Cervical cancer MRI (Magnetic Resonance Imaging) image segmentation method for improving U-Net structure based on feed-forward channel double-attention mechanism

Abstract

The invention relates to the technical field of computer image processing and medical image analysis, in particular to a cervical cancer MRI (Magnetic Resonance Imaging) image segmentation method for improving a U-Net structure based on a feed-forward channel double-attention mechanism. In order to solve the problems that a traditional U-Net structure is weak in multi-scale information extraction capability, inaccurate in boundary fuzzy region recognition and the like in processing female abdominal cervical cancer MRI images, the method proposes that a feedforward connection mechanism and a double-attention mechanism are embedded into an encoder and a bottleneck module to form a novel U-Net segmentation model; the recognition and segmentation precision of the cervical cancer focus area is improved, and a high-quality image basis is provided for subsequent clinical diagnosis and treatment. A double-attention mechanism is integrated into a plurality of key nodes of the model, a joint channel-space attention module is added to the tail of each convolution module in an encoder, a cavity space pyramid pooling structure is introduced to a bottleneck position, a channel attention mechanism is embedded, and joint modeling of cross-scale, multi-channel and space context information is achieved.

Description

TECHNICAL FIELD

[0001] The application relates to the fields of computer technology and medical image analysis, and particularly relates to a cervical cancer MRI image segmentation method based on a feedforward channel double attention mechanism improved U-Net structure. BACKGROUND

[0002] Cervical cancer is one of the common high-incidence malignant tumors that endanger women's health, and its early screening, diagnosis and accurate positioning are of great significance for improving the cure rate, optimizing the treatment path and reducing the mortality rate. Especially in the early stage of tumor treatment, doctors need to rely on high-precision cervical lesion segmentation results to develop individualized radiotherapy plans and surgical plans, therefore, automatic and intelligent cervical cancer image segmentation technology has become an important research direction in the field of medical image processing in recent years.

[0003] Medical image segmentation methods can be broadly divided into two categories: traditional image processing algorithms and intelligent methods based on deep learning. Traditional methods mainly rely on the bottom layer pixel information of images and classic image processing methods such as threshold segmentation, edge detection, region growing, K-means clustering and graph cut algorithm. In the aspect of feature extraction, gray level co-occurrence matrix (GLCM) and local binary pattern (LBP) are often used to model image texture and structure. These methods have certain effects in the case of clear image structure and obvious boundary, and have low computational complexity and simple implementation, but they are highly sensitive to image quality and pre-processing, especially in cervical cancer MRI images, due to the complex structures such as uneven gray scale, tissue adhesion and blurred lesion edge, which leads to large segmentation error and discontinuous edge of traditional algorithms, seriously affecting the accuracy of subsequent clinical diagnosis.

[0004] With the rapid development of artificial intelligence, deep learning image segmentation methods based on convolutional neural network (CNN), graph convolution network (GCN), self-attention mechanism (Self-Attention) and Transformer structure have gradually replaced traditional image processing techniques and become the mainstream path of medical image segmentation. Typical networks such as U-Net, FCN, SegNet and their variants have significantly improved image segmentation accuracy through deep feature extraction and skip connection mechanism. Among them, the U-Net structure has been widely used in cervical cancer segmentation tasks due to its symmetrical encoder-decoder design and effective fusion of local and global information. However, the original U-Net structure still has the following technical bottlenecks in dealing with complex pathological conditions such as multi-scale targets, small volume lesions and blurred boundaries: first, the ability to extract features in detail is insufficient; second, it cannot fully utilize the context global semantic information; third, it lacks understanding of anatomical structure priors, leading to target positioning deviation or tissue recognition error.

