Medical image segmentation method and device based on differential perception

Abstract

The invention provides a medical image segmentation method and device based on differential perception, and the method comprises the steps: S1, obtaining a medical image data set, and carrying out the preprocessing of the data set; s2, constructing a coding model based on ResNet50, and generating an uncertainty graph and a key region sensing prompt; s3, constructing a context awareness compensation module CAC for enhancing, correcting and compensating encoder features and providing higher-quality characterization for a subsequent decoding stage; s4, constructing a hierarchical uncertainty Transform module SUTrans, and explicitly distributing attention intensity according to the segmentation difficulty difference of the region to realize key modeling of the key region and effectively reduce redundancy; and S5, constructing a double-branch decoder, respectively decoding the shallow-layer features and the deep-layer semantic information, and generating a final segmentation result through feature fusion. According to the invention, through a differential perception strategy, the model has a stronger distinguishing capability when processing areas with fuzzy lesion boundaries, low contrast and the like, so that the segmentation precision is improved.

Description

Technical Field

[0001] This invention relates to the field of medical imaging, and more specifically to a medical image segmentation method and apparatus based on differential perception. Background Technology

[0002] Medical image segmentation is a key component of computer-aided diagnosis and treatment planning. Its goal is to perform semantic-level classification of image pixels to achieve accurate depiction of anatomical structures or lesion regions. High-quality segmentation results are of great significance for quantitative organ assessment, precise lesion localization, and clinical decision support. However, due to the complex imaging mechanism of medical images, the segmentation task still faces many challenges: (1) There is often a blurred transition area between the lesion and the surrounding tissue, resulting in low contrast and insignificant differences between the foreground and background; (2) The complex texture of the background tissue is easily confused with the lesion in terms of color and brightness distribution; (3) The significant differences in lesion scale make it difficult for the model to achieve a balance between global semantic modeling and local detail depiction.

[0003] Early medical image segmentation methods primarily relied on traditional techniques such as template matching, thresholding, and clustering analysis. These methods typically required manually designed features and depended on empirical parameter settings, thus limiting their applicability to simple scenarios with clear boundaries and significant foreground-background differences. With the development of deep learning technology, convolutional neural network-based segmentation models have made significant progress in the automatic learning and representation of medical image features, gradually becoming the mainstream approach in this field. Although existing methods have made significant progress in feature extraction and attention optimization, when faced with images exhibiting significant differences in segmentation difficulty, most still employ uniform feature enhancement and attention allocation strategies, ignoring the mismatch between segmentation difficulty and attention intensity. This weakens the focus on key regions and introduces a large amount of redundant computation. Therefore, how to introduce a differential perception mechanism into the segmentation network to enhance the model's selective response to different regions, thereby further improving the accuracy of medical image segmentation, has become a crucial problem urgently needing to be solved in current research. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a medical image segmentation method and apparatus based on differential perception, which can enable users to accurately segment lesions in medical images.

[0005] Specifically, the method includes the following steps: S1: Obtain the medical image dataset and perform preprocessing operations on the dataset; S2: Construct a ResNet50-based encoding model to generate uncertainty maps and key region awareness cues; Preferably, S2 includes the following steps: S2.1: Construct an encoding model consisting of a ResNet50-based encoder and a segmentation module. The segmentation module comprises two... The module consists of a sigmoid activation function; S2.2: Transfer the dataset Input the encoding model to obtain the predicted probability map , predict probability map Binarization provides key area perception prompts. ; S2.3: Using information entropy as an uncertainty measure, the prediction probability map is analyzed. Various spatial locations The uncertainty of the prediction is estimated to obtain the uncertainty diagram. ; S3: Construct a context-aware compensation module (CAC) to enhance, correct, and compensate encoder features, providing higher-quality representations for subsequent decoding stages; Preferably, S3 includes the following steps: S3.1: Construct the Context-Aware Compensation (CAC) module, which consists of... Module, Module, Spatial Attention Module Spatial Attention and Channel Attention Module (ChannelAttention) is composed of; S3.2: Extract the features from the encoder With key area perception prompts Inputting the data into the CAC module performs feature enhancement, correction, and compensation to obtain enhanced features. ; S4: Construct a hierarchical uncertainty Transformer module SUTrans, which explicitly allocates attention intensity according to the difference in segmentation difficulty of regions, so as to achieve key modeling of critical regions and effectively reduce redundancy; Preferably, S4 includes the following steps: S4.1: Construct the Stratified UncertaintyTransformer module SUTrans. S4.2: Uncertainty diagram Enhanced features and features extracted by the previous SUTrans module The SUTrans module is used as a common input to perform differentiated global modeling, thereby obtaining differentiated features of interest. ; S5: Construct a dual-branch decoder to decode shallow features and deep semantic information respectively, and generate the final segmentation result through feature fusion; Preferably, S5 includes the following steps: S5.1: Construct a dual-branch decoder. The features extracted by the two shallow SUTrans modules are input into the shallow decoder to obtain... The features extracted by the two deep SUTrans modules are input into the deep decoder to obtain... ; S5.2: Construct a feature fusion module to integrate the output features of the dual-branch decoder. and Input the feature fusion module to obtain the final segmentation result. .

