A medical image automatic segmentation method based on SDAttNet

Abstract

The present application belongs to the field of medical image segmentation, and particularly relates to a medical image automatic segmentation method based on SDAttNet; the method comprises the following steps: obtaining a medical image to be processed and inputting the medical image to a trained medical image segmentation model based on SDAttNet for processing to obtain a medical image segmentation result; the medical image segmentation model based on SDAttNet is a U-shaped encoder-decoder structure, the U-shaped encoder-decoder structure adopts three SDAtt modules for skip connection; the encoder comprises five encoding subunits, and the decoder comprises five decoding subunits; the present application is more efficient when processing medical images, and the generalization ability is improved, so that the present application can better adapt to different types of medical image tasks.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical image segmentation, and particularly relates to a medical image automatic segmentation method based on an SDAttNet. BACKGROUND

[0002] Medical image segmentation is a key technology in modern medicine for extracting specific tissues or lesion regions from complex images. Image segmentation can help doctors identify the morphology, size and location of lesions, providing important basis for clinical decision-making, and has important applications in medical diagnosis, treatment planning and postoperative monitoring. Traditional medical image segmentation methods include edge detection, region growing, level set, image thresholding, etc. These segmentation methods based on prior knowledge and image characteristics have achieved some success in specific application scenarios, but mostly rely on manual design of features and are easily affected by noise, low contrast and other unfavorable factors, making it difficult to achieve consistent high-precision performance on complex and diverse medical images.

[0003] With the rapid development of artificial intelligence, especially deep learning technology, medical image segmentation has entered a new stage. Convolutional Neural Network (CNN) as a deep learning model has made breakthrough progress in image classification, detection and segmentation tasks. U-Net, FCN (Fully Convolutional Network) and other models through encoder-decoder structure can not only automatically extract multi-level features from images, but also retain high-resolution spatial information through skip connections, becoming the mainstream technology solution for medical image segmentation. These CNN-based segmentation methods have greatly improved the efficiency and accuracy of medical image segmentation, and have performed particularly outstanding in tumor detection, organ segmentation and other tasks.

[0004] However, existing deep learning methods still face some challenges in medical image segmentation. First, medical images often have complex anatomical structures, multi-scale changes and fuzzy boundaries, making it easy for the model to lack local information when extracting and processing different levels of features. Second, the multi-modal characteristics of medical images (such as CT, MRI, ultrasound, etc.) result in significant differences in texture features exhibited by images under different modalities, and conventional convolution operations are prone to miss key contextual information when capturing these features, thus affecting segmentation accuracy.

[0005] To address these issues, in recent years, attention mechanisms have been widely introduced into the field of medical image segmentation. Attention mechanisms can effectively enhance the model's feature expression ability between different scales and regions by dynamically adjusting the model's attention to different feature regions. In image processing tasks, spatial attention and channel attention are two common attention mechanisms. Spatial attention adjusts the weight distribution of the model in the spatial dimension, allowing it to focus more on important regions in the image; while channel attention optimizes information interaction between different channels, improving the model's performance in capturing detailed features. This multi-level, multi-dimensional feature optimization strategy provides a new solution for medical image segmentation, helping to improve the model's segmentation accuracy for complex structures and detailed features.

[0006] Although attention mechanisms have played a positive role in improving segmentation accuracy, existing attention mechanisms still have certain limitations in feature processing. Traditional attention mechanisms usually process spatial and channel dimension feature interactions independently, failing to fully utilize cross-level information fusion. In addition, how to improve model segmentation accuracy while maintaining computational efficiency is still an important problem faced by current deep learning medical image segmentation methods.

[0007] In this context, further optimizing feature processing in space and channel, and improving cross-level feature interaction capability, has become a key direction to promote the progress of medical image automatic segmentation technology. SUMMARY

[0008] To address the deficiencies in the prior art, the present application proposes a medical image automatic segmentation method based on SDAttNet, which comprises: obtaining a medical image to be processed and inputting it into a trained medical image segmentation model based on SDAttNet for processing to obtain a medical image segmentation result.

[0009] The medical image segmentation model based on SDAttNet is a U-shaped encoder-decoder structure, which uses three SDAtt modules for skip connection; the encoder includes five encoding subunits, and the decoder includes five decoding subunits.