[0005] Researchers have optimized the U-Net architecture with various structural optimizations and embedded mechanisms, such as attention mechanisms, multi-scale pyramid structures, and dilated convolutions, to enhance the model's feature representation and generalization capabilities. As described in patent CN119693646A, a multi-angle enhancement method for cervical cancer image segmentation is proposed. This method, based on the Swin-Transformer architecture, embeds a multi-angle attention mechanism in the backbone network, extracting and fusing features from multiple angles, including the key lesion region, its surrounding area, horizontal direction, and vertical direction. By designing a module to enhance the positive and negative lesion regions, the model's ability to perceive changes in tissue boundaries and lesion gaps is improved, thus achieving higher segmentation accuracy in complex image scenes. This method not only improves the model's responsiveness to specific lesions through structural design but also enhances network robustness through information augmentation mechanisms, exhibiting strong adaptability to multi-angle data and varied pathological structures.

[0006] Another improved approach is the cervical cancer target segmentation method based on anatomical prior location information disclosed in patent CN119810118A. This method utilizes cervical cancer CT images and their clinical CTV target area data to construct a multi-task bi-branch feature fusion network (MTF-Net). While extracting image features, this model introduces a distance map between the CTV and image edges as auxiliary input, effectively enhancing the modeling ability of anatomical boundaries through a spatial attention mechanism, thereby achieving structural-level target segmentation. This model employs a bi-branch structure to learn target area and anatomical guidance features in parallel, taking into account both global location information and local texture features during segmentation, greatly improving the localization accuracy in anatomically complex regions. It is particularly suitable for scenarios with dense tissues and organs and significant structural interference in cervical region CT images.

[0007] Another approach, such as CN102999917A, provides an automatic cervical image segmentation method that combines image feature enhancement with a multi-region synthesis strategy. This technique employs a multi-channel image feature extraction mechanism and introduces a collaborative optimization strategy among multiple regions during the region synthesis stage, enabling the model to maintain high segmentation consistency and boundary clarity when handling heterogeneous lesion regions. By constructing a region-level synthetic map based on inter-regional similarity and anatomical structure priors, this method effectively overcomes the problem of insufficient robustness in traditional methods due to large differences in lesion distribution, making it particularly suitable for scenarios with large morphological differences and drastic grayscale changes in cervical images.

[0008] Despite the significant improvements made by the aforementioned methods in structural design, feature extraction, and mechanism fusion, existing segmentation models still generally suffer from the following technical challenges: 1. Insufficient multi-scale feature fusion capability: Most network structures employ shallow or fixed-scale feature fusion methods during segmentation, making it difficult to comprehensively capture information about lesions at multiple scales in the image. This is especially true when the cervical region exhibits diverse tissue morphology and significant differences in lesion size. Traditional models struggle to simultaneously handle large lesions and fine boundary features, resulting in discontinuous segmentation output shapes or artifacts. 2. Weak ability to identify blurred edge regions: Low contrast, blurred boundaries, and overlapping adhesions are common features in cervical cancer images and are key factors leading to segmentation failure. While existing methods have improved this issue to some extent through attention mechanisms, they still cannot completely resolve the defects of inaccurate edge localization and insufficient lesion penetration, leading to missed detections or over-segmentation in the segmentation results. 3. Insufficient prior modeling of structural anatomy: Most deep learning models rely solely on image features during the design process, failing to effectively integrate prior clinical knowledge such as structural guidance and spatial constraints. This results in poor performance when adjacent organs or lesions overlap, making it difficult to meet the accuracy requirements of clinical radiotherapy target areas. 4. Poor generalization ability across devices and centers: Existing segmentation models perform excellently on data from specific centers or devices, but due to biased training data or differences in image acquisition conditions, their performance on external validation sets is often unstable, limiting their usability in large-scale clinical deployment. 5. Large training data dependency and low inference efficiency: Current models rely heavily on large amounts of labeled data to achieve ideal performance. However, medical image labeling is time-consuming and labor-intensive, and high-quality labeled data is difficult to obtain in actual clinical practice. High model complexity and long inference time also restrict their real-time deployment and clinical application. To address these common technical problems in the field and comprehensively improve the segmentation performance of models in multi-source, multi-scene cervical images, thereby enhancing their application value in clinical-assisted radiotherapy and preoperative planning, this invention is presented. Summary of the Invention

