive field is effectively enlarged by stacking more layers.
[0047] Step 3. Then, input the features at each level into a newly proposed structured medical lesion camouflage detection module, namely the Asymmetric Receptive Field (ARF) module, for lesion detection. The idea of the ARF module is as follows: Use a set of dilated convolutions with different kernel sizes to simulate the area near the center of the retina, which is sensitive to small spatial displacements and helps to incorporate more discriminative features of the lesion when it is difficult to distinguish from the background. During this process, the module reduces the loss of feature information and expands the model's extraction of lesion features. For details, please refer to Figure 4 ;
[0048] In Step 3, the ARF module is improved from the RF module in SINet and has two aspects of improvement, specifically:
[0049] Step 3.1. Considering that excessive dimensionality reduction or feature contraction will cause a certain degree of information loss, perform dimensionality reduction through two 1×1 convolutional kernels, reducing parameters and feature information loss in a more stable manner before and after the dilated convolution respectively.
[0050] Step 3.2. Considering that large convolutional kernels can increase the network's receptive field in classification tasks, and dilated convolutions with larger convolutional kernels in the shallow layer of the structure can enable the ARF module to obtain more global feature information, thereby expanding the model's extraction of lesion features. Therefore, adjust the position of the convolutional layer in the RF module, and place the dilated convolutions with 3×3, 5×5, and 7×7 convolutional kernels, which have larger convolutional kernels, in a shallower position.
[0051] Step 4. Finally, integrate the feature information at each level processed by the ARF module into the Multi-Level Feature Fusion (MLFF) module to generate fused features. This module fuses the feature information extracted from the middle and deep layers of the network with the high-resolution features of the shallow layer through upsampling operations, and further increases the local context information through two 3×3 convolutions, increasing the network depth and improving the non-linear expression ability, enabling the model to extract more effective lesion features and making the lesion features more discriminative. For details, please refer to Figure 5 。
[0052] In Step 4, a new Multi-Level Feature Fusion (MLFF) module is defined to better fuse multi-feature matrices. The specific steps of Step 4 are as follows:
[0053] Step 4.1. Drawing on the fusion method of features at different levels in the feature pyramid, perform bilinear interpolation upsampling of different multiples on the features in the middle and deep layers of the network, and splice the deep and shallow layer features through the concept of horizontal connection, so as to combine the high and low layer feature matrices at a higher resolution.
[0054] Step 4.2 Process the middle and deep feature information before and after fusion through a 3×3 convolutional layer respectively. While deepening the network, the expressive ability of the network is correspondingly enhanced.
[0055] Step 5. Further extract information from the fused features of each layer through two residual blocks, and finally perform classification through a fully connected layer.
[0056] On the other hand, the present invention provides a residual network in a multi-parameter nuclear magnetic resonance image classification method for liver cancer based on a residual network, as Figure 2 shown in the structural diagram. The residual network includes the following modules:
[0057] Multi-parameter Feature Extraction Module (MFE Module): It is used to bring a larger receptive field and is used to increase the feature extraction of the fused image. Select a 5×5 convolutional kernel to bring a larger receptive field, and at the same time use a 3×3 convolutional kernel to increase the depth of the structure, increasing the non-linearity of the structure, so as to obtain more feature information.
[0058] Asymmetric Receptive Field Module (ARF Module): It is used to use a set of dilated convolutions with different convolutional kernel sizes to simulate the area near the center of the retina, which is sensitive to small spatial displacements and helps to combine more discriminative features of the lesion in the case where it is difficult to distinguish the lesion from the background. During the process, this module reduces the loss of feature information and expands the model's feature extraction of the lesion.
[0059] Multi-level Feature Fusion Module (MLFF Module): It is used to generate fused features. It is used to better fuse the feature information extracted from the middle and deep layers of the network with the high-resolution features of the shallow layer through upsampling operations, and further increase the local context information through two 3×3 convolutions, increasing the depth of the network, improving the non-linear expression ability, enabling the model to extract more effective lesion features, and making the lesion features more discriminative.