[0006] Preferably, S1 includes the following steps: S1.1: Obtain the medical image segmentation dataset and divide it into a training set and a test set in an 8:2 ratio; S1.2: Randomly shuffle the image samples in the training set to eliminate the influence of sample order on the model training process; S1.3: Adjust the images and their corresponding labeled masks in the training and test sets to a uniform size of 256×256 to ensure that different samples have a consistent resolution when inputting into the model; S1.4: Perform random data augmentation operations on the medical images in the training set, including random angle rotation, horizontal and vertical flipping, and random region occlusion.

[0007] Preferably, S2.2 includes the following steps: By using a convolutional neural network to perform convolution calculations, the structure of the encoding model is obtained:

[0008] In the above formula, This represents a probability prediction graph. Representative based on residual network, Represents the sigmoid activation function. It consists of three consecutive operations, consisting of a convolutional layer, batch normalization, and ReLU activation function; Using a binarization method, for Threshold segmentation is performed to obtain key area perception cues. :

[0009] In the above formula, The preset threshold is preferred. , This represents the set of spatial pixels in the predicted probability map; S2.3 includes the following steps: Using information entropy calculation, for Uncertainty estimation is performed to obtain the uncertainty diagram. :

[0010] In the above formula, Represents pixels Belongs to the first Predicted probability of class For the number of categories, A small constant to prevent numerical underflow.

[0011] Preferably, S3.2 includes the following steps: Using a convolutional neural network to perform convolution calculations, the structure of CAC is obtained:

[0012]

[0013]

[0014] In the above formula, Represents the features extracted by the encoder. This represents the output characteristics of the CAC module. The hole rate r of the convolutional layer is represented by Module, , The result of the intermediate module, Represents the spatial attention module. Represents the channel attention module. This represents a channel-level splicing operation. This represents the ReLU activation function.

[0015] Preferably, S4.2 includes the following steps: For uncertainty diagram Perform statistical analysis and calculate its mean. Standard deviation and median This characterizes the overall confidence distribution of the prediction results. Based on this, high and low percentile thresholds are dynamically determined:

[0016] In the above formula, and These represent the first and second digits of the entropy distribution. and Percentiles. Since the encoding and CAC module processing have enabled the model to accurately identify most regions, the initial settings... And adaptively fine-tuned based on global statistical characteristics: when or A higher confidence level indicates a lower overall confidence level for the model, so it should be appropriately reduced. and To expand the scope of the high uncertainty region; conversely, to increase the threshold to further reduce redundant calculations; when When it is large, it will decrease at the same time. and This is to cover more potential misclassifications. Ultimately, the uncertainty region is divided into three levels based on the entropy value range:

[0017] In the above formula, This is a multi-level uncertainty map, corresponding to high, medium, and low uncertainty regions. The computational mechanism adaptively captures the dynamic changes in entropy distribution under different statistical characteristics, thereby achieving fine-grained uncertainty modeling and regionally differentiated representation in the spatial dimension. Multi-level uncertainty diagram Enhanced features and features extracted by the previous SUTrans module After processing, the common input is the hierarchical uncertainty Transformer module SUTrans, resulting in the SUTrans structure:

[0018] In the above formula, This indicates the differentiated features of interest extracted by SUTrans. This represents a channel-level splicing operation. This indicates an upsampling operation; where, a multi-level uncertainty graph is used. The attention weight matrix is ​​used to extract the multi-head self-attention mechanism in the weighted SUTrans module, thereby achieving differentiated attention.