[0009] The purpose of this invention is to address the problems of insufficient feature extraction ability, poor boundary blurring and recognition, and weak multi-scale information fusion ability in the current automatic segmentation of cervical cancer MRI images. To propose a cervical cancer segmentation method based on a feedforward dual attention improved U-Net model, the method aims to significantly improve the segmentation accuracy and robustness of the model in complex scenarios.

[0010] To overcome the shortcomings of the prior art, the present invention adopts the following technical solution:

[0011] A method for segmenting cervical cancer MRI images based on a feedforward dual attention mechanism and an improved U-Net structure includes the following steps:

[0012] S1: Collect MRI image data of cervical cancer and preprocess the collected images;

[0013] S2: Construct an improved U-Net network model and apply it to image training and validation;

[0014] S3: Input the validation set of cervical cancer images into the trained model and output the lesion segmentation results.

[0015] Furthermore, the specific steps of image preprocessing in S1 are as follows:

[0016] S1.1, MRI image data of 120 female patients with cervical cancer were collected, with a total of 1415 images;

[0017] S1.2, a grayscale processing method is used to convert all cervical cancer images from color to single-channel images, simplifying processing complexity and preserving structural details;

[0018] S1.3, crop all images to a uniform 256×256 resolution, standardizing the size to meet the model input requirements.

[0019] Furthermore, the steps in S2 for constructing the feedforward channel dual-attention improved U-Net model are as follows:

[0020] S2.1: Based on the traditional U-Net network, a basic network structure including encoders, decoders and skip connections is constructed;

[0021] S2.2: Embed a Joint Spatial and Channel Attention Module (JSA) after each downsampling layer in the network encoding stage to simultaneously enhance the model's responsiveness to inter-channel information and spatial region features;

[0022] S2.3: Introduce a Spatial Enhanced Pyramid Pooling (SEP) module at the bottleneck layer to enhance the network's ability to extract and fuse multi-scale contextual information;

[0023] S3.1, The dataset is divided into training set and validation set in a ratio of 7:3;

[0024] S3.2, The Adam optimizer is used during training, with the initial learning rate set to 0.0001 and the batch size set to 4;

[0025] S3.3 uses three metrics to evaluate model performance: Dice coefficient, IoU, and accuracy. Dice coefficient is used to evaluate the degree of overlap of segmented regions, IoU measures the degree of matching between the prediction and the real region, and accuracy reflects the accuracy of lesion identification.

[0026] Furthermore, the specific design of the JSA module is as follows:

[0027] The channel attention branch first obtains the global features of each channel through global average pooling and inputs them into two cascaded fully connected layers. The layers are first activated by ReLU and then output channel weighting coefficients through the sigmoid function. The spatial attention branch compresses the feature map through the channels and inputs it into the convolutional layer and the sigmoid activation function to generate a spatial location weight map. Finally, the input feature map is multiplied channel by channel by channel attention weights and then element by element by spatial attention weights to achieve double weighting and output a weighted and enhanced feature map.

[0028] Furthermore, the SEP module is constructed as follows:

[0029] This module comprises five parallel branches. Four of these branches employ dilated convolution operations with dilation rates of 1, 6, 12, and 18, respectively, combined with 1×1 convolutions, BatchNorm, and ReLU activation layers. The fifth branch is a global context feature extraction branch, which performs global average pooling on the input feature map before convolution, normalization, and activation processing, ultimately restoring it to its original size through bilinear interpolation. The feature maps output from the five branches are concatenated along the channel dimension to form a multi-scale context fusion feature map. This fusion feature map is then input into a channel attention module for channel-dimensional weighted filtering to enhance the expressiveness of key semantic regions. Finally, a 1×1 convolution is used to reduce the dimensionality of the concatenated feature map channels, resulting in a semantically rich and structurally clear final feature map.

