Lightweight network architecture for medical image segmentation

Abstract

The invention discloses a lightweight network architecture for medical image segmentation, and the specific implementation comprises the steps: (1) obtaining original medical image data and a segmentation label of a target region, and constructing a training data set suitable for a segmentation task; (2) a lightweight neural network architecture is designed, the segmentation performance is improved through multi-scale feature interaction and cross-layer information fusion, and meanwhile features are effectively extracted and key areas are enhanced; (3) introducing a dynamic feature enhancement mechanism, and adaptively optimizing detail information and semantic expression according to different segmentation requirements; and (4) performing efficient segmentation of a target area on a test image through a model trained by matching a preprocessing technology and a post-processing technology. The invention provides an end-to-end medical image segmentation scheme, which has the characteristics of high efficiency, accuracy, strong robustness and support of real-time processing, is suitable for various medical image segmentation scenes, and provides powerful support for disease auxiliary diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and medical image processing, and in particular to a lightweight network architecture for medical image segmentation. The architecture aims to improve the computational efficiency and accuracy of medical image segmentation tasks by optimizing deep learning models, while reducing computing resource consumption and adapting to application requirements in edge devices or low-power environments. The technology can be widely used in the fields of medical image analysis, disease diagnosis, auxiliary treatment planning, etc., and has important application value, especially for resource-constrained devices such as mobile medical devices and embedded medical hardware. Background Art

[0002] With the rapid development of medical imaging technology, medical image segmentation, as a key task in medical image analysis, has been widely used in lesion area detection, disease diagnosis and treatment planning. Medical image segmentation can automatically identify and mark specific areas in images, helping doctors improve the accuracy of diagnosis, develop personalized treatment plans, and plan surgeries. In these applications, deep learning, especially convolutional neural networks (CNNs), has become an important technical means to solve the problem of medical image segmentation.

[0003] In recent years, deep learning models such as U-Net and its variants, DeepLab series, etc. have achieved remarkable results in the field of medical image segmentation. U-Net is widely used, especially because of its encoder-decoder structure and jump connection, which can effectively extract and restore image features and improve segmentation accuracy. Other advanced models such as DeepLab also show good performance in multi-scale feature extraction and region segmentation, and show strong capabilities in processing complex images. However, although these models have achieved good results in terms of accuracy and robustness, they still face some challenges in actual clinical applications.

[0004] Existing deep learning models are often complex in structure and contain a large number of parameters and computing operations, which makes them have high requirements for computing resources and storage space. This feature limits the application of models on resource-limited devices (such as mobile medical devices, embedded medical hardware, etc.), especially in some scenarios that require low power consumption and high efficiency. In addition, many existing networks have slow inference speeds, which is obviously not ideal in clinical environments with high real-time requirements. For applications such as emergency and surgery that require immediate feedback, model delays will affect the decision-making efficiency of doctors and the timeliness of diagnosis and treatment for patients.

[0005] Medical image data is usually highly complex and diverse. Existing models are prone to overfitting during training, especially when the amount of data is limited, which further affects the generalization ability and reliability of the model in practical applications. In order to maintain good segmentation accuracy, existing deep learning methods often contain complex network structures and redundant calculation steps, which not only consumes a lot of computing resources, but also may lead to unsatisfactory performance when processing fine-grained features, especially when processing medical images with blurred edges or more noise, which may lead to inaccurate segmentation results.

[0006] In addition, the deployment and maintenance of existing models often require professional technicians and high hardware resources, which is particularly difficult when they need to be applied on different platforms and devices. This problem limits the feasibility of large-scale promotion and popularization of deep learning models in clinical settings.

[0007] In order to meet the above challenges, the field of medical image segmentation urgently needs a new network architecture that can significantly reduce the consumption of computing resources while ensuring segmentation accuracy, improve reasoning speed, and have good generalization capabilities. This architecture should also be able to run efficiently in various devices and environments, promoting the widespread use of medical image analysis technology in a variety of application scenarios, especially in edge devices and low-power medical devices. Summary of the invention

[0008] The purpose of this invention is to provide a lightweight network architecture for medical image segmentation, aiming to solve the technical problems existing in the prior art, such as large consumption of computing resources, slow inference speed, overfitting problems, and difficulty in implementation on low-power devices. Specific objectives include:

[0009] 1. Reduce computing resource consumption: By adopting a lightweight network structure and optimizing module design, the computational complexity and parameter quantity of the network model are reduced, enabling it to run efficiently on devices with limited computing resources.