[0010] Preferably, among the five encoding subunits of the encoder, the first encoding subunit includes two feature extraction units, and the remaining four encoding subunits each include a max-pooling layer and two feature extraction units; each feature extraction unit is composed of a convolution layer, a batch normalization layer (BN) and a RELU activation function.

[0011] Preferably, the first four decoding subunits of the five decoding subunits of the decoder each comprise an upsampling layer and a feature extraction unit, and the last decoding subunit is a feature extraction unit; each feature extraction unit is composed of a convolution layer, a batch normalization layer and a RELU activation function.

[0012] Preferably, the SDAtt module is composed of an SSA module and an SCA module, the input of the first SDAtt module is the output of the first to third encoding subunits, the input of the second SDAtt module is the output of the second to fourth encoding subunits, and the input of the third SDAtt module is the output of the third to fifth encoding subunits; the output features of the SDAtt module are obtained by fusing the output features of the SSA module and the SCA module, and the three SDAtt modules are connected to the second to fourth decoding subunits.

[0013] Preferably, the processing process of the SSA module on data is represented as:

[0014] Q s =K s and V qs i , V s =W vs ·X i-1

[0015]

[0016] X′ s =SAtt(A s V s )

[0017] wherein Q s , K S and V s represent the query, key and value of the SSA module respectively, W qs represents a first weight matrix, W vs represents a second weight matrix, X i represents the output feature of the i-th encoding subunit, X i-1 represents the output feature of the i-1-th encoding subunit, X′ s represents the spatial attention feature, A s represents the spatial attention weight, d ks represents the dimension of the key vector of the SSA module, and SAtt represents the spatial attention mechanism.

[0018] Preferably, the processing process of the SCA module on data is represented as:

[0019] Q c =W qc ·X i+1 , K c =W​kc ·X i , V c = W vc ·X i

[0020]

[0021] X' c = CAtt(A c V c )

[0022] wherein Q c , K c and V c represent the query, key and value of the SCA module respectively, W qc , W kc and W vc represent the third, fourth and fifth weights respectively, X i represents the output feature of the i-th encoding subunit, X i+1 represents the output feature of the i+1-th encoding subunit, A c represents the channel attention weight, X' c represents the channel attention feature, d kc represents the dimension of the SCA module key vector, and CAtt represents the channel attention mechanism.

[0023] Preferably, the output features of the fused SSA module and SCA module are represented as:

[0024] X' = γ1·X' s + γ2·X' c

[0025] wherein X' represents the enhanced feature map, i.e., the output feature of the SDAtt module, γ1 represents the spatial weight, γ2 represents the channel weight, X' s represents the spatial attention feature, and X' c represents the channel attention feature.

[0026] Preferably, the loss function for training the medical image segmentation model based on the SDAttNet is:

[0027]

[0028] wherein Loss represents the total loss of the model, n represents the number of training samples, y i represents the real label of the i-th training sample, and σ represents the sigmoid function.

[0029] The beneficial effects of the present application are:

[0030] By using the upper layer features as values in spatial processing, the present application enables the model to utilize a wider range of contextual information to optimize the feature processing of the current layer. This approach not only expands the model's perception field, but also helps to more accurately resolve context-dependent problems in complex medical images.

[0031] In channel processing, the present application uses lower layer features as queries, enabling the model to capture and reconstruct detailed information in medical images more meticulously, thereby improving the accuracy of segmentation. Through this approach, the model can better identify and utilize those tiny features and changes that may be overlooked in conventional processing.

[0032] By changing the attention mechanism on space and channel, the present application provides a novel way to reorganize and optimize the network structure. This optimization in structure not only makes the model more efficient in processing medical images, but also improves its generalization ability, enabling the model to better adapt to different types of medical image tasks. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 Structure diagram of the SDAttNet-based medical image segmentation model in the present application.

[0034] Figure 2 Structure diagram of the SDAtt module in the present application;

[0035] Figure 3 Structure diagram of the SSA module in the present application;

[0036] Figure 4 Structure diagram of the SCA module in the present application; DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0038] The present application proposes a SDAttNet-based medical image automatic segmentation method, which comprises:

[0039] Obtain the medical image to be processed and input it into the trained SDAttNet-based medical image segmentation model for processing to obtain the medical image segmentation result.

