Medical image segmentation method with adaptive receptive field and feature correction

Abstract

The invention relates to the technical field of medical image processing, and particularly discloses a medical image segmentation method with adaptive receptive field and feature correction, which comprises the following steps: (1) acquiring an original medical image and a segmentation label thereof, and constructing a training and testing data set; (2) carrying out size normalization and enhancement processing on the image; (3) establishing an improved U-shaped encoder-decoder segmentation network, introducing an adaptive branch mixed shape convolution module in a shallow layer, and improving edge and texture feature modeling capability by adopting a multi-branch banded convolution and channel attention mechanism; (4) a residual directional feature interaction module is introduced into a deep layer, a spatial dependency relationship is modeled through an information interaction structure in the horizontal and vertical directions, and the direction sensing ability of the heterostructure is enhanced; and (5) completing network training and reasoning, and outputting a segmentation result. The method gives consideration to the calculation efficiency and the segmentation precision, and is suitable for the automatic segmentation task of various types of medical images with complex structures.

Description

Technical Field

[0001] This invention belongs to the field of medical image segmentation technology, and relates to a medical image segmentation method with adaptive receptive field and feature correction. Background Technology

[0002] With the widespread application of medical imaging equipment, computer-aided diagnostic (CAD) systems are playing an increasingly important role in clinical medicine. Medical image segmentation, as a core component of CAD systems, aims to automatically and accurately extract target regions, such as lesions, organs, or tissue structures, from complex medical images, providing doctors with reliable auxiliary information. The accuracy of the segmentation results is crucial for early disease detection, surgical planning, and evaluation of treatment outcomes.

[0003] Traditional medical image segmentation methods, such as thresholding, region growing, graph cutting, and active contour models, are limited by their ability to handle image noise and complex morphological structures, making it difficult to meet the dual requirements of accuracy and stability in clinical applications. In recent years, deep learning-based segmentation methods have developed rapidly, especially convolutional neural networks (CNNs), which have demonstrated powerful feature extraction capabilities in image representation. Among them, the symmetric encoder-decoder structure, represented by U-Net, has become the mainstream framework in the field of medical image segmentation.

[0004] Building upon this foundation, researchers have proposed various improved models to enhance segmentation performance. For example, U-Net++ introduces dense skip connections to enhance feature fusion; Attention U-Net applies an attention mechanism to enhance the network's response to key regions; ResUNet combines residual structures to alleviate the training difficulties of deep networks; and models such as TransUNet introduce Transformer structures to improve global modeling capabilities. These methods have improved segmentation accuracy to varying degrees, but still have the following limitations: 1. Most network structures use fixed convolutional kernels, resulting in limited receptive fields and difficulty adapting to multi-scale and morphologically varied lesion regions in medical images; 2. The fusion methods of the connections lack sufficient selectivity, easily introducing redundant background information, leading to blurred boundaries in the segmentation results; 3. Tissue structures in images exhibit obvious directional distributions (such as blood vessels, tumor boundaries, etc.), while conventional convolutional operations have weak modeling capabilities for directional features, resulting in insufficient utilization of spatial structural information.

[0005] Therefore, there is an urgent need to design an image segmentation method with adaptive receptive field, orientation modeling capability and feature correction mechanism to better address the problems of structural blurring, low contrast and noise interference in actual medical images and improve the performance of segmentation models in complex scenarios. Summary of the Invention

[0006] This invention aims to overcome the shortcomings of existing medical image segmentation methods in multi-scale feature representation, directional information modeling, and feature fusion, and proposes a medical image segmentation method with adaptive receptive field and feature correction capabilities. This method constructs an improved U-shaped encoder-decoder network structure, integrating an Adaptive-Branch Mixed-Shape Convolution (ABMSC) module and a Residual Directional Feature Interaction (RDFI) module. While maintaining the network structure's high efficiency, it significantly improves the ability to recognize detailed textures, directional features, and blurred regions, making it suitable for automatic segmentation tasks of various complex medical image structures.

[0007] To achieve the above objectives, this invention provides a medical image segmentation method with adaptive receptive field and feature correction, implemented through the following technical solution:

[0008] S1. Data Preprocessing

[0009] In the medical image segmentation method of this invention, the original medical images are first subjected to systematic data preprocessing operations to improve the model's adaptability to medical images from different sources and of different qualities, and to enhance the stability and generalization performance of the segmentation results during training. This preprocessing process includes, but is not limited to, the following steps:

[0010] S1-a) Data Acquisition: Obtain medical image samples with clearly labeled lesion areas from clinical medical databases or publicly available medical image datasets. These images include, but are not limited to, dermoscopic images (for skin lesion identification) and breast ultrasound images (for breast mass analysis). All images are calibrated with equipment parameters before acquisition to ensure consistent spatial resolution. Labeling information is reviewed by professional radiologists to ensure label accuracy and clinical reliability.

