Adaptive gradient preserving feature propagation network method for medical image segmentation

Abstract

The invention belongs to the technical field of image processing, and particularly relates to a self-adaptive gradient preserving feature propagation network method for medical image segmentation, which comprises the following steps of: constructing a self-adaptive gradient preserving feature propagation network, introducing the self-adaptive gradient preserving feature propagation network into a medical image segmentation model, then acquiring a medical image, and preprocessing to obtain an image set; and training the medical image segmentation model in which the adaptive gradient preserving feature propagation network is introduced. After training is completed, a medical image to be segmented is input into the trained medical image segmentation model for image segmentation, the adaptive gradient preserving feature propagation network calculates the calculation dynamic gradient weight of the corresponding layer at different layers of the network by using the gradient norm of the layer, and then adaptive adjustment of the features is achieved. Therefore, the gradient propagation strategy is dynamically adjusted at different depths. The method can dynamically adjust the importance of the feature channel based on the gradient information, effectively optimizes the gradient flow of the deep neural network, and enables the shallow feature learning to be more stable.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to an adaptive gradient-preserving feature propagation network method for medical image segmentation. Background Art

[0002] Medical image segmentation technology is crucial for disease diagnosis and treatment planning, but it remains challenging due to factors such as anatomical variability, complex tissue boundaries, and multi-scale structures. However, existing deep learning methods still face the following key problems in medical image segmentation:

[0003] Vanishing gradient and exploding gradient problems: When a deep neural network performs backpropagation, the gradient may exponentially decay (vanishing gradient) or exponentially grow (exploding gradient) as the number of layers increases. Vanishing gradient makes it difficult for shallow layers to learn low-level edge features, while exploding gradient may cause unstable model convergence. Traditional batch normalization or residual connection can alleviate this problem to some extent, but they cannot dynamically adapt to the gradient distribution of different feature layers.

[0004] Instability of feature propagation: Due to the complex organizational structure of medical images, different-scale targets may require different information transfer strategies. Existing methods lack an adaptive feature propagation mechanism and cannot dynamically adjust the importance of feature channels based on gradient information. Summary of the Invention

[0005] The purpose of the present invention is to provide an adaptive gradient-preserving feature propagation network method for medical image segmentation to solve the technical problems of vanishing gradient and exploding gradient existing in the prior art.

[0006] The described adaptive gradient-preserving feature propagation network method for medical image segmentation constructs an adaptive gradient-preserving feature propagation network, introduces the adaptive gradient-preserving feature propagation network into a medical image segmentation model, then obtains medical images, gets an image set through preprocessing, and then trains the medical image segmentation model introduced with the adaptive gradient-preserving feature propagation network. After training is completed, the medical image to be segmented is input into the trained medical image segmentation model for image segmentation. The adaptive gradient-preserving feature propagation network calculates the corresponding layer's computational dynamic gradient weight using the gradient norm of that layer at different layers of the network, thereby realizing the adaptive adjustment of features and dynamically adjusting the gradient propagation strategy at different depths.

[0007] Preferably, in the adaptive gradient-preserving feature propagation network, the feature map X input to the l-th layer is defined

[0005] , , l , l , , ,

[0007] , ,

[0006] The gradient norm G at the l-th layer of the network l , and the calculation formula is: Among them represents the output of the l-th layer of the network;

[0008] Then, calculate the dynamic gradient weight W l , which is used to adjust the gradient contribution of feature propagation; the calculation formula of the dynamic gradient weight W l is as follows:

[0009] W l =σ(W1G l +b1)

[0010] Among them, W1 and b1 respectively represent the gradient weight and bias term in the first layer of the network, and σ() is the Sigmoid activation function, which is used to map the gradient weight to the range of (0,1);

[0011] Use the dynamic gradient weight to perform feature fusion between the upper and lower layers to achieve adaptive adjustment. Finally, the adaptively adjusted feature transmission is as follows: X′ l =W l ·X l +(1-W l )·X l-1 , where, through the dynamic gradient weight W l , the features of the l-th layer and the l-1-th layer are fused to obtain the adjusted feature X′ l of the l-th layer of the network.

[0012] Preferably, an attention gating feature propagation mechanism is introduced and processed through an attention gate module after multiple convolutional processes.