[0040] As Figure 1As shown, the medical image segmentation model based on the SDAttNet is a U-shaped encoder-decoder structure, which adopts three SDAtt modules (hierarchical attention modules) for skip connection; the encoder includes 5 encoding subunits, and the decoder includes 5 decoding subunits.

[0041] Among the 5 encoding subunits of the encoder, the first encoding subunit includes two feature extraction units, and the remaining 4 encoding subunits each include a max-pooling layer and two feature extraction units; each feature extraction unit is composed of a convolution layer, a batch normalization layer and a RELU activation function; preferably, the convolution layer adopts 3x3 hollow convolution.

[0042] Among the 5 decoding subunits of the decoder, the first 4 decoding subunits each include an up-sampling layer and a feature extraction unit, and the last decoding subunit is a feature extraction unit; each feature extraction unit is composed of a convolution layer, a batch normalization layer and a RELU activation function; preferably, the convolution layer adopts 3x3 hollow convolution.

[0043] Preferably, the skip connection of the present application is constructed by using the SDAtt module. The SDAtt module is composed of an SSA module and an SCA module, and for the input of skip connection of each layer, two-dimensional feature maps wherein X i represents the output feature map of the current layer encoder, X i-1 represents the output feature map of the previous layer encoder, and X i+1 represents the output feature map of the next layer encoder; specifically, the input of the first SDAtt module is the output of the first to third encoding subunits, the input of the second SDAtt module is the output of the second to fourth encoding subunits, and the input of the third SDAtt module is the output of the third to fifth encoding subunits; the output feature of the SDAtt module is obtained by fusing the output features of the SSA module and the SCA module, and the three SDAtt modules are connected to the second to fourth decoding subunits.

[0044] As shown in Figure 2 , the hierarchical attention module is mainly composed of a spatial attention module (SSA) and a channel attention module (SCA) in parallel. Among them, the spatial attention uses the upper layer feature X i-1 as the value (V), the current layer feature X i as the query (Q) and the key (K), performs spatial attention processing to enhance the spatial correlation of the image. The channel attention uses the lower layer feature X i+1 as the query (Q), and the current layer feature X i as the key (K) and the value (V), performs channel attention processing to enhance the channel correlation of the feature.

[0045] AsFigure 3 As shown, the data processing procedure of the SSA module is represented as follows:

[0046] Q s =K s =W qs ·X i V s =W vs ·X i-1

[0047]

[0048] X′ s =SAtt(A s V s )

[0049] Among them, Q s K s and V s These represent the query, key, and value of the SSA module, respectively. qs Let W represent the first weight matrix. vs Let X represent the second weight matrix. i X represents the output feature of the i-th coding subunit. i-1 Let X′ represent the output feature of the (i-1)th coding subunit. s Representing spatial attention features, A s Denotes spatial attention weights, d ks This indicates the dimension of the key vector in the SSA module. Specifically, it refers to the dimension of the key vector used when calculating the dot product in the attention mechanism. Attention mechanisms often involve matrix dot products. To avoid excessively large dot product values ​​(leading to gradient instability), the dot product values ​​are scaled to achieve normalization. SAtt stands for Spatial Attention.

[0050] like Figure 4 As shown, the data processing procedure of the SCA module is represented as follows:

[0051] Q c =W qc ·X i+1 K c =W kc ·X i V c =W vc ·X i

[0052]

[0053] C′ c =CAtt(A c Vc )

[0054] Among them, Q c K c and V c These represent the query, key, and value of the SCA module, respectively. qc W kc and W vc X represents the third, fourth, and fifth weights, respectively. i X represents the output feature of the i-th coding subunit. i+1 A represents the output feature of the (i+1)th coding subunit. c X′ represents the channel attention weight. c Describing channel attention features, d kc This represents the dimension of the SCA module key vector, and CAtt represents the channel attention mechanism.

[0055] Finally, the SDAtt module fuses the processed spatial attention features and channel attention features to generate the final enhanced feature map, i.e.: X′=γ1·X′ s +γ2·X′ c , where γ1 and γ2 are the learned spatial and channel weights, used to adjust the relative importance of the two attention contributions.