[0011] S1-b) Random Flip: Apply random horizontal or vertical flipping to the input image to simulate changes in patient position in actual clinical imaging, improve the network's robustness to directional changes, and mitigate the risk of overfitting due to insufficient sample size.

[0012] S1-c) Random rotation: Perform random rotation processing on the image in 90-degree units (including 0°, 90°, 180° and 270°) to expand the orientation diversity of the image and further improve the model's ability to adapt to orientation uncertainty in clinical images.

[0013] S1-d) Resizing and Normalization: All images are uniformly cropped or bilinearly interpolated to a specified input size (e.g., 256×256 pixels) to meet the network structure input requirements and maintain scale consistency. Simultaneously, pixel values ​​are linearly normalized to map them to the standard numerical range of [0,1] or [-1,1], thereby reducing the impact of different imaging devices, exposure conditions, or lighting differences on image grayscale distribution and improving the consistency and stability of feature extraction.

[0014] S2. Segmentation Network Architecture Design and Feature Extraction Strategy

[0015] In this embodiment, an improved U-shaped network structure is constructed, which includes an encoder, a decoder, and multi-level skip connection paths to achieve multi-scale feature extraction and spatial detail restoration. Specifically, it includes the following steps:

[0016] S2-a) Overall Architecture Design: In this embodiment, the proposed medical image segmentation network adopts a symmetrical U-shaped structure, consisting of an encoder, a decoder, and skip connections across layers. The encoder is used to downsample layer by layer and extract multi-scale semantic features of the image, while the decoder gradually restores spatial resolution through upsampling to achieve fine reconstruction of the target region. Skip connections are set between corresponding layers of the encoder and decoder to fuse shallow spatial details with deep semantic information to enhance feature representation capabilities. The number of output channels in each layer increases progressively, and max pooling and bilinear interpolation are used to implement downsampling and upsampling operations, respectively, thereby maintaining the computational efficiency and structural integrity of the overall network.

[0017] S2-b) Encoder Design: The encoder extracts features from the input image layer by layer and performs downsampling processing. It contains multiple coding units of different depths. The shallow coding units integrate an Adaptive Branch Hybrid Shape Convolutional Module (ABMSC), which enhances the modeling ability of local details such as edges and textures through a variable-direction multi-branch structure. The deep coding units introduce a Residual Directional Feature Interaction Module (RDFI), which enhances the model's perception of horizontal and vertical spatial dependencies through orthogonal feature interaction structures, thereby improving the segmentation accuracy of complex structural regions. The spatial dimension is progressively compressed through downsampling operations between layers, enhancing the semantic level of feature representation and laying the foundation for subsequent decoding and reconstruction.

[0018] S2-c) Decoder Design: The decoder adopts a symmetrical structure to the encoder, responsible for progressively restoring the spatial resolution of the feature maps and generating segmentation results. Specifically, deeper features in the decoder first undergo orientation awareness and feature correction through the Residual Directional Feature Interaction (RDFI) module, followed by upsampling and further enhancement of feature representation through the RDFI module. Next, shallow features are processed through three consecutive upsampling steps combined with the Adaptive Branch Hybrid Shape Convolution (ABMSC) module, further improving the ability to capture local details and edge information. Each stage of the decoder also fuses with the corresponding encoder layer through skip connections, achieving an effective combination of shallow spatial details and deep semantic information, enhancing the accuracy and robustness of segmentation.

[0019] S2-d) Skip Connection Design: Skip connections are set between corresponding layers of the encoder and decoder to pass multi-scale features extracted by the encoder, thereby enhancing the decoder's ability to recover spatial details. Skip connections employ a feature concatenation method, fusing shallow features from the encoder with features from the current layer of the decoder. This design not only preserves rich edge and texture information from the shallow layers but also incorporates deep semantic features, improving the model's accuracy in recognizing and segmenting complex lesion structures. Through skip connections, the model can better balance local details with global context, achieving high-quality medical image segmentation.

[0020] S2-e) Output Layer Design: The feature map of the last layer of the decoder is compressed by 1×1 convolution to generate a predicted probability map corresponding to the target segmentation category. Subsequently, a thresholding operation is used to convert the probability map into a binary segmentation result to clearly distinguish the lesion region from the background. This design ensures the accuracy and interpretability of the segmentation results and is suitable for subsequent analysis and diagnosis of various medical image segmentation tasks.

