Adaptive gradient preserving feature propagation network method for medical image segmentation

Abstract

The application belongs to the technical field of image processing, and particularly relates to an adaptive gradient preserving feature propagation network method for medical image segmentation. An adaptive gradient preserving feature propagation network is constructed, the adaptive gradient preserving feature propagation network is introduced into a medical image segmentation model, then a medical image is acquired, an image set is obtained through preprocessing, and the medical image segmentation model with the introduced adaptive gradient preserving feature propagation network is trained. After the 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 utilizes the gradient norm of a layer to calculate the dynamic gradient weight of the corresponding layer at different layers of the network, and further realizes adaptive adjustment of features, so as to dynamically adjust the gradient propagation strategy at different depths. The application can dynamically adjust the importance of feature channels based on gradient information, effectively optimize the gradient flow of the deep neural network, and make the shallow feature learning more stable.

Description

TECHNICAL FIELD

[0001] The application 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

[0002] Medical image segmentation technology is crucial for disease diagnosis and treatment planning, but it is still challenging due to 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] Gradient vanishing and gradient explosion: When deep neural networks perform backpropagation, the gradient may exponentially decay (gradient vanishing) or exponentially grow (gradient explosion) as the number of layers increases. Gradient vanishing makes it difficult for shallow layers to learn low-level edge features, while gradient explosion can lead to unstable model convergence. Traditional batch normalization (Batch Normalization) or residual connection (Residual Connection) can alleviate this problem, but cannot dynamically adapt to the gradient distribution of different feature layers.

[0004] Instability of feature propagation: Due to the complex structure of medical images, different scales of targets may require different information transmission strategies. Existing methods lack adaptive feature propagation mechanisms and cannot dynamically adjust the importance of feature channels based on gradient information. SUMMARY

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

[0006] The 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 acquires a medical image, obtains an image set through preprocessing, and trains the medical image segmentation model with the introduced adaptive gradient-preserving feature propagation network. After training, 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 dynamic gradient weight of the corresponding layer using the gradient norm of the layer at different layers of the network, and then realizes adaptive adjustment of the features, thereby dynamically adjusting the gradient propagation strategy at different depths.

[0007] Preferably, in the adaptive gradient-preserving feature propagation network, the input feature map X l The gradient norm G l at the lth layer of the network is calculated as: wherein represents the output of the l-th layer of the network;

[0008] Then, the dynamic gradient weight W l is calculated, 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] wherein W1 and b1 represent the gradient weight and the bias term of the first layer of the network respectively, and σ() is a Sigmoid activation function, which is used to map the gradient weight to the range of (0, 1);

[0011] The features of the upper and lower layers are fused by using the dynamic gradient weight, and the adaptive adjustment is realized, and finally, the feature transmission after the adaptive adjustment is as follows: X′ l =W l ·X l +(1-W l )·X l-1 , wherein the features of the l-th layer and the (l-1)-th layer are fused by using the dynamic gradient weight W l , and the adjusted feature X′ l of the l-th layer of the network is obtained.

[0012] Preferably, an attention gate feature propagation mechanism is introduced, and after the multi-layer convolution processing, the features are processed by the attention gate module.

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

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

[0015] Preferably, in the DeepLabV3+ architecture model, the adaptive gradient preserving feature propagation network is arranged after the three 3*3 dilated convolution modules with different dilated rates in the ASPP module, and the output of the dilated convolution module is processed; in the ASPP module, the 1*1 dilated convolution result and the global pooling result are connected with the output of the three adaptive gradient preserving feature propagation networks, and then 1*1 convolution is performed to output.

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

[0017] The present application has the following advantages:

[0018] 1. The present application introduces an adaptive gradient preserving mechanism, which 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 the feature channel based on the gradient information, effectively optimizing the gradient flow of the deep neural network, making the shallow feature learning more stable, and avoiding the gradient vanishing of the deep feature, that is, avoiding overfitting and gradient instability at the same time.

[0019] 2. After introducing the attention gate strategy, the present application introduces a feature screening mechanism in the skip connection to enhance the propagation of useful features and suppress inefficient features. The model can 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 DRAWINGS

[0020] Figure 1 FIG. 1 is a schematic diagram of an adaptive gradient preserving feature propagation network in the present application for medical image segmentation.