[0010] 2. Improve the reasoning speed: By reducing redundant calculations and optimizing the network structure, the reasoning speed of the model is improved to meet the real-time requirements of medical image segmentation applications, especially the need for fast feedback in clinical diagnosis and surgery.

[0011] 3. Improve segmentation accuracy: While ensuring the network is lightweight, retain sufficient segmentation accuracy, especially when processing complex medical images (such as malignant melanoma, breast cancer, etc.), ensuring accurate segmentation of key areas.

[0012] 4. Enhance the generalization ability of the model: Through innovative design solutions, reduce the risk of overfitting of the model during training, and improve the adaptability and generalization ability of the model on different types of medical images and devices.

[0013] 5. Adapt to low-power devices and edge device applications: The network architecture can be efficiently deployed on mobile medical devices, embedded hardware, and edge computing platforms, thereby promoting the application of medical image segmentation technology in more resource-constrained environments.

[0014] The purpose of the present invention is to achieve the above purpose, and the technical solution adopted by the present invention is:

[0015] A lightweight network architecture for medical image segmentation, comprising the following steps:

[0016] S1. Data Preprocessing

[0017] In the lightweight network architecture for medical image segmentation of the present invention, firstly, a comprehensive data preprocessing is performed on the input medical image to ensure the quality of the input data and improve the robustness and generalization ability of the model. This step includes the following specific operations:

[0018] S1-a) Data collection: Various types of medical image data, including but not limited to dermoscopy images (malignant melanoma) and breast ultrasound images, are obtained from clinical medical databases or public medical image datasets. All data are collected in a standardized manner to ensure the accuracy of pixel resolution and annotation.

[0019] S1-b) Random flipping: Randomly flip the input image horizontally or vertically to simulate the effects of shooting at different angles. This enhancement operation not only expands the training data, but also improves the model's ability to recognize left-right symmetrical and top-bottom symmetrical lesions, which is particularly suitable for symmetrical structures commonly seen in skin lesions and breast images.

[0020] S1-c) Random rotation: A 90-degree rotation operation is randomly applied to the input image. This strategy helps the model to achieve robustness in direction and avoid the degradation of segmentation performance caused by changes in the orientation of the imaging device.

[0021] S1-d) Random change of HSI channels: Randomly perturb the hue, saturation and brightness channels of the image.

[0022] S1-e) Cropping and normalization: The input image is cropped or interpolated to the fixed size required by the network (such as 256×256, 512×512) to ensure the consistency of the input data. The pixel values ​​are then normalized and mapped to the range of [0,1] or [-1,1] to reduce the impact of grayscale differences caused by different imaging devices on the network.

[0023] S2. Feature extraction:

[0024] In the lightweight network architecture for medical image segmentation of the present invention, a lightweight neural network named JLU-Net (JointLightweight U-shaped Network) is constructed to meet the requirements of medical image segmentation tasks for efficiency and accuracy. By combining multi-scale feature fusion, reconstructed jump connections and attention mechanisms, efficient feature extraction is achieved. This step includes the following specific designs:

[0025] S2-a) Overall architecture design: In the present invention, JLU-Net is a lightweight neural network based on a U-shaped structure, and the overall architecture consists of three parts: an encoder, a decoder, and a reconstructed jump connection path. First, the encoder is responsible for multi-level feature extraction of the input image, gradually capturing deep semantic information, thereby effectively characterizing complex lesion areas in medical images. Secondly, the decoder gradually restores the spatial resolution of the features by upsampling layer by layer, while reconstructing the segmentation target, and finally generates a segmentation probability map. Finally, the reconstructed jump connection path establishes cross-layer feature transfer between the encoder and the decoder.

[0026] S2-b) Encoder design: The encoder consists of multiple encoding units, each of which includes an RLKG block and a JDS module. The RLKG block adopts a lightweight convolution module design, combining a hybrid normalization method and a depth-separable convolution, aiming to efficiently extract features of the input image while reducing computational complexity. The JDS module implements joint downsampling, which can effectively retain the key spatial information of the input image by combining pooling operations with convolution operations.

[0027] S2-c) Skip connection path design: A reconstructed skip connection path is set between the encoder and the decoder, which integrates the SKConv module and the JASM module in the feature transfer process. The SKConv module introduces a dynamic kernel selection mechanism that can be performed according to the semantic requirements of the input features, optimizing the extraction capabilities of local and global features. The JASM module uses the attention shift mechanism to focus on the significant areas in the input features, effectively enhancing the expression capabilities of small objects and edge areas.