[0021] S3. Model Training and Segmentation Inference

[0022] To achieve optimal performance of the medical image segmentation network proposed in this embodiment, the constructed network structure is trained under supervision using preprocessed medical image data. The model training phase is based on a standard segmentation dataset, employing a combination of cross-entropy loss and Dice loss as optimization objectives, with parameters updated using the Adam optimizer. The training process is set to 300 epochs to ensure that the network achieves gradient stability and continuous enhancement of feature extraction capabilities through sufficient iterations. After training, the model parameters are fixed and applied to medical image segmentation tasks on an independent test set. During the testing phase, no additional manual intervention is required for the input images. The model can automatically extract key anatomical regions or lesion regions from the images and output the corresponding segmentation results. It has good boundary preservation ability and tissue recognition accuracy, which can effectively support subsequent clinical analysis and assisted diagnostic applications.

[0023] Furthermore, the specific structure of the Adaptive Branch Hybrid Shape Convolutional Module (ABMSC) in the network described in step S2 is as follows:

[0024] a) Construction of multi-branch convolutional structure: The input feature map is divided into multiple sub-branches along the channel dimension. Different branches use depth-separable convolutional kernels of different shapes for feature extraction. The convolutional kernels include, but are not limited to: standard square convolution (e.g., 3×3), wide strip convolution (e.g., 1×k), and tall strip convolution (e.g., k×1), which are used to capture local structural information of the image in multiple directions and scales.

[0025] b) Multi-scale channel attention introduction: The output of each convolutional branch is connected to the multi-scale channel attention module (MSSE). This module performs global average pooling, max pooling and standard deviation pooling on the branch outputs respectively, and extracts the global response features of the channel dimension through a lightweight perceptron to generate weights to adaptively emphasize key semantic channels and suppress redundant features.

[0026] c) Feature concatenation and fusion: Combining the output and input features of all convolutional branches. Figure 1 Along the same channel dimension, feature compression and information fusion are achieved through 1×1 convolutional layers, thereby effectively integrating perceptual information in various scale directions;

[0027] d) Enhanced expressive power and suppression of interference: Through the above-mentioned hybrid convolutional branch structure and multi-scale attention mechanism, the ABMSC module achieves efficient fusion of multi-directional and multi-scale features, enhances the network's ability to perceive the texture structure of the target region, and suppresses background noise interference, thereby improving the edge localization and small target detection effect in medical image segmentation.

[0028] Furthermore, the specific structure of the Residual Directional Feature Interaction Module (RDFI) in the network described in step S2 is as follows:

[0029] a) Orthogonal feature shift: The input feature map is divided into several subgroups according to the channel dimension, and the feature map is shifted in the horizontal (width) and vertical (height) directions respectively to form two orthogonal feature branches, which are used to simulate the spatial information interaction in different directions, thereby enhancing the local context modeling capability without introducing additional convolution parameters.

[0030] b) Channel Attention Fusion (CAF): The features of the two directional branches are concatenated along the channel dimension. The global channel representation is extracted by statistical methods such as average pooling, max pooling and standard deviation pooling, and then input into a multilayer perceptron to generate channel attention weights, which are used to guide the weighted fusion of the features of the two directions, effectively strengthening the channel response of semantically salient regions.

[0031] c) Spatial Attention Fusion (SAF): The fusion features output by the CAF module are spliced ​​through channels again to introduce a spatial attention mechanism. A spatial weight map is generated through a lightweight perceptron. The spatial information of the two branches is weighted and fused according to the cross-residual connection method to achieve fine-grained weighting of features at different spatial locations and improve the network’s sensitivity to key structural regions (such as edges, lesions, etc.).

[0032] d) Controlled Fusion (DGF): The features processed by spatial attention are re-fused. A lightweight gated network is constructed by global average pooling and double 1×1 convolution. The importance of the features in the two branches is adaptively learned to achieve dynamic selection and fusion of multi-directional semantics.

[0033] e) Spatial Refinement Enhancement (SAR): The features are then concatenated by max pooling and average pooling, connected to a 7×7 convolution and activated by Sigmoid, to generate a spatial attention map to further refine the representation of salient regions and enhance the segmentation ability of complex structures such as blurred edges and small targets.

[0034] f) Residual connections enhance stability: The final output features are added element-wise to the initial input through residuals, which preserves the original low-level features and alleviates the gradient vanishing problem in deep networks, thereby improving the model's training stability and generalization ability. Attached Figure Description

[0035] Figure 1 A schematic diagram of a medical image segmentation method with adaptive receptive field and feature correction provided in this embodiment of the invention.

[0036] Figure 2 A schematic diagram of the ABMSC structure provided in this embodiment of the invention.

[0037] Figure 3 RDFI structure diagram provided in this embodiment of the invention

[0038] Figure 4 Skin lesion image segmentation results provided in this embodiment of the invention

[0039] Figure 5 Breast ultrasound image segmentation results provided in this embodiment of the invention Detailed Implementation

[0040] To more clearly illustrate the technical solution, concept, and expected effects of the present invention, the present invention will be described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that the described embodiments are only used to explain the principles and methods of the present invention, and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various adjustments, substitutions, or combinations can be made to the technical features without departing from the spirit and essence of the present invention, and these should all be considered as additional embodiments of the present invention. The accompanying drawings are only illustrative in nature, used to assist in understanding the structure and process of the present invention, and are not necessarily drawn to scale. Unless otherwise specified, the technical features in the following embodiments can also be reasonably combined and used in other embodiments.