[0056] In the training process of the medical image segmentation model based on SDAttNet, the optimizer used is the Adam optimizer, with an initial learning rate of 0.001, Adam's β1 parameter of 0.9, Adam's β2 parameter of 0.99, a training epoch of 40, a batch size of 6, and a loss function combining cross-entropy and sigmoid, expressed as:

[0057]

[0058] Where Loss represents the total model loss, n represents the number of training samples, and y i Let represent the true label of the i-th training sample, and σ represent the sigmoid function.

[0059] After the medical image segmentation model based on SDAtNet is trained, the medical image to be processed is input into the trained medical image segmentation model to obtain the medical image segmentation result.

[0060] The above examples further illustrate the objects, technical solutions and advantages of the present application. It should be understood that the above examples are only preferred embodiments of the present application and are not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made to the present application within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for automatic medical image segmentation based on SDAttNet, characterized in that, include: The medical image to be processed is acquired and input into a trained SDAttNet-based medical image segmentation model to obtain the medical image segmentation result. The medical image segmentation model based on SDAttNet has a U-shaped encoder-decoder structure, which uses three SDAtt modules for skip connections; the encoder includes 5 encoding subunits and the decoder includes 5 decoding subunits. Of the five coding subunits of the encoder, the first coding subunit includes two feature extraction units, and the remaining four coding subunits each include a max pooling layer and two feature extraction units. Each feature extraction unit consists of a convolutional layer, a batch normalization layer, and a ReLU activation function; Of the five decoding subunits of the decoder, the first four decoding subunits each include an upsampling layer and a feature extraction unit, while the last decoding subunit is a feature extraction unit. Each feature extraction unit consists of a convolutional layer, a batch normalization layer, and a ReLU activation function; The SDAt module consists of an SSA module and an SCA module. The input of the first SDAt module is the output of the first to third coding sub-units, the input of the second SDAt module is the output of the second to fourth coding sub-units, and the input of the third SDAt module is the output of the third to fifth coding sub-units. The output features of the SDAt module are obtained by fusing the output features of the SSA and SCA modules. The three SDAt modules are respectively connected to the second to fourth decoding sub-units.

2. The automatic medical image segmentation method based on SDAttNet according to claim 1, characterized in that, The SSA module's data processing procedure is represented as follows: Q s =K s =W qs ·X i ,V s =W vs ·X i-1 X′ s =SAtt(A s V s ) Among them, Q s K s and V s These represent the query, key, and value of the SSA module, respectively. qs Let W represent the first weight matrix. vs Let X represent the second weight matrix. i X represents the output feature of the i-th coding subunit. i-1 Let X′ represent the output feature of the (i-1)th coding subunit. s Representing spatial attention features, A s Denotes spatial attention weights, d ks This represents the dimension of the SSA module key vector, and SAtt represents the spatial attention mechanism.

3. The automatic medical image segmentation method based on SDAttNet according to claim 1, characterized in that, The data processing procedure of the SCA module is represented as follows: Q c =W qc X i+1 ,K c =W kc ·X i ,V c =W vc ·X i X′ c =CAtt(A c V c ) Among them, Q c K c and V c These represent the query, key, and value of the SCA module, respectively. qc W kc and W vc X represents the third, fourth, and fifth weights, respectively. i X represents the output feature of the i-th coding subunit. i+1 A represents the output feature of the (i+1)th coding subunit. c X′ represents the channel attention weight. c Describing channel attention features, d kc This represents the dimension of the SCA module key vector, and CAtt represents the channel attention mechanism.

4. The automatic medical image segmentation method based on SDAttNet according to claim 1, characterized in that, The output characteristics of the fused SSA and SCA modules are represented as follows: X′=γ1·X′ s +γ2·X′ c Where X′ represents the enhanced feature map, i.e., the output feature of the SDAtt module, γ1 represents the spatial weight, γ2 represents the channel weight, and X′ s Representing spatial attention features, X′ c This represents the channel attention feature.

5. The automatic medical image segmentation method based on SDAttNet according to claim 1, characterized in that, The loss function for training the SDAttNet-based medical image segmentation model is: Where Loss represents the total model loss, n represents the number of training samples, and y i Let represent the true label of the i-th training sample, and σ represent the sigmoid function.

Images