[0030] Furthermore, the model proposed in this invention has the following beneficial effects:

[0031] 1. This approach significantly enhances feature extraction and representation capabilities by guiding the network through a dual attention mechanism, particularly in the identification of complex backgrounds and lesions with blurred edges, which is significantly superior to existing models;

[0032] 2. The JSA module integrates channel and spatial attention mechanisms, enabling the model to focus more on key areas of cervical cancer during the encoding stage, while suppressing irrelevant background information;

[0033] 3. The SEP module effectively achieves joint perception of multi-scale lesion regions by introducing dilated convolution and global context information, maintaining the integrity of segmentation edges while avoiding information loss; Attached Figure Description

[0034] The invention will be further understood from the following description taken in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but rather the emphasis is on illustrating the principles of the embodiments. In different views, the same reference numerals designate the same parts.

[0035] Figure 1 This is a schematic diagram of the feedforward channel dual-attention improved U-Net network of the present invention.

[0036] Figure 2 This is a schematic diagram of the JSA module structure of the present invention.

[0037] Figure 3 This is a schematic diagram of the SEP module structure of the present invention.

[0038] Figure 4 This table shows the verification and comparison results of the feedforward channel dual-attention improved U-Net network segmentation of cervical cancer according to the present invention.

[0039] Figure 5 This is a schematic diagram comparing the results of segmenting cervical cancer using different models of the present invention. Detailed Implementation

[0040] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0041] Example: According to Figures 1 to 3 This embodiment provides a cervical cancer MRI image segmentation method based on a feedforward dual attention mechanism to improve the U-Net structure, including the following steps:

[0042] S1. Collect MRI image data of cervical cancer and preprocess the collected images;

[0043] Specifically, this embodiment uses MRI image data of 120 female patients with cervical cancer from Liuzhou Workers' Hospital as the research object. The image source is uniform and the image quality is stable, ensuring the effectiveness and scientific nature of the model training.

[0044] S1.1: MRI image data of 120 female patients with cervical cancer were collected, with a total of 1415 images;

[0045] The acquired MRI image slices are all cross-sectional images of T2-weighted sequences in PNG format. The images have the typical medical imaging features of real clinical complexity, such as blurred tumor boundaries, interference from adjacent tissue structures, and uneven grayscale, which can effectively simulate the actual clinical diagnostic environment.

[0046] S1.2: A grayscale conversion method is used to convert all cervical cancer images from color to single-channel images, simplifying the processing complexity and preserving structural details;

[0047] In the image preprocessing stage, all original color images are first converted to grayscale using a weighted average formula 1:

[0048] Gray(i,j)=0.299*R(i,j)+0.578*G(i,j)0.114*B(i,j) (1)

[0049] This grayscale conversion method can effectively preserve the most important brightness information in the image while removing interfering color components, which helps subsequent deep learning networks focus on the structural differences in the lesion area.

[0050] S1.3: Crop all images to a uniform 256×256 resolution, standardizing the size to meet the model input requirements.

[0051] Image size standardization is a crucial step in ensuring consistent model input dimensions. In this invention, all images are uniformly cropped to 256×256 pixels. During the cropping process, priority is given to preserving the content of the central region of the image, and suspected lesion areas are preserved as much as possible. For some images, a strategy of center cropping combined with mirror expansion is used to both meet the network input dimension requirements and preserve the spatial distribution information of lesions in the image to the greatest extent.