[0060] According to the above implementation process, in combination with Figure 2 , the working principle of the present invention is summarized as follows:
[0061] 1. For multi-parameter nuclear magnetic resonance images, first perform data preprocessing in Step 1, including image dimensionality reduction, lesion segmentation, size unification, data augmentation, and splicing by channels.
[0062] 2. Train and predict the image data under the Figure 2 improved model. The improvements include three new modules, namely, the multi-parameter feature extraction module MFE, the asymmetric receptive field module ARF, and the multi-level feature fusion module MLFF, covering the content of Step 2, Step 3, and Step 4.
[0063] Thus, an effective classification model of IMCC / HCC for multi-parameter nuclear magnetic resonance images can be obtained to assist doctors in discriminating the conditions of the two diseases through nuclear magnetic resonance images before surgery.
[0064] To further illustrate the experimental results of the present invention, in addition to Figures 1 - 5 the model display part, the experimental results of the present invention are now shown:
[0065] First of all, the experimental environment of the present invention is a processor Intel(R) Core(TM) i7-10700F CPU@2.90GHz, 64.0GB of memory, a graphics card of NVDIA GeForce RTX3060, and a video memory of 12G. The optimizer uses SGD, and a learning rate decay strategy is adopted. The learning rate is updated with a multiplication factor of 0.1 every 30 epochs.
[0066] According to the experimental steps, it is tested on 103 cases of HCC and ICC nuclear magnetic resonance images, including 55 cases of HCC patients and 48 cases of IMCC patients. The overall classification accuracy of the experiment is 0.9394, and the AUC value is 0.9790. The specific experimental evaluation indexes of IMCC and HCC are shown in Table 1. The ROC curve of this experiment is as Figure 6 shown:
[0067] Table 1 Experimental results of the present invention
[0068]
[0069] It can be seen that the method of the present invention has a good diagnostic classification effect on multi-parameter nuclear magnetic resonance images of IMCC / HCC, and relatively ideal results are also obtained in the experiment with a small amount of data. Moreover, the excellent performance of IMCC / HCC in the AUC value indicates that the method of the present invention can well assist doctors in the imaging diagnosis of the two diseases before surgery. Thus, based on the differences in the diagnosis and prognosis of the two diseases, the present invention has certain clinical significance.
[0070] The specific implementation manners of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above implementation manners, and various changes can be made without departing from the gist of the present invention within the scope of knowledge possessed by those of ordinary skill in the art.
Claims
1. A method for classifying multi-parametric nuclear magnetic resonance images of liver cancer based on a residual network, characterized in that: The method comprises the following steps: Step1. First, for the preprocessing process, it is divided into two parts; The first part is to preprocess each parameter nuclear magnetic resonance image, including reducing the dimension of the nuclear magnetic resonance images of three 3D sequences respectively and screening clear 2D lesion images, cutting the tumor using the maximum radius and in a square specification; Scaling the image proportionally to protect the unified size of the tumor shape features; randomly adjusting the contrast, contrast-limited adaptive histogram equalization, median filtering, and flipping to enhance the image; The second part is to splice the processed nuclear magnetic resonance images of the three parameters by channels to form a fused image; Step2. Secondly, input the fused multi-parametric nuclear magnetic resonance imaging into the proposed multi-parametric feature extraction module MFE, which is used to bring a larger receptive field and is used to enhance the feature extraction of the fused image; select a 5×5 convolutional kernel to bring a larger receptive field, and at the same time use a 3×3 convolutional kernel to increase the depth of the structure, increasing the non-linearity of the structure, so as to obtain more feature information; Step3. Then, input the features at each level into a newly proposed structured medical lesion camouflage detection module, that is, the asymmetric receptive field ARF module for lesion detection; the idea of the asymmetric receptive field ARF module is: use a group of dilated convolutions with different convolutional kernel sizes to simulate the area near the center of the retina, which is sensitive to tiny spatial displacements and helps to merge more discriminative features of the lesion in the case where the lesion and the background are difficult to distinguish. During the process, this module reduces the loss of feature information and expands the model's feature extraction of the lesion; Step4. Finally, integrate the feature information at each level processed by the ARF module into the multi-level feature fusion module MLFF for generating fused features; this module fuses the feature information extracted from the middle and deep layers of the network with the high-resolution features of the shallow layer better through upsampling operations, and further increases the local context information through two 3×3 convolutions, increasing the network depth and improving the non-linear expression ability, enabling the model to extract more effective lesion features and making the lesion features more discriminative; Step5. Further extract information from the fused features at each level through two residual blocks, and finally classify through a fully connected layer.