[0021] Figure 2 FIG. 3 is a schematic diagram of applying an adaptive gradient preserving feature propagation network in a DeepLabV3+ architecture model.

[0022] Figure 3 FIG. 4 is a schematic diagram of applying an adaptive gradient preserving feature propagation network in a U-Net architecture model.

[0023] Figure 4 FIG. 6 is a visualization comparison diagram of the segmentation results of the present application and the prior art applied to a COVID-19 lung and infection segmentation dataset (3D CT).

[0024] Figure 5is a visualization comparison chart of the segmentation results of the application and the prior art applied to the CAS2023 dataset (3D MRA). DETAILED DESCRIPTION

[0025] The specific embodiments of the application will be further described in detail below with reference to the accompanying drawings and by describing the embodiments, to help the skilled in the art have a more complete, accurate and in-depth understanding of the inventive concept and technical solutions of the application.

[0026] As Figures 1-5 shown, the application provides an adaptive gradient preserving feature propagation network method for medical image segmentation, comprising: 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 can contain a U-Net architecture model or a DeepLabV3+ architecture model. Then, a medical image is acquired, an image set is obtained through preprocessing, and the medical image segmentation model introduced with the adaptive gradient preserving feature propagation network is trained. After training, a 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 problem of gradient vanishing or gradient explosion, and the specific implementation method is as follows.

[0028] First, define the input feature map X l of the lth layer. l The gradient norm G of the lth layer in the network is calculated as follows: wherein X represents the output of the lth layer of the network.

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

[0030] W l = σ(W1G l +b1)

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

[0032] The dynamic gradient weight is used for feature fusion between upper and lower layers, and adaptive adjustment is realized. Finally, the feature transmission after adaptive adjustment is as follows: X′ l = W l ·Xl +(1-W l )·X l-1 , wherein the features of the l-th layer and the (l-1)-th layer are fused by a dynamic gradient weight W l to obtain the adjusted feature X' of the l-th layer of the network l .

[0033] In this way, the shallow features can obtain more gradient signals in the early learning stage, and the deep features can adaptively reduce the gradient influence in the later learning stage to avoid overfitting or gradient instability. The adaptive gradient preserving feature propagation network is introduced into the medical image segmentation model of the corresponding architecture, which 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 introduced to process the features after multiple convolutional processing through an attention gate module. The adaptive gradient preserving feature propagation network can be applied to a medical image segmentation model of a U-Net architecture or a DeepLabV3+ architecture.

[0035] In the U-Net architecture model, the adaptive gradient preserving feature propagation network is arranged on the skip connection between the encoder and the decoder. The conventional skip connection cannot effectively distinguish between important and irrelevant features, which may cause noise propagation. The feature screening mechanism is introduced by the attention gate module, which can enhance the propagation of useful features while suppressing inefficient features.

[0036] In the U-Net architecture model, first, the gating weight G is calculated based on the encoder feature X enc and the decoder feature X dec , and the calculation formula is G=σ(W g X enc +b g ), wherein 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 while reducing the interference of irrelevant information while maintaining semantic consistency. Then the gradient propagation strategy is dynamically adjusted based on the different depths of the network for the optimized skip connection feature, so that the U-Net architecture model enables the low-level features to be better propagated under the optimization of the gradient flow, thereby improving the recognition ability of the tissue boundary.

[0037] In the DeepLabV3+ architecture model, the adaptive gradient preserving feature propagation network is arranged after the three dilated rate different 3*3 dilated convolution modules of the atrous spatial pyramid pooling (ASPP) module, and the output of the dilated convolution module is processed. In the ASPP module, the 1*1 dilated convolution result and the global pooling result are connected with the output of the three adaptive gradient preserving feature propagation networks, and then output after 1*1 convolution. This scheme 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, at this time the DeepLabV3+ architecture model is used to process the deepest layer feature information of the U-Net architecture model. And the adaptive gradient preserving feature propagation network can be arranged on the skip connection and in the ASPP module at the same time.

[0039] The medical images of different modalities that can be processed by the present application include CT, MRI, ultrasound, pathological images, etc. Especially in low-contrast ultrasound images, the AGP mechanism can enhance the tissue boundary features and reduce missegmentation. Experimental results show that the medical image segmentation model introduced AGP-Net has achieved performance improvement better than traditional methods on different data sets.