[0028] S2-d) Decoder design: The decoder consists of multiple decoding units, each of which includes an upsampling module and an RLKG block. The upsampling module restores the resolution of the feature map layer by layer using bilinear interpolation to reconstruct the same size as the original input image.

[0029] S2-e) Output layer: The final output of the decoder is processed into a segmentation probability map through a 1×1 convolutional layer, and then the probability map is binarized according to the set threshold T (T=0.5) to obtain the final segmentation result.

[0030] S3. Network training and testing:

[0031] In the lightweight network architecture for medical image segmentation of the present invention, the JLU-Net network is trained using a training data set. The training process continues for 300 rounds to ensure that the model effectively converges during multiple rounds of optimization and learns efficient feature representation and segmentation patterns. The fully trained network will be used to test the preprocessed medical images. This test step can automatically detect and obtain the organ or diseased tissue segmentation area in the image, thereby achieving the accurate segmentation task of the medical image.

[0032] Furthermore, the specific structure of the RLKG module in the JLU-Net network is as follows:

[0033] a) Module composition: The RLKG module adopts a layered convolution structure, first performing a 7×7 separable convolution operation and combining it with group normalization to improve the stability of the convolution output and reduce the computational complexity. Subsequently, a 5×5 separable convolution operation is performed and combined with batch normalization to further optimize the standardization and fluency of the features.

[0034] b) Feature fusion: After the convolution operation, the output features are fused with the original input to preserve the low-level features of the input and the high-level features after network processing. This fusion is performed by element-by-element addition and the ReLU activation function is applied to introduce nonlinear characteristics, thereby increasing the expressive power of the network.

[0035] c) Output features: The fused and activated feature maps are fed into subsequent layers of the network for further processing, ensuring that the model can effectively capture and utilize the multi-level features of the input data.

[0036] Furthermore, the specific structure of the JDS module in the JLU-Net network is as follows:

[0037] a) Branch processing: The JDS module receives the input feature map and divides it into two parallel branches: the first branch first performs a normal convolution operation on the input feature map to extract features, and then applies a dilated convolution to capture contextual features of different scales. The dilated convolution enhances the expressiveness of long-distance dependent features by expanding the receptive field. The second branch performs point-by-point convolution on the input feature map to compress the channel dimension, and then further splits it into two sub-paths. One sub-path highlights significant features through a maximum pooling operation, and the other sub-path retains global background information through an average pooling operation. b) Feature fusion: All output feature maps of the first and second branches are integrated through a splicing operation.

[0038] c) Output features.

[0039] Furthermore, the specific structure of the JASM module in the JLU-Net network is as follows:

[0040] a) Input feature map processing: First, the input feature map is reshaped into multiple small image blocks (patches), each of which contains local feature information of a fixed size. This can effectively decompose the input image into small units that are easier to process, laying the foundation for subsequent feature extraction.

[0041] b) Linear Projection: Apply linear transformations to these small image patches to generate three types of feature maps: query, key, and value. The query, key, and value are linearly projected through a learnable weight matrix to generate feature representations suitable for the attention mechanism.

[0042] c) Adaptive input method: In the traditional attention mechanism, there is a direct matching relationship between the query and the key. In the JASM module, this direct correspondence is changed through the adaptive input method, allowing more relative position information to be transmitted between the query, key and value.

[0043] d) Feature synthesis: The processed query, key and value are interacted, the feature weights are calculated, and these features are synthesized together to form the final feature representation. This step fuses this information through convolution operations to obtain an enhanced feature map.

[0044] e) Shift MLP operation: The JASM module then applies the Shift MLP operation, first using depthwise separable convolution to extract features, then using the AugShuffleNet module to shuffle the channels in the feature map to promote cross-channel mixing of information. Next, the GELU activation function is used to perform a nonlinear transformation on the result to further enhance the diversity and nonlinear characteristics of feature expression.

[0045] f) Final output: Finally, the JASM module uses residual connections to merge the input features with the processed output to ensure effective information transfer and stabilize the training process through layer normalization operations.

[0046] Furthermore, the specific structure of the SKConv module in the JLU-Net network is as follows:

[0047] a) Input feature map processing: Given an input feature map, firstly, a preliminary convolution operation is performed on the input feature map to extract basic features and obtain an intermediate feature map.

[0048] b) Channel separation: In SKConv, the channel separation operation first processes the intermediate feature map, and then processes the features of different channels independently through deep separable convolution, reducing the amount of calculation and improving efficiency. Each channel undergoes independent convolution to extract useful feature information.