2. The method for classifying multi-parametric nuclear magnetic resonance images of liver cancer based on a residual network according to claim 1, characterized in that: In Step2, a multi-parametric feature extraction module MFE is defined for enhancing the feature extraction of the fused image. The specific steps are as follows: Step2.
1. The input fused image first passes through a 5×5 convolutional kernel to obtain more global feature information with a larger receptive field than a 3×3 convolutional kernel; Step2.
2. In subsequent feature extraction, further enhance the depth of the network through two serially connected 3×3 convolutional layers to effectively expand the receptive field by stacking more layers.
3. The method for classifying multi-parametric nuclear magnetic resonance images of liver cancer based on a residual network according to claim 1, characterized in that: In Step 3, the Asymmetric Receptive Field (ARF) module is improved from the RF module in SINet, and there are two aspects of improvement, specifically: Step 3.
1. Considering that excessive dimensionality reduction or feature contraction will cause a certain degree of information loss, dimensionality reduction is performed through two 1×1 convolutional kernels, reducing parameters and minimizing the loss of feature information in a more stable manner before and after dilated convolution respectively. Step 3.
2. Considering that large convolutional kernels can increase the receptive field of the network in classification tasks, and dilated convolutions with larger convolutional kernels in the shallow layers of the structure can enable the ARF module to obtain more global feature information, thereby expanding the model's extraction of lesion features; therefore, the positions of the convolutional layers in the RF module are adjusted, and 3×3, 5×5, and 7×7 dilated convolutions with larger convolutional kernels are placed in shallower positions.
4. The method for classifying multi-parameter nuclear magnetic resonance images of liver cancer based on a residual network according to claim 1, characterized in that: In Step 4, a new Multi-Level Feature Fusion (MLFF) module is defined to better fuse multi-feature matrices; the specific steps of Step 4 are as follows: Step 4.
1. Drawing on the feature fusion method of different levels in the feature pyramid, bilinear interpolation upsampling of different multiples is performed on the middle and deep features in the network, and the deep and shallow features are concatenated through the concept of horizontal connection, so as to combine the high and low-level feature matrices at a higher resolution. Step 4.
2. The middle and deep feature information is processed through 3×3 convolutional layers before and after fusion, deepening the network while enhancing the network's expressive ability accordingly.
5. A residual network in a method for classifying multi-parameter nuclear magnetic resonance images of liver cancer based on a residual network, characterized in that the residual network includes the following modules: Multi-Parameter Feature Extraction (MFE) module: used to bring a larger receptive field and to enhance the feature extraction of fused images; a 5×5 convolutional kernel is selected to bring a larger receptive field, and at the same time a 3×3 convolutional kernel is used to increase the depth of the structure, increasing the non-linearity of the structure, thereby obtaining more feature information. Asymmetric Receptive Field (ARF) module: used to simulate the area near the center of the retina with a set of dilated convolutions with different convolutional kernel sizes. This area is sensitive to small spatial displacements, which helps to combine more discriminative features of the lesion in cases where the lesion and the background are difficult to distinguish. During the process, this module reduces the loss of feature information and expands the model's extraction of lesion features. Multi-Level Feature Fusion (MLFF) module: used to generate fused features; used to better fuse the feature information extracted from the middle and deep layers of the network with the high-resolution features in the shallow layer through upsampling operations, and further increase the local context information through two 3×3 convolutions, increasing the network depth and improving the non-linear expressive ability, enabling the model to extract more effective lesion features and making the lesion features more discriminative.