[0041] The present invention will now be described in further detail with reference to the accompanying drawings.

[0042] Example 1

[0043] A medical image segmentation method with adaptive receptive field and feature correction includes the following steps:

[0044] S1. Data Preprocessing

[0045] In the medical image segmentation method of this invention, the original medical images are first subjected to systematic data preprocessing operations to improve the model's adaptability to medical images from different sources and of different qualities, and to enhance the stability and generalization performance of the segmentation results during training. This preprocessing process includes, but is not limited to, the following steps:

[0046] S1-a) Data Acquisition: Obtain medical image samples with clearly labeled lesion areas from clinical medical databases or publicly available medical image datasets. These images include, but are not limited to, dermoscopic images (for skin lesion identification) and breast ultrasound images (for breast mass analysis). All images are calibrated with equipment parameters before acquisition to ensure consistent spatial resolution. Labeling information is reviewed by professional radiologists to ensure label accuracy and clinical reliability.

[0047] S1-b) Random Flip: Apply random horizontal or vertical flipping to the input image to simulate changes in patient position in actual clinical imaging, improve the network's robustness to directional changes, and mitigate the risk of overfitting due to insufficient sample size.

[0048] S1-c) Random rotation: Perform random rotation processing on the image in 90-degree units (including 0°, 90°, 180° and 270°) to expand the orientation diversity of the image and further improve the model's ability to adapt to orientation uncertainty in clinical images.

[0049] S1-d) Resizing and Normalization: All images are uniformly cropped or bilinearly interpolated to a specified input size (e.g., 256×256 pixels) to meet the network structure input requirements and maintain scale consistency. Simultaneously, pixel values ​​are linearly normalized to map them to the standard numerical range of [0,1] or [-1,1], thereby reducing the impact of different imaging devices, exposure conditions, or lighting differences on image grayscale distribution and improving the consistency and stability of feature extraction.

[0050] S2. Segmentation Network Architecture Design and Feature Extraction Strategy

[0051] In the medical image segmentation method described in this invention, an improved U-shaped network structure is constructed, such as... Figure 1 As shown, this structure includes an encoder, a decoder, and multi-level skip connection paths to achieve multi-scale feature extraction and spatial detail restoration, specifically including the following steps:

[0052] S2-a) Overall Architecture Design: In this invention, the proposed medical image segmentation network adopts a symmetrical U-shaped structure, consisting of an encoder, a decoder, and skip connections across layers. The encoder is used to downsample layer by layer and extract multi-scale semantic features of the image, while the decoder gradually restores spatial resolution through upsampling to achieve fine reconstruction of the target region. Skip connections are set between corresponding layers of the encoder and decoder to fuse shallow spatial details with deep semantic information to enhance feature representation capabilities. The number of output channels in each layer increases progressively, and max pooling and bilinear interpolation are used to implement downsampling and upsampling operations, respectively, thereby maintaining the computational efficiency and structural integrity of the overall network.

[0053] S2-b) Encoder Design: The encoder extracts features from the input image layer by layer and performs downsampling processing. It contains multiple coding units of different depths. The shallow coding units integrate an Adaptive Branch Hybrid Shape Convolutional Module (ABMSC), which enhances the modeling ability of local details such as edges and textures through a variable-direction multi-branch structure. The deep coding units introduce a Residual Directional Feature Interaction Module (RDFI), which enhances the model's perception of horizontal and vertical spatial dependencies through orthogonal feature interaction structures, thereby improving the segmentation accuracy of complex structural regions. The spatial dimension is progressively compressed through downsampling operations between layers, enhancing the semantic level of feature representation and laying the foundation for subsequent decoding and reconstruction.

[0054] S2-c) Decoder Design: The decoder adopts a symmetrical structure to the encoder, responsible for progressively restoring the spatial resolution of the feature maps and generating segmentation results. Specifically, deeper features in the decoder first undergo orientation awareness and feature correction through the Residual Directional Feature Interaction (RDFI) module, followed by upsampling and further enhancement of feature representation through the RDFI module. Next, shallow features are processed through three consecutive upsampling steps combined with the Adaptive Branch Hybrid Shape Convolution (ABMSC) module, further improving the ability to capture local details and edge information. Each stage of the decoder also fuses with the corresponding encoder layer through skip connections, achieving an effective combination of shallow spatial details and deep semantic information, enhancing the accuracy and robustness of segmentation.