[0013] Preferably, in the U-Net architecture model, the adaptive gradient retention feature propagation network is set on the skip connection between the encoder and the decoder.

[0014] Preferably, in the U-Net architecture model, first, based on the encoder feature X enc and the decoder feature X dec calculate the gating weight G, and the calculation formula is: G = σ(W g X enc +b g ), where W g and b g are the weight and bias term of the corresponding network layer; then generate the optimized skip connection feature X skip , and the calculation formula is: X skip =G·X enc +(1-G)·X dec ; then, based on the different depths of the network, dynamically adjust the gradient propagation strategy for the optimized skip connection feature.

[0015] Preferably, in the DeepLabV3+ architecture model, the adaptive gradient-preserving feature propagation network is set after the three 3×3 atrous convolution modules with different dilation rates in the ASPP module to process the output of the atrous convolution modules. In the ASPP module, the results of the 1×1 atrous convolution and the global pooling result are connected to the outputs of the three adaptive gradient-preserving feature propagation networks, and then output after a 1×1 convolution.

[0016] Preferably, the medical image segmentation model includes a U-Net architecture model and a DeepLabV3+ architecture model.

[0017] The present invention has the following advantages:

[0018] 1. By introducing an adaptive gradient-preserving mechanism, the present invention dynamically adjusts the gradient propagation strategy at different depths of the adaptive gradient-preserving feature propagation network. In this way, shallow features can obtain more gradient signals in the early learning stage, while deep features can adaptively reduce the gradient influence in the later learning stage. This realizes the feature propagation process and can dynamically adjust the importance of feature channels based on gradient information, effectively optimizing the gradient flow of the deep neural network, making the shallow feature learning more stable, and at the same time avoiding the gradient disappearance of deep features, that is, it can avoid overfitting and gradient instability at the same time.

[0019] 2. After introducing the attention gating strategy, the present invention introduces a feature screening mechanism in the skip connection to enhance the propagation of useful features while suppressing inefficient features. This enables the model to dynamically adjust the effectiveness of the skip connection, reduce the interference of irrelevant information while maintaining semantic consistency, and make the feature interaction between the encoder and the decoder more efficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the adaptive gradient-preserving feature propagation network in the method of the adaptive gradient-preserving feature propagation network for medical image segmentation of the present invention.

[0021] Figure 2 It is a schematic diagram of applying the adaptive gradient-preserving feature propagation network in the DeepLabV3+ architecture model.

[0022] Figure 3 It is a schematic diagram of applying the adaptive gradient-preserving feature propagation network in the U-Net architecture model.

[0023] Figure 4 It is a visualization comparison diagram of the segmentation results of the present invention and the prior art applied to the COVID-19 lung and infection segmentation dataset (3D CT).

[0024] Figure 5It is a visualization comparison chart of the segmentation results of the present invention and the prior art applied to the CAS2023 dataset (3D MRA). Detailed implementation manners

[0025] Next, with reference to the accompanying drawings, through the description of the embodiments, the specific implementation manners of the present invention will be further described in detail to help those skilled in the art have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0026] As Figures 1-5 shown, the present invention provides an adaptive gradient-preserving feature propagation network method for medical image segmentation, including: constructing an adaptive gradient-preserving feature propagation network (AGP-Net) and introducing the adaptive gradient-preserving feature propagation network into a medical image segmentation model. The medical image segmentation model may include a U-Net architecture model or a DeepLabV3+ architecture model. Then, a medical image is obtained, an image set is obtained through preprocessing, and then the medical image segmentation model introduced with the adaptive gradient-preserving feature propagation network is trained. After training is completed, the medical image to be segmented is input into the trained medical image segmentation model for image segmentation.

[0027] The core idea of constructing the adaptive gradient-preserving feature propagation network (AGP-Net) is to dynamically adjust the gradient propagation strategy at different depths of the network to prevent the problems of gradient disappearance or gradient explosion. The specific implementation method is as follows.

[0028] First, define the feature map X input to the l-th layer l and the gradient norm G at the l-th layer of the network l , and the calculation formula is: where represents the output of the l-th layer of the network.