[0052] S2: Construct an improved U-Net network model and apply it to image training and validation;

[0053] To improve the ability to identify complex cervical cancer lesions, an improved U-Net structure model based on a feedforward dual attention mechanism is proposed, including the following steps:

[0054] S2.1: Based on the traditional U-Net network, a basic network structure including encoders, decoders and skip connections is constructed;

[0055] This invention employs a symmetric encoder-decoder U-Net architecture as the basic network framework. The network structure consists of five downsampling layers (encoder) and five upsampling layers (decoder), with skip connections placed at each pair of symmetrical positions to ensure that low-level features can participate in high-level semantic reconstruction. All convolutional operations use 3×3 kernels, and the ReLU activation function is employed. A weighted fusion of channel attention and spatial attention (JSA) attention module and a dilated convolutional module with channel attention added (SEP) are added to the U-Net network structure, such as... Figure 1 As shown, both modules are applied to the entire segmentation network. The model maintains the U-Net encoder-decoder framework in its overall structure, while introducing a joint attention mechanism with the bottleneck layer in the encoding stage. This will improve the model's segmentation ability and significantly enhance the information representation ability of the channel and spatial dimensions in the feature extraction stage.

[0056] S2.2: Embed a Joint Spatial and Channel Attention Module (JSA) after each downsampling layer in the network encoding stage to simultaneously enhance the model's responsiveness to inter-channel information and spatial region features;

[0057] The JSA module design is inspired by the SE attention mechanism and the CBAM attention mechanism. The module structure is as follows: Figure 2 As shown, it includes the following steps:

[0058] The channel attention part first performs global average pooling on the input features to extract the global statistics of each channel; then it generates weight tensors through a two-layer fully connected network to control the response level of each channel in the next stage; the spatial attention part models the spatial distribution based on the channel weighting results of the feature map through a 7×7 convolutional layer; the channel weighting and spatial weighting are used together to finally generate a JSA attention map, which is used to weight the output of the input feature map.

[0059] The JSA module is a new attention module that combines channel attention and spatial attention with weighted computation. Image features are taken as input and fed into both attention modules. The channel attention module first uses an adaptive average pooling layer (AdaptiveAvgPool2d) to pool each channel of the input feature map and calculate the average value for each channel. The pooling operation for the input tensor is as follows:

[0060]

[0061] Where B is the batch size, C is the number of channels, H and W are the height and width of the image, and X is the width and height of the image. cThis is the average value of channel c, representing the global information of that channel. The input high-dimensional feature map is compressed into a 1×1 feature map using global average pooling, and the output is a tensor of shape B×C×1×1. Then, a fully connected (FC) layer is used, with two linear layers weighting the channels. The input to the fully connected layer is the pooled result X. c After passing through the first fully connected layer, we get:

[0062] y1=W1X c +b1 (3)

[0063] Where W1 is the weight of the first layer, b1 is the bias term, and y1 is the output. Then, it undergoes non-linear processing using the ReLU activation function, followed by channel weights generated using the Sigmoid activation function. Spatial attention employs a convolutional layer and an activation layer to generate spatial attention. The input is the average value of the channel-weighted feature map. The convolution operation helps learn the spatial weights. The spatial attention weights are obtained through a convolutional layer, which captures the relationships between spatial locations, generating a spatial weight matrix of shape B×1×H×W, which is then activated using the Sigmoid activation function. Finally, the channel-weighted feature map is multiplied by the spatial attention weights to obtain the final weighted feature map. After this processing, the spatially weighted feature map The shape is the same as the input feature map, still B×C×H×W.

[0064] By introducing the JSA module, the model can more accurately focus on the tumor region during the encoding stage, suppress background and adjacent tissue interference, and improve the robustness of the network.

[0065] S2.3: Introduce a Spatial Enhanced Pyramid Pooling (SEP) module at the bottleneck layer to enhance the network's ability to extract and fuse multi-scale contextual information;

[0066] The SEP module mainly consists of four dilated convolution branches with different dilation rates and a global context branch, with the following structure: Figure 3 As shown. The specific processing steps are as follows:

[0067] The four dilated convolutional branches use 3×3 convolutional kernels with dilation rates of 1, 6, 12, and 18, respectively. Each branch contains a convolutional layer, a BatchNorm layer, and a ReLU activation function. The global branch extracts overall features through global average pooling, and then restores them through 1×1 convolution and upsampling. The outputs of all branches are concatenated along the channel dimension. The concatenated feature map is fed into the channel attention module, where information filtering and enhancement are performed through multiplication operations. Channel compression is achieved through 1×1 convolution to ensure that the output dimension is consistent with that of the encoder.