[0049] c) Detail stripping and enhancement: Detail feature maps are extracted by performing detail stripping operations on the feature maps after channel separation. This process is completed by a specially designed detail stripping module. The stripped detail map can effectively capture local information such as edges and textures in the image. Then, the enhancement operation strengthens these detail information, making the local features more prominent and improving the segmentation ability of small targets.

[0050] d) Feature enhancement: The feature map after detail enhancement is weighted fused with the original input feature map to form the final output feature map.

[0051] e) Output: The final output feature map contains enhanced local features and global information, providing rich feature representation for subsequent network layers. Through this module, JLU-Net can effectively capture features of different scales and levels, thereby enhancing the segmentation effect of lesion areas in medical images. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the JLU-Net network structure of the embodiment of the present invention

[0053] Figure 2 Schematic diagram of encoder structure in an embodiment of the present invention

[0054] FIG. 3 is a schematic diagram of the JASM structure in an embodiment of the present invention.

[0055] Figure 4 Schematic diagram of SKConv structure in an embodiment of the present invention

[0056] Figure 5 Skin lesion image segmentation result obtained by the embodiment of the present invention

[0057] Figure 6 Breast ultrasound segmentation results obtained by the embodiment of the present invention DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0059] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0060] Example 1

[0061] A lightweight network architecture for medical image segmentation, comprising the following steps:

[0062] S1. Data preprocessing

[0063] In the lightweight network architecture for medical image segmentation of this embodiment, firstly, a comprehensive data preprocessing is performed on the input medical image to ensure the quality of the input data and improve the robustness and generalization ability of the model. This step includes the following specific operations:

[0064] S1-a) Data collection: Various types of medical image data, including but not limited to dermoscopy images (malignant melanoma) and breast ultrasound images, are obtained from clinical medical databases or public medical image datasets. All data are collected in a standardized manner to ensure the accuracy of pixel resolution and annotation.

[0065] S1-b) Random flipping: Randomly flip the input image horizontally or vertically to simulate the effects of shooting at different angles. This enhancement operation not only expands the training data, but also improves the model's ability to recognize left-right symmetrical and top-bottom symmetrical lesions, which is particularly suitable for symmetrical structures commonly seen in skin lesions and breast images.

[0066] S1-c) Random rotation: A 90-degree rotation operation is randomly applied to the input image. This strategy helps the model to achieve robustness in direction and avoid the degradation of segmentation performance caused by changes in the orientation of the imaging device.

[0067] S1-d) Random change of HSI channels: Randomly perturb the hue, saturation and brightness channels of the image.

[0068] S1-e) Cropping and normalization: The input image is cropped or interpolated to the fixed size required by the network (such as 256×256, 512×512) to ensure the consistency of the input data. The pixel values ​​are then normalized and mapped to the range of [0,1] or [-1,1] to reduce the impact of grayscale differences caused by different imaging devices on the network.

[0069] S2. Feature extraction:

[0070] In the lightweight network architecture for medical image segmentation of this embodiment, a lightweight neural network named JLU-Net (JointLightweight U-shaped Network) is constructed to meet the requirements of medical image segmentation tasks for efficiency and accuracy. By combining multi-scale feature fusion, reconstructed jump connections and attention mechanisms, efficient feature extraction is achieved. This step includes the following specific designs:

[0071] S2-a) Overall architecture design: In this embodiment, JLU-Net is a lightweight neural network based on a U-shaped structure, and the overall architecture consists of three parts: an encoder, a decoder, and a reconstructed jump connection path. First, the encoder is responsible for multi-level feature extraction of the input image, gradually capturing deep semantic information, thereby effectively characterizing complex lesion areas in medical images. Secondly, the decoder gradually restores the spatial resolution of the features by upsampling layer by layer, while reconstructing the segmentation target, and finally generates a segmentation probability map. Finally, the reconstructed jump connection path establishes cross-layer feature transfer between the encoder and the decoder.

[0072] S2-b) Encoder design: The encoder consists of multiple encoding units, each of which includes an RLKG block and a JDS module. The RLKG block adopts a lightweight convolution module design, combining a hybrid normalization method and a depth-separable convolution, aiming to efficiently extract features of the input image while reducing computational complexity. The JDS module implements joint downsampling, which can effectively retain the key spatial information of the input image by combining pooling operations with convolution operations.

[0073] S2-c) Skip connection path design: A reconstructed skip connection path is set between the encoder and the decoder, which integrates the SKConv module and the JASM module in the feature transfer process. The SKConv module introduces a dynamic kernel selection mechanism that can be performed according to the semantic requirements of the input features, optimizing the extraction capabilities of local and global features. The JASM module uses the attention shift mechanism to focus on the significant areas in the input features, effectively enhancing the expression capabilities of small objects and edge areas.