[0055] S2-d) Skip Connection Design: Skip connections are set between corresponding layers of the encoder and decoder to pass multi-scale features extracted by the encoder, thereby enhancing the decoder's ability to recover spatial details. Skip connections employ a feature concatenation method, fusing shallow features from the encoder with features from the current layer of the decoder. This design not only preserves rich edge and texture information from the shallow layers but also incorporates deep semantic features, improving the model's accuracy in recognizing and segmenting complex lesion structures. Through skip connections, the model can better balance local details with global context, achieving high-quality medical image segmentation.

[0056] S2-e) Output Layer Design: The feature map of the last layer of the decoder is compressed by 1×1 convolution to generate a predicted probability map corresponding to the target segmentation category. Subsequently, a thresholding operation is used to convert the probability map into a binary segmentation result to clearly distinguish the lesion region from the background. This design ensures the accuracy and interpretability of the segmentation results and is suitable for subsequent analysis and diagnosis of various medical image segmentation tasks.

[0057] S3. Model Training and Segmentation Inference

[0058] To achieve optimal performance of the proposed medical image segmentation network, supervised training of the constructed network structure is performed using preprocessed medical image data. The model training phase is based on a standard segmentation dataset, employing a combination of cross-entropy and Dice loss functions as optimization objectives, and updating parameters using the Adam optimizer. The training process is set to 300 epochs to ensure that the network achieves gradient stability and continuous enhancement of feature extraction capabilities through sufficient iterations.

[0059] After training, the model parameters are fixed and applied to medical image segmentation tasks on an independent test set. During the testing phase, no additional manual intervention is required for the input images. The model can automatically extract key anatomical regions or lesion regions from the images and output the corresponding segmentation results. It has good boundary preservation ability and tissue recognition accuracy, which can effectively support subsequent clinical analysis and assisted diagnostic applications.

[0060] Furthermore, such as Figure 2 As shown, the specific structure of the Adaptive Branch Hybrid Shape Convolutional Module (ABMSC) in step S2 within the network is as follows:

[0061] a) Construction of multi-branch convolutional structure: The input feature map is divided into multiple sub-branches along the channel dimension. Different branches use depth-separable convolutional kernels of different shapes for feature extraction. The convolutional kernels include, but are not limited to: standard square convolution (e.g., 3×3), wide strip convolution (e.g., 1×k), and tall strip convolution (e.g., k×1), which are used to capture local structural information of the image in multiple directions and scales.

[0062] b) Multi-scale channel attention introduction: The output of each convolutional branch is connected to the multi-scale channel attention module (MSSE). This module performs global average pooling, max pooling and standard deviation pooling on the branch outputs respectively, and extracts the global response features of the channel dimension through a lightweight perceptron to generate weights to adaptively emphasize key semantic channels and suppress redundant features.

[0063] c) Feature concatenation and fusion: Combining the output and input features of all convolutional branches. Figure 1 Along the same channel dimension, feature compression and information fusion are achieved through 1×1 convolutional layers, thereby effectively integrating perceptual information in various scale directions;

[0064] d) Enhanced expressive power and suppression of interference: Through the above-mentioned hybrid convolutional branch structure and multi-scale attention mechanism, the ABMSC module achieves efficient fusion of multi-directional and multi-scale features, enhances the network's ability to perceive the texture structure of the target region, and suppresses background noise interference, thereby improving the edge localization and small target detection effect in medical image segmentation.

[0065] Furthermore, such as Figure 3 As shown, the specific structure of the Residual Directional Feature Interaction (RDFI) module in the network described in step S2 is as follows:

[0066] a) Orthogonal feature shift: The input feature map is divided into several subgroups according to the channel dimension, and the feature map is shifted in the horizontal (width) and vertical (height) directions respectively to form two orthogonal feature branches, which are used to simulate the spatial information interaction in different directions, thereby enhancing the local context modeling capability without introducing additional convolution parameters.

[0067] b) Channel Attention Fusion (CAF): The features of the two directional branches are concatenated along the channel dimension. The global channel representation is extracted by statistical methods such as average pooling, max pooling and standard deviation pooling, and then input into a multilayer perceptron to generate channel attention weights, which are used to guide the weighted fusion of the features of the two directions, effectively strengthening the channel response of semantically salient regions.

[0068] c) Spatial Attention Fusion (SAF): The fusion features output by the CAF module are spliced ​​through channels again to introduce a spatial attention mechanism. A spatial weight map is generated through a lightweight perceptron. The spatial information of the two branches is weighted and fused according to the cross-residual connection method to achieve fine-grained weighting of features at different spatial locations and improve the network’s sensitivity to key structural regions (such as edges, lesions, etc.).

[0069] d) Dynamic Gated Fusion (DGF): Features processed by spatial attention are re-fused. Global average pooling and double 1×1 convolution are introduced to construct a lightweight gated network. The importance of features in the two branches is adaptively learned to achieve dynamic selection and fusion of multi-directional semantics.