[0019] Preferably, S5.1 includes the following steps: A convolutional neural network is used to perform convolution calculations, resulting in the decoder structure:

[0020]

[0021] In the above formula, , and , These represent the differential interest features extracted by the two shallow SUTrans modules and the two deep SUTrans modules, respectively. This represents the features extracted by the shallow decoder. This represents the features extracted by the deep decoder; S5.2 includes the following steps: A convolutional neural network is used to perform convolution calculations, resulting in the feature fusion module structure:

[0022] In the above formula, This indicates the final segmentation result.

[0023] The beneficial effects of the method and apparatus of this invention are as follows: First, the invention preprocesses medical image data to construct a dataset for model training and testing. Based on this, a ResNet50-based encoding model is constructed to generate uncertainty maps and key region awareness cues, which are then used as inputs to subsequent network structures, guiding the model to focus on potential key regions during feature extraction. Subsequently, a context-aware compensation module enhances, corrects, and compensates the features extracted by the encoder, providing more comprehensive and high-quality feature representations for the subsequent decoding stage. Furthermore, this invention introduces a hierarchical uncertainty Transformer module, which performs hierarchical modeling of regions with different segmentation difficulties in the image based on uncertainty information, achieving differentiated perception and global context modeling, thereby effectively reducing computational redundancy while highlighting the ability to model key regions. Finally, a dual-branch decoder performs parallel decoding and fusion of the multi-level features output by the hierarchical uncertainty Transformer module to generate the final segmentation result, achieving fine segmentation of lesions in medical images.

[0024] Meanwhile, this invention can effectively overcome the challenges commonly faced by existing deep learning models in medical image segmentation tasks: on the one hand, through the context-aware compensation module, the model's feature extraction and fusion capabilities in the encoding stage are enhanced, effectively enriching the local detail feature representation of key regions; on the other hand, through the hierarchical uncertainty Transformer module, the model's global context modeling capability is improved, and differentiated perception of regions with different segmentation difficulties is achieved; in addition, with the help of the dual-branch decoder, the model can achieve effective decoding and reconstruction of multi-layer features, thereby improving the overall accuracy and stability of medical image segmentation. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following is a brief introduction to the drawings used in the prior art and embodiments. The following drawings are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 is a flowchart illustrating a medical image segmentation method based on differential perception according to the present invention. Figure 2 is a diagram of the overall architecture of a medical image segmentation method based on differential perception according to the present invention. Figure 3 is a schematic diagram of the CAC module structure of a medical image segmentation method based on differential perception according to the present invention; Figure 4 is a schematic diagram of the SUTrans module structure of a medical image segmentation method based on differential perception according to the present invention. Detailed Implementation

[0027] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0028] This application provides a medical image segmentation method and apparatus based on differential perception, which is used to effectively extract feature information of lesions in medical images, thereby achieving accurate segmentation of lesions.