[0074] S2-d) Decoder design: The decoder consists of multiple decoding units, each of which includes an upsampling module and an RLKG block. The upsampling module restores the resolution of the feature map layer by layer using bilinear interpolation to reconstruct the same size as the original input image.

[0075] S2-e) Output layer: The final output of the decoder is processed into a segmentation probability map through a 1×1 convolutional layer, and then the probability map is binarized according to the set threshold T (T=0.5) to obtain the final segmentation result.

[0076] S3. Network training and testing:

[0077] In the lightweight network architecture for medical image segmentation of this embodiment, the JLU-Net network is trained using a training data set. The training process continues for 300 rounds to ensure that the model effectively converges during multiple rounds of optimization and learns efficient feature representation and segmentation patterns. The fully trained network will be used to test the preprocessed medical images. This test step can automatically detect and obtain the segmented areas of organs or diseased tissues in the image, thereby achieving accurate segmentation tasks for medical images.

[0078] Furthermore, the specific structure of the RLKG module in the JLU-Net network is as follows:

[0079] a) Module composition: The RLKG module adopts a layered convolution structure, first performing a 7×7 separable convolution operation and combining it with group normalization to improve the stability of the convolution output and reduce the computational complexity. Subsequently, a 5×5 separable convolution operation is performed and combined with batch normalization to further optimize the standardization and fluency of the features.

[0080] b) Feature fusion: After the convolution operation, the output features are fused with the original input to preserve the low-level features of the input and the high-level features after network processing. This fusion is performed by element-by-element addition and the ReLU activation function is applied to introduce nonlinear characteristics, thereby increasing the expressive power of the network.

[0081] c) Output features: The fused and activated feature maps are fed into subsequent layers of the network for further processing, ensuring that the model can effectively capture and utilize the multi-level features of the input data.

[0082] Furthermore, the specific structure of the JDS module in the JLU-Net network is as follows:

[0083] a) Branch processing: The JDS module receives the input feature map and divides it into two parallel branches: the first branch first performs a normal convolution operation on the input feature map to extract features, and then applies a dilated convolution to capture contextual features of different scales. The dilated convolution enhances the expressiveness of long-distance dependent features by expanding the receptive field. The second branch performs point-by-point convolution on the input feature map to compress the channel dimension, and then further splits it into two sub-paths. One sub-path highlights significant features through a maximum pooling operation, and the other sub-path retains global background information through an average pooling operation. b) Feature fusion: All output feature maps of the first and second branches are integrated through a splicing operation.

[0084] c) Output features

[0085] Furthermore, the specific structure of the JASM module in the JLU-Net network is as follows:

[0086] a) Input feature map processing: First, the input feature map is reshaped into multiple small image blocks (patches), each of which contains local feature information of a fixed size. This can effectively decompose the input image into small units that are easier to process, laying the foundation for subsequent feature extraction.

[0087] b) Linear Projection: Apply linear transformations to these small image patches to generate three types of feature maps: query, key, and value. The query, key, and value are linearly projected through a learnable weight matrix to generate feature representations suitable for the attention mechanism.

[0088] c) Adaptive input method: In the traditional attention mechanism, there is a direct matching relationship between the query and the key. In the JASM module, this direct correspondence is changed through the adaptive input method, allowing more relative position information to be transmitted between the query, key and value.

[0089] d) Feature synthesis: The processed query, key and value are interacted, the feature weights are calculated, and these features are synthesized together to form the final feature representation. This step fuses this information through convolution operations to obtain an enhanced feature map.

[0090] e) Shift MLP operation: The JASM module then applies the Shift MLP operation, first using depthwise separable convolution to extract features, then using the AugShuffleNet module to shuffle the channels in the feature map to promote cross-channel mixing of information. Next, the GELU activation function is used to perform a nonlinear transformation on the result to further enhance the diversity and nonlinear characteristics of feature expression.

[0091] f) Final output: Finally, the JASM module uses residual connections to merge the input features with the processed output to ensure effective information transfer and stabilize the training process through layer normalization operations.

[0092] Furthermore, the specific structure of the SKConv module in the JLU-Net network is as follows:

[0093] a) Input feature map processing: Given an input feature map, firstly, a preliminary convolution operation is performed on the input feature map to extract basic features and obtain an intermediate feature map.