[0070] e) Spatial Refinement Enhancement (SAR): By performing max pooling and average pooling on the fused features, connecting them to a 7×7 convolution and activating them with Sigmoid, a spatial attention map is generated to further refine the representation of salient regions and enhance the segmentation ability of complex structures such as blurred edges and small targets.

[0071] f) Residual connections enhance stability: The final output features are added element-wise to the initial input through residuals, which preserves the original low-level features and alleviates the gradient vanishing problem in deep networks, thereby improving the model's training stability and generalization ability.

[0072] Example 2

[0073] To verify the actual segmentation performance of the medical image segmentation method with adaptive receptive field and feature correction proposed in this embodiment, the method described in Example 1 was used, and the ISIC2017 publicly available skin lesion segmentation dataset was selected as the experimental dataset. Performance evaluation experiments were conducted on the skin lesion region segmentation task. The specific steps are as follows:

[0074] S1. Experimental Environment and Data Preparation

[0075] In the experimental phase, the ISIC2017 skin lesion dataset was selected as the basic testing platform. It contains dermoscopic images collected in real clinical scenarios and their corresponding lesion area masks. The segmentation task was binary segmentation of lesion areas and normal skin areas in the skin images. The dataset was randomly divided into training and testing sets, with a ratio of 80% and 20%, respectively. The experiments were conducted using PyTorch on an NVIDIA RTX 3060Ti GPU with 8GB of VRAM.

[0076] S2. Data Preprocessing Strategy

[0077] All input images were uniformly resized to a fixed size of 256×256 pixels and underwent a series of enhancement processes, including but not limited to random 90° rotation and horizontal flipping, to enhance the model's generalization ability and robustness. Simultaneously, image pixel values ​​were normalized to ensure a uniform distribution and reduce interference from different acquisition devices.

[0078] S3. Model Training Process

[0079] The model was optimized using the AdamW optimization algorithm, with an initial learning rate of 0.01 and a batch size of 8. Training lasted for 300 epochs, during which key metrics such as the Dice coefficient, IoU exponent, and cross-entropy loss were monitored in real time to adjust the training strategy. After training, predictions were made on the test set images, and threshold segmentation was performed based on the output probability map to generate the final result.

[0080] S4. Experimental Results and Performance Evaluation

[0081] The proposed network architecture was comprehensively evaluated on the ISIC 2017 skin lesion segmentation dataset. Experimental results show that the proposed method achieves excellent performance on multiple evaluation metrics. Specifically, the Intersection over Union (IoU) reaches 84.49% ± 0.13, and the Dice coefficient reaches 91.40% ± 0.09, placing it at the forefront of the compared methods and fully validating the accuracy and stability of the model in segmenting complex skin lesion regions.

[0082] Furthermore, this model maintains high accuracy while having reasonable computational resource requirements and good inference efficiency. The network contains only 8.07M parameters, with a computational complexity of 0.83 GFLOPs. At a 256×256 resolution input, it can complete forward inference at a relatively fast speed, making it suitable for medical image processing scenarios with certain requirements for real-time performance and deployment friendliness.

[0083] S5. Visualization Analysis

[0084] like Figure 5 As shown, the segmentation map generated by the model exhibits good structural preservation capabilities in terms of boundary contours and lesion area integrity, and can accurately capture the spatial location of skin lesion areas, effectively supporting clinical auxiliary diagnosis.

[0085] Example 3

[0086] To verify the actual segmentation performance of the medical image segmentation method with adaptive receptive field and feature correction proposed in this embodiment, the method described in Example 1 was used, and the BUSI2018 publicly available skin lesion segmentation dataset was selected as the experimental dataset. Performance evaluation experiments were conducted on the skin lesion region segmentation task. The specific steps are as follows:

[0087] S1. Experimental Environment and Data Preparation

[0088] In the experimental phase, the BUSI2018 breast ultrasound image dataset was selected as the basic testing platform. This dataset contains breast ultrasound images acquired in real clinical settings and their corresponding tumor region annotation masks. The segmentation task was to accurately separate the tumor lesion region from the normal breast tissue region in the ultrasound image, which is a typical binary segmentation problem. The dataset was randomly divided into training and testing sets, with a ratio of 80% and 20%, respectively. The experiments were conducted using PyTorch on an NVIDIA RTX 3060Ti GPU with 8GB of video memory.

[0089] S2. Data Preprocessing Strategy

[0090] All input images were uniformly resized to a fixed size of 256×256 pixels and underwent a series of enhancement processes, including but not limited to random 90° rotation and horizontal flipping, to enhance the model's generalization ability and robustness. Simultaneously, image pixel values ​​were normalized to ensure a uniform distribution and reduce interference from different acquisition devices.