[0068] The SEP module divides the input feature map into multiple parallel processing branches and introduces a channel attention mechanism in the fusion stage to enhance feature representation capabilities. The overall structure consists of five main branches. The SEP module network structure is as follows: Figure 3 As shown, there are four dilated convolution branches with different dilation rates, and one global context feature extraction branch. The first four branches use 11-fold convolution (without dilation) and 33-fold dilated convolution operations, with dilation rates of 1, 6, 12, and 18, respectively, to capture contextual information under different receptive fields. Dilated convolution effectively improves the model's ability to recognize multi-scale targets by adjusting the receptive field size without increasing additional computation. Each branch consists of convolution, BatchNorm, and ReLU nonlinear activation, maintaining the stability and nonlinear expression of local features. The essence of dilated convolution is the use of a gapped convolution kernel. Mathematically, it is:

[0069]

[0070] Where y(p) is the value of the output feature map at position p (i.e., the p-th pixel) after the convolution operation, x is the input feature map, p is the position coordinate of the current output pixel, w(k) is the weight (value) at position k in the convolution kernel, i.e., the k-th element in the convolution kernel, k is the position of the convolution kernel, and r is the dilatation rate. The fifth branch is used to extract global context information. This branch first performs global average pooling on the input feature map to compress the spatial information into 1×1 features; then it uses 1×1 convolution, BatchNorm, and ReLU to further compress the channel dimension; finally, it restores the global features to the same spatial size as the original input feature map through bilinear interpolation upsampling. This step supplements the global information that is difficult to obtain through local convolution and provides global background assistance.

[0071] The SEP module can significantly improve the model's ability to perceive multi-scale, irregular tumor structures, especially in regions with blurred boundaries or complex tumor morphology, where it has good adaptability.

[0072] S3: Input the validation set of cervical cancer images into the trained model and output the lesion segmentation results.

[0073] S3.1: The dataset is divided into training and validation sets in a ratio of 7:3;

[0074] This invention divides 1415 cervical cancer images into a training set and a validation set at a random ratio. The training set is used for updating model parameters, and the validation set is used for testing the model's generalization ability. The training set contains 990 images, and the validation set contains 425 images. The division method ensures randomness and representativeness, and minimizes cross-interference between the training and validation sets.

[0075] S3.2: The Adam optimizer is used during training, with the initial learning rate set to 0.0001 and the batch size set to 4;

[0076] The Adam optimizer was used during training to accelerate model convergence. The initial learning rate was set to 0.0001, the batch size was 4, and the training epochs were 200. To prevent overfitting, an Early Stopping mechanism was introduced during training, which automatically terminated training when the accuracy on the validation set did not improve for 20 consecutive epochs.

[0077] The beneficial effects of this invention are as follows: The introduction of the Joint Spatial Channel Attention (JSA) module enables the model to adjust attention weights simultaneously in both spatial and channel dimensions, significantly improving the ability to identify lesion regions; the SEP module enhances the model's understanding of multi-scale targets and contextual information by fusing dilated convolution and channel attention mechanisms, thereby improving the accuracy of segmentation results; the overall network structure retains the efficient skip connection mechanism of the U-Net structure, and the improvements effectively enhance the model's performance in complex structural images;

[0078] The above-disclosed content is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the description and drawings of the present invention are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops. The above units are merely examples, and those skilled in the art can adopt corresponding units according to actual needs when implementing this solution.