[0094] b) Channel separation: In SKConv, the channel separation operation first processes the intermediate feature map, and then processes the features of different channels independently through deep separable convolution, reducing the amount of calculation and improving efficiency. Each channel undergoes independent convolution to extract useful feature information.

[0095] c) Detail stripping and enhancement: Detail feature maps are extracted by performing detail stripping operations on the feature maps after channel separation. This process is completed by a specially designed detail stripping module. The stripped detail map can effectively capture local information such as edges and textures in the image. Then, the enhancement operation strengthens these detail information, making the local features more prominent and improving the segmentation ability of small targets.

[0096] d) Feature enhancement: The feature map after detail enhancement is weighted fused with the original input feature map to form the final output feature map.

[0097] e) Output: The final output feature map contains enhanced local features and global information, providing rich feature representation for subsequent network layers. Through this module, JLU-Net can effectively capture features of different scales and levels, thereby enhancing the segmentation effect of lesion areas in medical images.

[0098] Example 2

[0099] The method in Example 1 was used to experimentally verify the ISIC2018 skin lesion public segmentation dataset. This experiment aims to evaluate the performance of the lightweight medical image segmentation network JLU-Net in the skin lesion segmentation task.

[0100] S1. Example Setup

[0101] S1-a) Dataset: The ISIC2018 dataset is used, which contains a large number of clearly labeled skin lesion images. The segmentation categories include normal skin area and diseased skin area.

[0102] S1-b) Segmentation task: Perform binary segmentation on the input image, with the goal of clearly separating the diseased area from the normal skin area.

[0103] S1-c) Training and test set division: The dataset is randomly divided in a ratio of 80% / 20%, where 80% of the images are used for training and 20% of the images are used for testing.

[0104] S1-d) Operating environment: The experiment was conducted on a computer system equipped with an NVIDIA RTX 4090 graphics card, the operating system was Ubuntu, the deep learning framework was PyTorch, and the CUDA version was 12.2.

[0105] S2. Implementation Flow

[0106] S2-a) Data preprocessing: The images in the ISIC2018 dataset are standardized and data augmented, including size normalization (adjusted to 512×512), random rotation, horizontal flipping, and HSI adjustment.

[0107] S2-b) Model training: The AdamW optimizer is used, the initial learning rate is set to 0.1, and the batch size is 16. The training process is carried out for 300 rounds, and the cross entropy loss, IoU index and Dice coefficient are monitored in real time on the validation set.

[0108] S2-c) Model testing: Use the trained JLU-Net model to segment the images in the test set, output the segmentation probability map of the disease area, and generate the final segmentation result through threshold processing.

[0109] S3. Example Results

[0110] S3-a) Performance indicators: On the test set, the JLU-Net of this embodiment achieved 91.75% and 84.81% in Dice coefficient and IoU coefficient, respectively, which is significantly better than traditional U-Net and other lightweight segmentation networks.

[0111] S3-b) Operation efficiency: Thanks to the lightweight design and efficient modules, JLU-Net has a computational load of 0.29M, a computational GFLOPs of 0.52, and an average FPS of 7.77 (512×512 resolution), which fully meets the needs of real-time medical applications.

[0112] S3-c) Visual effect: Some skin lesion segmentation results are shown in Figure 5 As shown in the figure, the generated skin lesion segmentation result has clear edges, complete details, and obvious segmentation boundaries between diseased areas and normal areas.

[0113] Example 3

[0114] The method in Example 1 is used to conduct an experimental verification on the BUSI breast ultrasound public segmentation dataset. This experiment aims to evaluate the performance of the lightweight medical image segmentation network JLU-Net in the skin lesion segmentation task.

[0115] S1. Example Setup

[0116] S1-a) Dataset: The BUSI dataset is used, which contains a large number of well-labeled breast ultrasound images. The segmentation categories include normal background areas and disease abnormal areas.

[0117] S1-b) Segmentation task: Perform binary segmentation on the input image, with the goal of clearly separating the diseased area from the normal area.

[0118] S1-c) Training and test set division: The dataset is randomly divided in a ratio of 80% / 20%, where 80% of the images are used for training and 20% of the images are used for testing.

[0119] S1-d) Operating environment: The experiment was conducted on a computer system equipped with an NVIDIA RTX 4090 graphics card, the operating system was Ubuntu, the deep learning framework was PyTorch, and the CUDA version was 12.2.

[0120] S2. Implementation Flow

[0121] S2-a) Data preprocessing: The images in the BUSI dataset are standardized and data augmented, including size normalization (adjusted to 256×256), random rotation, horizontal flipping, and HSI adjustment.