[0091] S3. Model Training Process

[0092] The model was optimized using the AdamW optimization algorithm, with an initial learning rate of 0.01 and a batch size of 8. Training lasted for 300 epochs, during which key metrics such as the Dice coefficient, IoU exponent, and cross-entropy loss were monitored in real time to adjust the training strategy. After training, predictions were made on the test set images, and threshold segmentation was performed based on the output probability map to generate the final result.

[0093] S4. Experimental Results and Performance Evaluation

[0094] The proposed network architecture was comprehensively evaluated on the BUSI2018 skin lesion segmentation dataset. Experimental results show that the proposed method achieves excellent performance on multiple evaluation metrics. Specifically, the average Intersection over Union (IoU) reaches 65.55% ± 0.22, and the average Dice coefficient reaches 78.09% ± 0.32, placing it at the forefront among the compared methods and fully validating the accuracy and stability of the model in segmenting complex skin lesion regions.

[0095] Furthermore, this model maintains high accuracy while having reasonable computational resource requirements and good inference efficiency. The network contains only 8.07M parameters, with a computational complexity of 0.83 GFLOPs. At a 256×256 resolution input, it can complete forward inference at a relatively fast speed, making it suitable for medical image processing scenarios with certain requirements for real-time performance and deployment friendliness.

[0096] S5. Visualization Analysis

[0097] like Figure 5 As shown, the segmentation results generated by the model perform excellently in terms of tumor boundary contour restoration and lesion area structural integrity. It can accurately locate the tumor area in breast ultrasound images, effectively alleviate background noise and artifact interference, and demonstrate good spatial resolution and detail preservation capabilities, providing strong support for the clinical auxiliary diagnosis of breast tumors.

[0098] It should be noted that the above embodiments are only used to illustrate the principles and implementation of the present invention, and are intended to help those skilled in the art to better understand the core ideas and technical solutions of the present invention. It should be understood that, without departing from the basic concept of the present invention, various adjustments and improvements made to its structural layout, module combination or algorithm flow should be considered as equivalent variations of the present invention and should also be included within the protection scope of the present invention.

[0099] The network architecture and key modules proposed in this invention can be flexibly expanded and optimized according to different medical image types or application scenarios. Any technical solution that achieves the same function through equivalent technical means under the guidance of this invention shall be protected by the claims of this invention.

[0100] Therefore, the contents disclosed in this specification should not be construed as limiting the scope of protection of this invention, and the final scope of protection shall be determined by the contents of the claims.

Claims

1. A medical image segmentation method with adaptive receptive field and feature correction, characterized in that, Comprising the following steps: S1. Obtain raw medical image data and pre-process to obtain a training data set, the pre-processing comprising: S1-a) Data acquisition: Obtain medical images and their corresponding segmentation annotations from a clinical medical database or a public medical image dataset, the images including but not limited to dermoscopy images and breast ultrasound images, the images being collected via standardization to ensure pixel resolution and annotation accuracy; S1-b) Random flip: Randomly perform horizontal or vertical flip operations on the original images to enhance the robustness of the model to different direction inputs; S1-c) Random rotation: Apply a random rotation of 90 degrees to the images to simulate the diversity of clinical shooting directions; S1-d) Size adjustment and normalization: Crop or interpolate the images to a uniform size to adapt to the network input requirements, the size being 256x256, and normalize the image pixels to map the pixel values to the interval [0, 1] or [-1, 1], reducing data bias caused by different imaging devices; S2. Segmentation network architecture design and feature extraction strategy: In the medical image segmentation method described in the present application, an improved U-shaped network structure is constructed, which includes an encoder, a decoder and a multi-level jump connection path, for realizing multi-scale feature extraction and spatial detail recovery, comprising the following steps: S2-a) Overall architecture design: In the present application, the proposed medical image segmentation network adopts a symmetrical U-shaped structure, which is composed of an encoder, a decoder and a cross-layer jump connection, the encoder is used for layer-by-layer downsampling and extracting multi-scale semantic features of the image, the decoder is used for gradually recovering the spatial resolution through upsampling to realize fine reconstruction of the target region, the corresponding layers of the encoder and the decoder are connected by jump connection to fuse the spatial details of the shallow layer and the semantic information of the deep layer to enhance the feature expression ability, the output channel number of each layer is increased layer by layer, and maximum pooling and bilinear interpolation are used to realize downsampling and upsampling operations respectively, so as to maintain the computational efficiency and structural integrity of the overall network; S2-b) Encoder design: The encoder is used to extract features from the input image layer by layer and perform downsampling processing, including multiple encoding units of different depths, the shallow encoding unit integrates an adaptive branch mixed shape convolution module (ABMSC), which improves the modeling ability of local details such as edges and textures through a multi-branch structure with variable direction, the deep encoding unit introduces a residual directional feature interaction module (RDFI) to enhance the perception ability of the model to horizontal and vertical spatial dependence through an orthogonal directional feature interaction structure, thereby improving the segmentation accuracy of complex structure regions, the spatial dimensions are gradually compressed between layers through downsampling operations to enhance the semantic level of feature expression, laying a foundation for subsequent decoding and reconstruction; S2-c) Decoder design: The decoder adopts a symmetric structure design with the encoder, responsible for gradually restoring the spatial resolution of the feature map and generating the segmentation result. Specifically, the deeper features in the decoder first pass through the residual directional feature interaction module (RDFI) for direction perception and feature correction, then are upsampled and pass through the RDFI module again to strengthen feature expression. Subsequently, the shallow features are processed by the adaptive branch mixed shape convolution module (ABMSC) through continuous three times of upsampling, further improving the ability to capture local details and edge information. The stages of the decoder are also fused with the corresponding encoder layers through jump connections, effectively combining shallow spatial details and deep semantic information to enhance the accuracy and robustness of segmentation. S2-d) Jump connection design: A jump connection is set between the corresponding layers of the encoder and the decoder to transfer the multi-scale features extracted by the encoder, enhancing the decoder's ability to recover spatial details. The jump connection uses feature concatenation to fuse the shallow features of the encoder with the features of the current layer of the decoder. This design not only preserves the rich edge and texture information of the shallow layer but also combines deep semantic features, improving the model's recognition accuracy and segmentation effect for complex lesion structures. Through the jump connection, the model can better balance local details and global context, achieving high-quality medical image segmentation. S2-e) Output layer design: The feature map of the last layer of the decoder is compressed in channel number through 1x1 convolution to generate a prediction probability map corresponding to the target segmentation class. Then, a thresholding operation is used to convert the probability map into a binary segmentation result to clearly distinguish the lesion area from the background. This design ensures the accuracy and interpretability of the segmentation result, making it suitable for subsequent analysis and diagnosis of various medical image segmentation tasks.