[0029] As a typical embodiment of the present invention, the Kvasir-SEG polyp segmentation dataset is used as an example for illustration, referring to... Figure 1 , Figure 2 The method includes the following steps: S1: Obtain the Kvasir-SEG polyp segmentation dataset and perform preprocessing operations on the dataset; S2: Construct a ResNet50-based encoding model to generate uncertainty maps and key region awareness cues; Furthermore, S2 includes the following steps: S2.1: Construct an encoding model consisting of a ResNet50-based encoder and a segmentation module. The segmentation module comprises two... The module consists of a sigmoid activation function; S2.2: Transfer the dataset Input the encoding model to obtain the predicted probability map , predict probability map Binarization provides key area perception prompts. ; S2.3: Using information entropy as an uncertainty measure, the prediction probability map is analyzed. Various spatial locations The uncertainty of the prediction is estimated to obtain the uncertainty diagram. ; S3: Construct a context-aware compensation module (CAC) to enhance, correct, and compensate encoder features, providing higher-quality representations for subsequent decoding stages; Furthermore, S3 includes the following steps: S3.1: Construct the Context-Aware Compensation (CAC) module, which consists of... Module, Module, Spatial Attention Module Spatial Attention and Channel Attention Module (ChannelAttention) is composed of; S3.2: Extract the features from the encoder With key area perception prompts Inputting the data into the CAC module performs feature enhancement, correction, and compensation to obtain enhanced features. ; S4: Construct a hierarchical uncertainty Transformer module SUTrans, which explicitly allocates attention intensity according to the difference in segmentation difficulty of regions, so as to achieve key modeling of critical regions and effectively reduce redundancy; Furthermore, S4 includes the following steps: S4.1: Construct the Stratified UncertaintyTransformer module SUTrans. S4.2: Uncertainty diagram Enhanced features and features extracted by the previous SUTrans module The SUTrans module is used as a common input to perform differentiated global modeling, thereby obtaining differentiated features of interest. ; S5: Construct a dual-branch decoder to decode shallow features and deep semantic information respectively, and generate the final segmentation result through feature fusion; Furthermore, S5 includes the following steps: S5.1: Construct a dual-branch decoder. The features extracted by the two shallow SUTrans modules are input into the shallow decoder to obtain... The features extracted by the two deep SUTrans modules are input into the deep decoder to obtain... ; S5.2: Construct a feature fusion module to integrate the output features of the dual-branch decoder. and Input the feature fusion module to obtain the final segmentation result. .

[0030] Furthermore, S1 includes the following steps: S1.1: Obtain the Kvasir-SEG polyp segmentation dataset and divide the dataset into training and test sets in an 8:2 ratio; S1.2: Randomly shuffle the image samples in the training set to eliminate the influence of sample order on the model training process; S1.3: Adjust the images and their corresponding labeled masks in the training and test sets to a uniform size of 256×256 to ensure that different samples have a consistent resolution when inputting into the model; S1.4: Perform random data augmentation operations on the medical images in the training set, including random angle rotation, horizontal and vertical flipping, and random region occlusion.

[0031] Furthermore, S2.2 includes the following steps: By using a convolutional neural network to perform convolution calculations, the structure of the encoding model is obtained:

[0032] In the above formula, This represents a probability prediction graph. Representative based on residual network, Represents the sigmoid activation function. It consists of three consecutive operations, consisting of a convolutional layer, batch normalization, and ReLU activation function; Using a binarization method, for Threshold segmentation is performed to obtain key area perception cues. :

[0033] In the above formula, The preset threshold is preferred. , This represents the set of spatial pixels in the predicted probability map; S2.3 includes the following steps: Using information entropy calculation, for Uncertainty estimation is performed to obtain the uncertainty diagram. :

[0034] In the above formula, Represents pixels Belongs to the first Predicted probability of class For the number of categories, A small constant to prevent numerical underflow.

[0035] Furthermore, refer to Figure 3 S3.2 includes the following steps: Using a convolutional neural network to perform convolution calculations, the structure of CAC is obtained:

[0036]

[0037]

[0038] In the above formula, Represents the features extracted by the encoder. This represents the output characteristics of the CAC module. The hole rate r of the convolutional layer is represented by Module, , The result of the intermediate module, Represents the spatial attention module. Represents the channel attention module. This represents a channel-level splicing operation. This represents the ReLU activation function.

[0039] Furthermore, refer to Figure 4 S4.2 includes the following steps: For uncertainty diagram Perform statistical analysis and calculate its mean. Standard deviation and median This characterizes the overall confidence distribution of the prediction results. Based on this, high and low percentile thresholds are dynamically determined:

[0040] In the above formula, and These represent the first and second digits of the entropy distribution. and Percentiles. Since the encoding and CAC module processing have enabled the model to accurately identify most regions, the initial settings... And adaptively fine-tuned based on global statistical characteristics: when or A higher confidence level indicates a lower overall confidence level for the model, so it should be appropriately reduced. and To expand the scope of the high uncertainty region; conversely, to increase the threshold to further reduce redundant calculations; when When it is large, it will decrease at the same time. and This is to cover more potential misclassifications. Ultimately, the uncertainty region is divided into three levels based on the entropy value range:

