Information-based meat segmentation method and system based on attention decoding

Abstract

The present application belongs to the technical field of medical image segmentation, and the method comprises the following steps: preprocessing a data set, constructing a reliable polyp segmentation model architecture based on attention decoding; using a combination of cross-entropy loss function, KL divergence loss function and Dice loss function to calculate the loss function of the constructed polyp segmentation architecture model; using the Adam optimization algorithm to train the constructed polyp segmentation architecture model; using the DiceScore and mIou indexes to measure the segmentation accuracy. The present application applies evidence deep learning to an end-to-end polyp segmentation model, aims to quantify the pixel-level uncertainty, and obtains a reliable polyp segmentation model. Meanwhile, when processing polyps with fuzzy boundaries and small sizes, the present application adopts an edge-guided feature module to mine the edge semantics related to the boundaries, so that better segmentation results can be obtained.

Claims

1. An informational meat segmentation method based on attention decoding, characterized in that: include, The dataset is preprocessed to construct a reliable polyp segmentation model architecture based on attention decoding; The loss function of the constructed polyp segmentation architecture model is calculated by combining the cross-entropy loss function, the KL divergence loss function, and the Dice loss function; A polyp segmentation architecture model was constructed using the Adam optimization algorithm. The DiceScore and mIou metrics were used to measure segmentation accuracy. The architecture for constructing a reliable polyp segmentation model based on attention decoding includes an encoder module 100, a decoder module 200, and an uncertainty estimation module 300. The encoder module 100 uses PVTv2 as an encoder for multi-level feature mapping. The input data is processed by the PVTv2 encoder to generate four stages of features X1 to X4. X1 contains detailed texture information of the target, while X2, X3 and X4 contain high-level semantic information. The decoder module 200, the cascaded attention decoding module, consists of three parts: an attention gate module for cascaded feature fusion, an edge-guided feature module for enhancing boundary representation, and a convolutional attention module for robustly enhancing the feature map. The features X1~X4 obtained by the encoder module are processed by the decoder module to gradually extract key information and suppress background information. The uncertainty estimation module 300 takes the result obtained by the decoder module and passes it through a non-negative activation function to obtain evidence output, and then constructs the probability and uncertainty of different categories through the Dirichlet distribution; The decoder module 200 includes a convolutional attention module 201, an attention gate module 202, and an edge-guided feature module 203; The convolutional attention module 201 includes a channel attention module, a spatial attention module, and a convolutional block. The channel attention module determines the channels that need to be focused on, the spatial attention module determines the positions that need to be focused on in the feature map, and the convolutional block is used to further enhance the features through the channel attention and spatial attention modules. The features of the four stages X1 to X4 generated by the encoder module are decoded: the X4 features of high-level semantic information are convolved by 1×1, the number of channels is adjusted, and the input is fed into the convolutional attention module 201 to refine the feature mapping. After the above operation, one path of the features is upsampled by 32 times to obtain the final segmentation result, and the other path is upsampled by 2 times and fused with the X3 stage features generated by the encoder module through the attention gate module 202. After feature fusion, the features are then input into the edge-guided feature module 203, which uses a local channel attention mechanism to multiply with the initial fused features and reduce the number of channels to enhance the boundary feature representation. Among them, the edge-guided feature module 203 provides input features. and the upsampled features of the previous layer First, element-wise multiplication is performed between them, followed by skip connections and 3×3 convolutions to obtain the initial fused features. To enhance feature representation, a channel-level global average pooling is used to aggregate convolutional features. Channel attention is obtained through one-dimensional convolution and the sigmoid function, and this channel attention is then integrated with the input features. Multiply the results and reduce the number of channels using a 1×1 convolution to obtain the final output. : in, For the final output feature fusion, For 1×1 convolution, For the Sigmoid function, The process involves a 1D convolution with kernel k, followed by a convolutional attention module 201. One path upsamples by 16 times to obtain the final segmentation result, while the other path continues to perform the above operation with lower-level features, resulting in four stages of output. The uncertainty estimation module 300 includes a credible segmentation framework defined by subjective logic theory, which is used to derive the probability and uncertainty of the segmentation problem: the output results of the four stages obtained by the decoder module 200 are processed by the non-negative activation function softplus to obtain four stages of evidence output. The output result of each stage is a Dirichlet distribution and is regarded as the conjugate prior of a multinomial distribution. Then, subjective logic theory combines the evidence with parameters that have a Dirichlet distribution. Connecting them: Uncertainty is derived from this distribution, and the belief quality and uncertainty of each pixel are represented: The more type of evidence a pixel (i, j) receives, the higher its probability; conversely, the greater the uncertainty of a pixel (i, j). For Dirichlet parameters, For the purpose of providing evidence, For the sake of belief quality, It is uncertain.

2. The information-based meat segmentation method based on attention decoding as described in claim 1, characterized in that: The calculated loss function includes the results of the constructed polyp segmentation architecture model. First, the predicted probabilities and image labels are used to calculate the cross-entropy loss. Since the predicted probabilities follow a Dirichlet distribution, within the framework of evidence theory, the Dirichlet distribution is linked to the belief distribution. Based on the evidence collected from the backbone, the probabilities of different categories and the uncertainty of different voxels are obtained to derive an improved cross-entropy loss function. Regularization techniques are used to introduce KL divergence loss, which reduces the evidence generated by incorrect labels, even to 0, ensuring that the true category is not considered 0. Furthermore, by using a Dice loss to optimize the network, the overall loss of the network is the sum of the three losses mentioned above: in, , The balancing factors are set to 0.2 and 0.

1. For digital matrix functions, Let cross-entropy be the loss function. Let KL divergence loss function be used. The Dice loss function, The correction parameters for the Dirichlet distribution are... This is a gamma function.

3. The information-based meat segmentation method based on attention decoding as described in claim 2, characterized in that: The polyp segmentation architecture model constructed by the Adam optimization algorithm includes training the network using the network structure constructed by the polyp segmentation architecture model and the loss function trained by the optimization algorithm, and using a multi-scale training strategy. The learning rate of the Adam optimization algorithm is set to 1e-4, the batch size is set to 16, and the number of iterations is set to 100.

4. The information-rich meat segmentation method based on attention decoding as described in claim 3, characterized in that: The DiceScore and mIou metrics are derived by inputting test data into a network trained using the Adam optimization algorithm to obtain the segmentation results and uncertainties for each image. The DiceScore and mIou metrics are then used to measure the segmentation accuracy, as specifically defined in the following formula: In this diagram, A represents the polyp lesion area marked by the doctor, and B represents the segmented polyp lesion area. Let be the loU of the i-th test image.

5. An information-segmenting system based on attention decoding as described in any one of claims 1 to 4, characterized in that: It includes a data preprocessing module, a model architecture construction module, a loss function statistics module, an optimization algorithm training module, and a segmentation performance evaluation module; The data preprocessing module uses five public datasets to convert the raw image data into a format acceptable to the model and divides the datasets into training and prediction sets. The model architecture building module constructs a meat segmentation model based on attention decoding after preprocessing, and constructs the probabilities and uncertainties of different categories; The loss function statistics module calculates the overall loss function of the network, helping to find the optimal parameters during model training and making the model's prediction results more accurate. The optimization algorithm training module uses optimization algorithms to update the model and employs a multi-scale training strategy to make the model's prediction results more accurate. The segmentation performance evaluation module evaluates the segmentation performance of the model, determines whether the model has achieved the expected results, and identifies the model's advantages, disadvantages, and directions for improvement.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images