[0122] S2-b) Model training: The AdamW optimizer is used, the initial learning rate is set to 0.1, and the batch size is 16. The training process is carried out for 300 rounds, and the cross entropy loss, IoU index and Dice coefficient are monitored in real time on the validation set.

[0123] S2-c) Model testing: Use the trained JLU-Net model to segment the images in the test set, output the segmentation probability map of the disease area, and generate the final segmentation result through threshold processing.

[0124] S3. Example Results

[0125] S3-a) Performance indicators: On the test set, the JLU-Net of this embodiment achieved 79.94% and 67.05% in Dice coefficient and IoU coefficient, respectively, which is significantly better than traditional U-Net and other lightweight segmentation networks.

[0126] S3-b) Operation efficiency: Thanks to the lightweight design and efficient modules, JLU-Net has a computational load of 0.29M, a computational GFLOPs of 0.52, and an average FPS of 37.20 (256×256 resolution), which fully meets the needs of real-time medical applications.

[0127] S3-c) Visual effect: Some breast ultrasound image segmentation results are shown in Figure 6 As shown in the figure, the generated lesion segmentation result has clear edges, complete details, and obvious segmentation boundaries between the diseased area and the normal area.

[0128] The specific settings are based on actual construction requirements and are not specifically limited in this application.

[0129] It should be noted that, in this application, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0130] The above description is only a specific implementation of the present application, so that those skilled in the art can understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest range consistent with the principles and novel features applied for herein.

Claims

1. A lightweight network architecture for medical image segmentation, characterized in that: The following steps are involved: S1. Data Preprocessing In the lightweight network architecture for medical image segmentation of the present invention, firstly, a comprehensive data preprocessing is performed on the input medical image to ensure the quality of the input data and improve the robustness and generalization ability of the model. This step includes the following specific operations: S1-a) Data collection: Various types of medical image data are obtained from clinical medical databases or public medical image datasets, including but not limited to dermoscopic images (malignant melanoma) and breast ultrasound images. All data are collected in a standardized manner to ensure the accuracy of pixel resolution and annotation; S1-b) Random flipping: randomly flip the input image horizontally or vertically to simulate the shooting effects at different angles; S1-c) Random rotation: Randomly apply a 90-degree rotation operation to the input image; S1-d) Random change of HSI channels: Randomly perturb the hue, saturation and brightness channels of the image; S1-e) Cropping and normalization: Cropping or interpolating the input image to the fixed size required by the network (such as 256×256, 512×512) to ensure the consistency of the input data, and then normalizing the pixel values ​​and mapping them to the range of [0,1] or [-1,1] to reduce the impact of grayscale differences caused by different imaging devices on the network; S2. Feature extraction: In the lightweight network architecture for medical image segmentation of the present invention, a lightweight neural network named JLU-Net (JointLightweight U-shaped Network) is constructed to meet the requirements of medical image segmentation tasks for efficiency and accuracy. This step includes the following specific designs: S2-a) Overall architecture design: In the present invention, JLU-Net is a lightweight neural network based on a U-shaped structure. The overall architecture consists of three parts: an encoder, a decoder, and a reconstructed jump connection path. First, the encoder is responsible for multi-level feature extraction of the input image, gradually capturing deep semantic information, thereby effectively characterizing the complex lesion areas in medical images. Secondly, the decoder gradually restores the spatial resolution of the features by upsampling layer by layer, and reconstructs the segmentation target at the same time, and finally generates a segmentation probability map. Finally, the reconstructed jump connection path establishes cross-layer feature transfer between the encoder and the decoder; S2-b) Encoder design: The encoder consists of multiple encoding units, each of which includes an RLKG block and a JDS module. The LKG block adopts a lightweight convolution module design, combining a hybrid normalization method and a depth-separable convolution, aiming to efficiently extract the features of the input image while reducing the computational complexity. The JDS module implements joint downsampling, which can effectively retain the key spatial information of the input image by combining the pooling operation with the convolution operation; S2-c) Skip connection path design: A reconstructed skip connection path is set between the encoder and the decoder. This path integrates the SKConv module and the JASM module in the feature transfer process. The SKConv module introduces a dynamic kernel selection mechanism, which can be performed according to the semantic requirements of the input features, optimizing the extraction capabilities of local and global features. The JASM module uses the attention shift mechanism to focus on the salient areas in the input features, effectively enhancing the expression capabilities of small objects and edge areas; S2-d) Decoder design: The decoder consists of multiple decoding units, each of which includes an upsampling module and an RLKG block. The upsampling module restores the resolution of the feature map layer by layer using bilinear interpolation to reconstruct the same size as the original input image; S2-e) Output layer: The final output of the decoder is processed into a segmentation probability map through a 1×1 convolutional layer, and then the probability map is binarized according to the set threshold T (T=0.5) to obtain the final segmentation result.