2. A method for medical image segmentation as claimed in claim 1, characterized in that, The adaptive branch mixed shape convolution module (ABMSC) in step S2 has the following structure: 2-a) The input feature map is divided into multiple branches along the channel dimension, and each branch uses different shape depth separable convolution operations, including standard square convolution, wide strip convolution, and high strip convolution, to extract multi-directional and multi-scale spatial features. 2-b) Each convolution branch is followed by a multi-scale channel attention module that generates channel attention weights through a perceptron by combining global average pooling and max pooling to adaptively adjust the importance of each channel. 2-c) The outputs of multiple convolution branches are concatenated with the original input features, and then fused through 1x1 convolution to adjust the feature dimension and integrate information. 2-d) This module effectively enhances the network's ability to perceive details, textures, and structural edges in medical images through multi-branch mixed shape convolution and channel attention fusion.

3. The network architecture for medical image segmentation of claim 1, wherein, The residual directional feature interaction module (RDFI) in step S2 has the following structure and functions: 3-a) Orthogonal Shift Branches: The input feature map is divided into multiple parts along the channel dimension, and the part of the channel feature is shifted along the high (vertical) direction and the wide (horizontal direction) respectively, and then two branch feature maps are obtained by splicing, realizing the cross-channel axis neighborhood information interaction without introducing additional parameters; 3-b) ChannelAttention Fusion (CAF): After the channel splicing of the two orthogonal branch feature maps, the channel global information is extracted by using multiple statistical pooling (average, maximum and standard deviation pooling), and the channel attention weight is generated by inputting the multilayer perceptron, and the weighted fusion and correction of the features in two directions are realized by the cross residual mechanism; 3-c) Spatial Attention Fusion (SAF): After the channel splicing of the two features output by CAF, the spatial attention weight is calculated using a lightweight multilayer perceptron, and the spatial information of the two branches is weighted and fused according to the cross residual connection method to improve the expression of important regions at the spatial level; 3-d) Dynamic Gate Fusion (DGF): After the spatial attention fusion of the two branch features, the two branch features are spliced by channel, and a lightweight gate network with global average pooling and double 1×1 convolution is used to dynamically learn the importance weight of the two branch features, realizing adaptive weighted fusion; 3-e) Spatial Attention Refinement (SAR): After the channel dimension maximum pooling and average pooling of the features fused by the dynamic gate fusion, the spatial attention weight map is generated by 7×7 convolution and Sigmoid activation, and the spatial weighted refinement of the fused features is performed to enhance the expression ability of boundary details and fuzzy areas; 3-f) Residual Connection: The output feature of the spatial attention refinement module is added to the input feature by using the residual connection structure, which relieves the gradient vanishing problem and improves the model training stability and generalization ability.

Images