[0029] Then, calculate the dynamic gradient weight W l , which is used to adjust the gradient contribution of feature propagation. The calculation formula of the dynamic gradient weight W l is:

[0030] W l =σ(W1G l +b1)

[0031] where W1 and b1 respectively represent the gradient weight and bias term at the first layer of the network, and σ() is the Sigmoid activation function, which is used to map the gradient weight to the range of (0,1).

[0032] Use the dynamic gradient weight to perform feature fusion between the upper and lower layers to achieve adaptive adjustment. Finally, the adaptively adjusted feature transmission is as follows: X' l =W l ·Xl +(1 - W l )·X l-1 , where the features of the l-th layer and the (l - 1)-th layer are fused through the dynamic gradient weight W l to obtain the adjusted feature X' of the l-th layer of the network l .

[0033] In this way, shallow features can obtain more gradient signals in the early learning stage, while deep features can adaptively reduce the gradient influence in the later learning stage to avoid overfitting or gradient instability. Introducing the above-mentioned adaptive gradient-preserving feature propagation network into the corresponding architecture of the medical image segmentation model solves the technical problems of the prior art.

[0034] In the adaptive gradient-preserving feature propagation network, an Attention-Gated Feature Propagation (AGFP) mechanism is also introduced and processed through an attention gate module after multiple convolutional processes. The adaptive gradient-preserving feature propagation network can be applied to medical image segmentation models with U-Net architecture or DeepLabV3+ architecture.

[0035] In the U-Net architecture model, the adaptive gradient-preserving feature propagation network is set on the skip connection between the encoder and the decoder. The traditional skip connection cannot effectively distinguish important and irrelevant features, which may lead to noise propagation. The present invention introduces a feature screening mechanism through the attention gate module, which can enhance the propagation of useful features while suppressing inefficient features.

[0036] In the U-Net architecture model, first, based on the encoder feature X enc and the decoder feature X dec the gating weight G is calculated, and the calculation formula is: G = σ(W g X enc + b g ), where W g and b g are the weights and bias terms of the corresponding network layer. Then the optimized skip connection feature X skip is generated, and the calculation formula is: X skip = G·X enc + (1 - G)·X dec . The above improvement enables the model to dynamically adjust the effectiveness of the skip connection, reducing the interference of irrelevant information while maintaining semantic consistency. Then, based on the different depths of the network, the gradient propagation strategy of the optimized skip connection feature is dynamically adjusted. The U-Net architecture model thus enables low-level features to better propagate under the optimization of the gradient flow, improving the ability to identify tissue boundaries.

[0037] In the DeepLabV3+ architecture model, the Adaptive Gradient Preserving Feature Propagation Network is set after three 3×3 atrous convolution modules with different dilation rates in the Atrous Spatial Pyramid Pooling (ASPP) module to process the output of the atrous convolution modules. In the ASPP module, the result of the 1×1 atrous convolution and the result of global pooling are connected to the outputs of the three Adaptive Gradient Preserving Feature Propagation Networks, and then a 1×1 convolution is performed for output. This solution enables multi-scale features to not only adaptively adjust their importance but also optimize the gradient influence of different-scale features, thereby enhancing the robustness of the segmentation network.

[0038] The DeepLabV3+ architecture model and the U-Net architecture model can also be combined. In this case, the DeepLabV3+ architecture model is used to process the feature information of the deepest layer of the U-Net architecture model. The Adaptive Gradient Preserving Feature Propagation Network can be set both on the skip connections and in the ASPP module.

[0039] The medical images of different modalities that the present invention can process include CT, MRI, ultrasound, pathological images, etc. Especially in low-contrast ultrasound images, the AGP mechanism can enhance tissue boundary features and reduce missegmentation. Experimental results show that the medical image segmentation model incorporating AGP-Net has achieved performance improvements superior to traditional methods on different datasets.

[0040] Apply the present invention to the COVID-19 Lung and Infection Segmentation Dataset (3D CT) and the CAS2023 Dataset (3D MRA) respectively, apply other existing technologies to the same datasets, and compare the segmentation results of each method. The visualization comparison diagrams are as shown in Figure 4 and Figure 5 shown. After analyzing the segmentation results of each method, the method provided by the present invention has continuously achieved better results compared to the existing technologies, surpassing various methods of the existing technologies. Despite the variations in imaging formats, anatomical structures, and pathological features among different types of medical images, the present invention still demonstrates strong generalization ability and establishes a new benchmark for medical image segmentation. These results highlight the adaptability and robustness of the present invention in different imaging formats and segmentation tasks, and demonstrate its effectiveness in accurately depicting complex medical structures. The optimized feature propagation strategy of the present invention ensures that low-level details (such as small blood vessel structures in MRA) are retained throughout the network, thereby significantly enhancing the segmentation accuracy in high-structural complexity regions.