[0040] The present application is applied to COVID-19 lung and infection segmentation data set (3D CT) and CAS2023 data set (3D MRA) respectively, other prior art is applied to the same data set, and the segmentation results of each method are compared, and the visual comparison chart is as shown in Figure 4 and Figure 5 After analyzing the segmentation results of each method, the method provided by the present application continuously achieves better effect compared with prior art, and surpasses various methods of prior art. Although different types of medical images have changes in imaging format, anatomical structure and pathological characteristics, the present application still shows strong generalization ability and establishes a new benchmark for medical image segmentation. These results highlight the adaptability and robustness of the present application in different imaging formats and segmentation tasks, and demonstrate its effectiveness in accurately delineating complex medical structures. The optimized feature propagation strategy of the present application ensures that low-level details (such as small blood vessel structures in MRA) are maintained throughout the network, thereby significantly enhancing the segmentation accuracy of high structural complexity regions.

[0041] The above describes the present application in conjunction with the drawings, and it is obvious that the specific implementation of the present application is not limited by the above manner, as long as various non-essential improvements are made by adopting the inventive concept and technical solution of the present application, or the inventive concept and technical solution of the present application is directly applied to other occasions without improvement, all within the protection scope of the present application.

Claims

1. An adaptive gradient preserving feature propagation network method for medical image segmentation, characterized in that: The adaptive gradient preserving feature propagation network is introduced into the medical image segmentation model, and then medical images are acquired, an image set is obtained through preprocessing, and the medical image segmentation model with the adaptive gradient preserving feature propagation network is trained; after the 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 dynamic gradient weight of the corresponding layer by using the gradient norm of the layer at different layers of the network, and then realizes the adaptive adjustment of the features, so as to dynamically adjust the gradient propagation strategy at different depths. In the adaptive gradient-preserving feature propagation network, the input is defined as... l Feature map of layer In the network l Gradient norm of the layer The calculation formula is: ,in Indicates the network number l Layer output; Then, the dynamic gradient weight is calculated for adjusting the gradient contribution of feature propagation; the calculation formula of the dynamic gradient weight is as follows: wherein, and denote the gradient weight and bias term at layer 1 of the network, respectively, ( ) is a Sigmoid activation function used to map the gradient weight to the range (0, 1); The feature fusion of the upper and lower layers is performed by using dynamic gradient weights, adaptive adjustment is realized, and finally, the adaptive adjusted features are transmitted as follows: wherein the dynamic gradient weights are The features of the first l layer and the first l layer are fused to obtain the adjusted features of the network first l layer. ​ 2. The adaptive gradient-preserving feature propagation network method for medical image segmentation according to claim 1, characterized in that: The attention gate feature propagation mechanism is introduced, and after being processed by multiple layers of convolution, the attention gate module is used for processing.

3. The adaptive gradient-preserving feature propagation network method for medical image segmentation of any one of claims 1-2, characterized in that: In the U-Net architecture model, the adaptive gradient preserving feature propagation network is arranged on the skip connection between the encoder and the decoder.

4. The adaptive gradient-preserving feature propagation network method for medical image segmentation of claim 3, wherein: In the U-Net architecture model, first, the gating weight is calculated based on the encoder feature and the decoder feature , the calculation formula is: wherein and are the weight and bias items of the corresponding network layer; then the optimized skip connection feature is generated, the calculation formula is: ; and then the gradient propagation strategy is dynamically adjusted based on the different depths of the network.​ 5. The adaptive gradient-preserving feature propagation network method for medical image segmentation of any one of claims 1-2, wherein: In the DeepLabV3+ architecture model, the adaptive gradient preserving feature propagation network is arranged after three 3*3 dilated convolution modules with different dilated rates in the ASPP module, and the output of the dilated convolution module is processed; in the ASPP module, the 1*1 dilated convolution result and the global pooling result are connected with the output of the three adaptive gradient preserving feature propagation networks, and then 1*1 convolution is performed to output.

6. The adaptive gradient-preserving feature propagation network method for medical image segmentation of claim 1, wherein: The medical image segmentation model includes a U-Net architecture model and a DeepLabV3+ architecture model.

Images