2. A lightweight network architecture for medical image segmentation according to claim 1, characterized in that: The RLKG module in step S2 has the following structure: 2-a) Module composition: The RLKG module adopts a hierarchical convolution structure, first performing a 7×7 separable convolution operation combined with group normalization to improve the stability of the convolution output and reduce the computational complexity. Subsequently, a 5×5 separable convolution operation is performed, combined with batch normalization to further optimize the standardization and fluency of features. 2-b) Feature fusion: After the convolution operation, the output features are fused with the original input to retain the low-level features of the input and the high-level features after network processing. This fusion is performed by element-by-element addition and the ReLU activation function is applied to introduce nonlinear characteristics, thereby increasing the network's expressive power. 2-c) Output features: The fused and activated feature maps are sent to the subsequent layers of the network for further processing, ensuring that the model can effectively capture and utilize the multi-level features of the input data.

3. A lightweight network architecture for medical image segmentation as claimed in claim 1, characterized in that: The JDS module in step S2 has the following structure: 3-a) Branch processing: The JDS module receives the input feature map and divides it into two parallel branches: the first branch first performs a normal convolution operation on the input feature map to extract features, and then applies a dilated convolution to capture contextual features of different scales. The dilated convolution enhances the expression of long-distance dependent features by expanding the receptive field. The second branch performs point-by-point convolution on the input feature map to compress the channel dimension, and then further splits it into two sub-paths. One sub-path highlights significant features through a maximum pooling operation, and the other sub-path retains global background information through an average pooling operation. 3-b) Feature fusion: All output feature maps of the first and second branches are integrated through a splicing operation; 3-c) Output features.

4. A lightweight network architecture for medical image segmentation as claimed in claim 1, characterized in that: The JASM module in step S2 has the following structure: 4-a) Input feature map processing: First, the input feature map is reshaped into multiple small image blocks (patches), each of which contains local feature information of a fixed size. This can effectively decompose the input image into small units that are easier to process, laying the foundation for subsequent feature extraction; 4-b) Linear projection: Linear transformation is applied to these small image blocks to generate three types of feature maps: query, key, and value. The query, key, and value are linearly projected through a learnable weight matrix to generate feature representations suitable for the attention mechanism; 4-c) Adaptive input method: In the traditional attention mechanism, there is a direct matching relationship between the query and the key. In the JASM module, this direct correspondence is changed through the adaptive input method, allowing more relative position information to be transmitted between the query, key and value; 4-d) Feature synthesis: The processed query, key and value are interacted, the feature weights are calculated, and these features are synthesized together to form the final feature representation. This step fuses this information through convolution operations to obtain an enhanced feature map; 4-e) Shift MLP operation: The JASM module then applies the Shift MLP operation, first using depthwise separable convolution to extract features, then using the AugShuffleNet module to shuffle the channels in the feature map to promote cross-channel mixing of information, and then using the GELU activation function to perform a nonlinear transformation on the result to further enhance the diversity and nonlinear characteristics of feature expression; 4-f) Final output: Finally, the JASM module uses residual connections to merge the input features with the processed output to ensure effective information transfer and stabilize the training process through layer normalization operations.

5. A lightweight network architecture for medical image segmentation according to claim 1, characterized in that: The SKConv module in step S2 has the following structure: 5-a) Input feature map processing: Given an input feature map, first, a preliminary convolution operation is performed on the input feature map to extract basic features and obtain an intermediate feature map; 5-b) Channel separation: In SKConv, the channel separation operation first processes the intermediate feature map, and the features of different channels are processed independently through deep separable convolution to reduce the amount of calculation and improve efficiency. Each channel undergoes independent convolution to extract useful feature information; 5-c) Detail stripping and enhancement: The detail feature map is extracted by performing detail stripping operation on the feature map after channel separation. This process is completed by a specially designed detail stripping module. The stripped detail map can effectively capture local information such as edges and textures in the image. Then, the enhancement operation strengthens these detail information, making the local features more prominent and improving the segmentation ability of small targets; 5-d) Feature enhancement: The feature map after detail enhancement is weighted fused with the original input feature map to form the final output feature map; 5-e) Output: The final output feature map contains enhanced local features and global information, providing rich feature representation for subsequent network layers.

Images