[0041] The present invention has been described exemplarily above in conjunction with the accompanying drawings. Obviously, the specific implementation of the present invention is not limited by the above-mentioned manner. As long as various non-substantive improvements are made by adopting the inventive concept and technical solution of the present invention, or the inventive concept and technical solution of the present invention are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. Adaptive Gradient-Preserving Feature Propagation Network Method for Medical Image Segmentation, characterized in that: Construct an adaptive gradient-preserving feature propagation network, introduce the adaptive gradient-preserving feature propagation network into the medical image segmentation model, then obtain medical images, get an image set through preprocessing, and then train the medical image segmentation model introduced with the adaptive gradient-preserving feature propagation network. After training, input the medical image to be segmented into the trained medical image segmentation model for image segmentation. The adaptive gradient-preserving feature propagation network calculates the corresponding layer's computational dynamic gradient weight using the gradient norm of that layer at different layers of the network, thereby realizing the adaptive adjustment of features, and thus dynamically adjusting the gradient propagation strategy at different depths.

2. The adaptive gradient-preserving feature propagation network method for medical image segmentation according to claim 1, wherein: In the adaptive gradient-preserving feature propagation network, the feature map \(X\) of the \(l\)-th input layer is defined l The gradient norm \(G\) of the \(l\)-th layer in the network l , and the calculation formula is: where represents the output of the \(l\)-th layer of the network; Then, calculate the dynamic gradient weight W l , which is used to adjust the gradient contribution of feature propagation; the calculation formula of the dynamic gradient weight W l is as follows: W l = σ(W1G l + b1) Among them, W1 and b1 respectively represent the gradient weight and bias term in the first layer of the network, and σ() is the Sigmoid activation function, which is used to map the gradient weight to the range of (0, 1); Feature fusion between upper and lower layers is performed using dynamic gradient weights to achieve adaptive adjustment. Finally, the adaptively adjusted features are transmitted as follows: X′ l = W l ·X l +(1 - W l )·X l-1 , where the features of the l-th layer and the (l - 1)-th layer are fused through the dynamic gradient weight W l to obtain the adjusted features X′ l of the l-th layer of the network.

3. The adaptive gradient-preserving feature propagation network method for medical image segmentation according to claim 2, characterized in that: An attention gating feature propagation mechanism is introduced and processed through an attention gate module after multiple convolutional processes.

4. The adaptive gradient-preserving feature propagation network method for medical image segmentation according to any one of claims 1-3, characterized in that: In the U-Net architecture model, the adaptive gradient-preserving feature propagation network is set on the skip connection between the encoder and the decoder.

5. The adaptive gradient-preserving feature propagation network method for medical image segmentation according to claim 4, wherein: In the U-Net architecture model, first, based on the encoder feature X enc and the decoder feature X dec calculate the gating weight G, and the calculation formula is: G = σ(W g X enc + b g ), where W g and b g are the weights and bias terms of the corresponding network layer; then generate the optimized skip connection feature X skip , and the calculation formula is: X skip = G · X enc + (1 - G) · X dec ; then dynamically adjust the gradient propagation strategy for the optimized skip connection feature based on different depths of the network.

6. The adaptive gradient-preserving feature propagation network method for medical image segmentation according to any one of claims 1-3, characterized in that: In the DeepLabV3+ architecture model, the adaptive gradient-preserving feature propagation network is set after three 3×3 atrous convolution modules with different dilation rates in the ASPP module to process the output of the atrous convolution module; in the ASPP module, after the 1×1 atrous convolution result and the global pooling result are connected to the output of the three adaptive gradient-preserving feature propagation networks, they are output after a 1×1 convolution.

7. The adaptive gradient-preserving feature propagation network method for medical image segmentation according to claim 1, characterized in that: The medical image segmentation model includes a U-Net architecture model and a DeepLabV3+ architecture model.

Images