Claims

1. A method for segmenting cervical cancer MRI images based on a feedforward dual-attention mechanism to improve the U-Net structure, characterized in that, Includes the following steps: S1, Preprocessing the acquired cervical cancer MRI images; S2, an improved U-Net network structure is constructed based on the feedforward channel dual attention mechanism, and the processed MRI image data is divided into training set and test set, and the improved model is trained using the training set; S3 applies the trained model to the test set images and outputs the cervical cancer lesion segmentation results.

2. The cervical cancer MRI image segmentation method based on the improved U-Net structure using a feedforward dual attention mechanism according to claim 1, characterized in that, Step S1 includes the following processing operations: S1.1, MRI image data of 120 female patients with cervical cancer were collected, with a total of 1415 images; S1.2, a grayscale processing method is used to convert all cervical cancer images from color to single-channel images, simplifying processing complexity and preserving structural details; S1.3, crop all images to a uniform 256×256 resolution, standardizing the size to meet the model input requirements.

3. The cervical cancer MRI image segmentation method based on the improved U-Net structure using a feedforward dual attention mechanism according to claim 2, characterized in that, The construction of the improved U-Net network model based on the feedforward channel dual attention mechanism in step S2 includes the following steps: S2.1: Based on the traditional U-Net network, a basic network structure including encoders, decoders and skip connections is constructed; S2.2: Embed a Joint Spatial and Channel Attention Module (JSA) after each downsampling layer in the network encoding stage to simultaneously enhance the model's responsiveness to inter-channel information and spatial region features; S2.3: A Spatial Enhanced Pyramid Pooling (SEP) module is introduced at the bottleneck layer to enhance the network's ability to extract and fuse multi-scale contextual information.

4. The cervical cancer MRI image segmentation method based on the improved U-Net structure using a feedforward dual attention mechanism according to claim 3, characterized in that, The JSA module has the following specific structure: The input feature map is first processed by adaptive average pooling, pooling each channel to extract global channel feature information. The pooling result is then input into a neural network consisting of two fully connected layers, where ReLU and Sigmoid activation functions are used to generate channel attention weights. The channel-weighted feature map is then input into the spatial attention branch, where a convolution operation is performed to generate a spatial weight map. The spatial weights are then multiplied element-wise with the channel-weighted feature map to generate the final weighted output feature map. This module simultaneously assigns response weights to both the channel and spatial dimensions of the original input map, enhancing the ability to focus on lesion regions.

5. The cervical cancer MRI image segmentation method based on the improved U-Net structure using a feedforward dual attention mechanism according to claim 4, characterized in that, The SEP module includes the following steps: the input feature map is divided into five independent branches; four of these branches are convolutional with dilated kernels of 1, 6, 12, and 18 dilation rates, respectively, to capture contextual information at different scales; the fifth branch compresses the spatial dimension through global average pooling, extracts global contextual information, and restores the image to its original size through 1×1 convolution and upsampling; the feature maps output from the five branches are concatenated along the channel dimension to obtain a multi-scale fused feature map; the fused feature map is input to the channel attention module, where key channels are enhanced by weighting to suppress redundant information; the weighted feature map is then subjected to 1×1 convolution for dimensionality reduction, outputting a semantically rich fused feature map for subsequent network use.

6. The cervical cancer MRI image segmentation method based on the improved U-Net structure using a feedforward channel dual attention mechanism according to claim 5, characterized in that, The following configuration parameters are used during the model training process: S3.1, The dataset is divided into training set and validation set in a ratio of 7:3; S3.2, The Adam optimizer is used during training, with the initial learning rate set to 0.0001 and the batch size set to 4; S3.3 uses three metrics to evaluate model performance: Dice coefficient, IoU, and accuracy. Dice coefficient is used to evaluate the degree of overlap of segmented regions, IoU measures the degree of matching between the prediction and the real region, and accuracy reflects the accuracy of lesion identification.

Images