[0041] In the above formula, This is a multi-level uncertainty map, corresponding to high, medium, and low uncertainty regions. The computational mechanism adaptively captures the dynamic changes in entropy distribution under different statistical characteristics, thereby achieving fine-grained uncertainty modeling and regionally differentiated representation in the spatial dimension. Multi-level uncertainty diagram Enhanced features and features extracted by the previous SUTrans module After processing, the common input is the hierarchical uncertainty Transformer module SUTrans, resulting in the SUTrans structure:

[0042] In the above formula, This indicates the differentiated features of interest extracted by SUTrans. This represents a channel-level splicing operation. This indicates an upsampling operation; where, a multi-level uncertainty graph is used. The attention weight matrix is ​​used to extract the multi-head self-attention mechanism in the weighted SUTrans module, thereby achieving differentiated attention.

[0043] Furthermore, S5.1 includes the following steps: A convolutional neural network is used to perform convolution calculations, resulting in the decoder structure:

[0044]

[0045] In the above formula, , and , These represent the differential interest features extracted by the two shallow SUTrans modules and the two deep SUTrans modules, respectively. This represents the features extracted by the shallow decoder. This represents the features extracted by the deep decoder; S5.2 includes the following steps: A convolutional neural network is used to perform convolution calculations, resulting in the feature fusion module structure:

[0046] In the above formula, This indicates the final segmentation result.

[0047] The above is a detailed description of the implementation of the present invention. However, the present invention is not limited to the described embodiments. Those skilled in the art should know that although the present invention has been described according to the above specific embodiments, the inventive concept of the present invention is not limited to this invention. Any modification that uses the inventive concept will be included within the scope of protection of this patent.

Claims

1. A medical image segmentation method and apparatus based on differential perception, characterized in that, The method includes the following steps: S1: Obtain the medical image dataset and perform preprocessing operations on the dataset; S2: Construct a ResNet50-based encoding model to generate uncertainty maps and key region awareness cues; S2 includes the following steps: S2.1: Construct an encoding model consisting of a ResNet50-based encoder and a segmentation module. The segmentation module comprises two... The module consists of a sigmoid activation function; S2.2: Transfer the dataset Input the encoding model to obtain the predicted probability map , predict probability map Binarization provides key area perception prompts. ; S2.3: Using information entropy as an uncertainty measure, the prediction probability map is analyzed. Various spatial locations The uncertainty of the prediction is estimated, and the uncertainty diagram is obtained. ; S3: Construct a context-aware compensation module (CAC) to enhance, correct, and compensate encoder features, providing higher-quality representations for subsequent decoding stages; S3 includes the following steps: S3.1: Construct the Context-Aware Compensation (CAC) module, which consists of... Module, Module, Spatial Attention Module Spatial Attention and Channel Attention Module (ChannelAttention) is composed of; S3.2: Extract the features from the encoder With key area perception prompts Inputting the data into the CAC module performs feature enhancement, correction, and compensation to obtain enhanced features. ; S4: Construct a hierarchical uncertainty Transformer module SUTrans, which explicitly allocates attention intensity according to the difference in segmentation difficulty of regions, so as to achieve key modeling of critical regions and effectively reduce redundancy; S4 includes the following steps: S4.1: Construct the Stratified UncertaintyTransformer module SUTrans. S4.2: Uncertainty diagram Enhanced features and features extracted by the previous SUTrans module The SUTrans module is used as a common input to perform differentiated global modeling, thereby obtaining differentiated features of interest. ; S5: Construct a dual-branch decoder to decode shallow features and deep semantic information respectively, and generate the final segmentation result through feature fusion; S5 includes the following steps: S5.1: Construct a dual-branch decoder. The features extracted by the two shallow SUTrans modules are input into the shallow decoder to obtain... The features extracted by the two deep SUTrans modules are input into the deep decoder to obtain... ; S5.2: Construct a feature fusion module to integrate the output features of the dual-branch decoder. and Input the feature fusion module to obtain the final segmentation result. .

2. The medical image segmentation method and apparatus based on differential perception according to claim 1, characterized in that, S1 includes the following steps: S1.1: Obtain the medical image segmentation dataset and divide it into a training set and a test set in an 8:2 ratio; S1.2: Randomly shuffle the image samples in the training set to eliminate the influence of sample order on the model training process; S1.3: Adjust the images and their corresponding labeled masks in the training and test sets to a uniform size of 256×256 to ensure that different samples have a consistent resolution when inputting into the model; S1.4: Perform random data augmentation operations on the medical images in the training set, including random angle rotation, horizontal and vertical flipping, and random region occlusion.

3. The medical image segmentation method and apparatus based on differential perception according to claim 1, characterized in that, S2.2 includes the following steps: By using a convolutional neural network to perform convolution calculations, the structure of the encoding model is obtained: In the above formula, This represents a probability prediction graph. Representative based on residual network, Represents the sigmoid activation function. It consists of three consecutive operations, consisting of a convolutional layer, batch normalization, and ReLU activation function; Using a binarization method, for Threshold segmentation is performed to obtain key area perception cues. : In the above formula, The preset threshold is preferred. , This represents the set of spatial pixels in the predicted probability map; S2.3 includes the following steps: Using information entropy calculation, for Uncertainty estimation is performed to obtain the uncertainty diagram. : In the above formula, Represents pixels Belongs to the first Predicted probability of class For the number of categories, A small constant to prevent numerical underflow.

4. The medical image segmentation method and apparatus based on differential perception according to claim 1, characterized in that, S3.2 includes the following steps: Using a convolutional neural network to perform convolution calculations, the structure of CAC is obtained: In the above formula, Represents the features extracted by the encoder. This represents the output characteristics of the CAC module. The hole rate r of the convolutional layer is represented by Module, , The result of the intermediate module, Represents the spatial attention module. Represents the channel attention module. This represents a channel-level splicing operation. This represents the ReLU activation function.

5. The medical image segmentation method and apparatus based on differential perception according to claim 1, characterized in that, S4.2 includes the following steps: For uncertainty diagrams Perform statistical analysis and calculate its mean. Standard deviation and median This characterizes the overall confidence distribution of the prediction results; based on this, the high and low percentile thresholds are dynamically determined. In the above formula, and These represent the first and second digits of the entropy distribution. and Percentiles; since the encoding and CAC module processing have enabled the model to accurately identify most regions, the initial setting is... And adaptively fine-tuned based on global statistical characteristics: when or A higher confidence level indicates a lower overall confidence level for the model, so it should be appropriately reduced. and To expand the scope of the high uncertainty region; conversely, to increase the threshold to further reduce redundant calculations; when When it is large, it will decrease at the same time. and This is to cover more potential misclassifications; ultimately, the uncertainty region is divided into three levels based on the entropy range: In the above formula, It is a multi-level uncertainty map, corresponding to high, medium and low uncertainty regions respectively; the computing mechanism can adaptively capture the dynamic changes of entropy distribution under different statistical characteristics, thereby realizing fine-grained uncertainty modeling and regional differentiation expression in the spatial dimension; Multi-level uncertainty diagram Enhanced features and features extracted by the previous SUTrans module After processing, the common input is the hierarchical uncertainty Transformer module SUTrans, resulting in the SUTrans structure: In the above formula, This indicates the differentiated features of interest extracted by SUTrans. This represents a channel-level splicing operation. This indicates an upsampling operation; where, a multi-level uncertainty graph is used. The attention weight matrix is ​​used to extract the multi-head self-attention mechanism in the weighted SUTrans module, thereby achieving differentiated attention.

6. The medical image segmentation method and apparatus based on differential perception according to claim 1, characterized in that, S5.1 includes the following steps: A convolutional neural network is used to perform convolution calculations, resulting in the decoder structure: In the above formula, , and , These represent the differential interest features extracted by the two shallow SUTrans modules and the two deep SUTrans modules, respectively. This represents the features extracted by the shallow decoder. This represents the features extracted by the deep decoder; S5.2 includes the following steps: A convolutional neural network is used to perform convolution calculations, resulting in the feature fusion module structure: In the above formula, This indicates the final segmentation result.

Images