CN-122023947-A - Pathological image classification method, electronic device, and readable storage medium

CN122023947ACN 122023947 ACN122023947 ACN 122023947ACN-122023947-A

Abstract

The embodiment of the application provides a pathological image classification method, electronic equipment and a readable storage medium, and belongs to the technical fields of medical image processing and artificial intelligence. Taking a pathological image block as a node, extracting node characteristics, carrying out pathological classification on the node characteristics through a lightweight branch network to obtain preliminary node classification probability, screening pathological nodes from the nodes according to the preliminary node classification probability, calculating the node attention score of each node, searching K neighbor nodes according to the preliminary node classification probability to create edges between the neighbors and the pathological nodes, generating edge weights based on the node characteristics of the neighbors and the pathological nodes to construct a target pathological node diagram, carrying out perception aggregation on the target pathological node diagram, and carrying out global pathological classification on the target pathological image according to updated node characteristics obtained by aggregation. The embodiment of the application can reduce the calculation cost on the premise of not losing the classification precision of the pathological image so as to improve the classification efficiency of the pathological image.

Inventors

FANG SHENYING
ZENG GUANGJIAN

Assignees

南方科技大学

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. A method of classifying pathological images, the method comprising: acquiring a target pathological image, dividing the target pathological image to obtain a plurality of pathological image blocks, and regarding each pathological image block as a node; Extracting the characteristics of each node through a pre-trained target pathological image classification model to obtain the node characteristics corresponding to each node; Performing pathological classification on the node characteristics through a lightweight branch network to obtain preliminary node classification probability of each node; node screening is carried out on each node according to the preliminary node classification probability, and pathological nodes are obtained; Performing attention processing on node characteristics of the pathological node through a weight branch network to obtain a node attention score; Performing neighbor searching on each pathological node according to the node attention score through a weight branch network to obtain K neighbor nodes of the pathological node, creating edges between the neighbor nodes and the pathological node, and generating weights of the edges according to node characteristics of the neighbor nodes and node characteristics of the pathological node to construct a target pathological node diagram; Performing perception aggregation on the target pathological node diagram through a weight branch network to obtain updated node characteristics of the pathological nodes; And carrying out global pathology classification on the target pathology image according to the updated node characteristics of each pathology node to obtain target pathology classification data.
2. The method according to claim 1, wherein the node screening of each node according to the preliminary node classification probability to obtain a pathological node comprises: performing normal approximate processing according to the preliminary node classification probability to obtain the upper limit number of pathological nodes; Screening a first number of nodes with highest probability from the nodes according to the preliminary node classification probability; quasi-bell grouping is carried out on the preliminary node classification probability to obtain a grouping node classification probability; Screening a second number of packet nodes from the nodes except the first number of nodes according to the packet node classification probability; and integrating the first number of nodes and the second number of grouping nodes to obtain the pathological node, wherein the sum of the first number and the second number does not exceed the upper limit number of the pathological node.
3. The method according to claim 2, wherein said performing a normal approximation process according to said preliminary node classification probability to obtain an upper limit number of pathological nodes comprises: determining the total number of pathological nodes judged to be positive in each node and the pathological node classification probability according to the preliminary node classification probability; Determining normal distribution data of the pathological nodes of the total number of the pathological nodes according to the probability mean value and the probability variance of the pathological node classification probability; And acquiring normal confidence critical values of the normal distribution data of the pathological nodes, and calculating according to the probability mean value, the probability variance and the normal confidence critical values to obtain the upper limit number of the pathological nodes.
4. The method according to claim 1, wherein the performing neighbor searching on each pathological node according to the node attention score through the weight branch network to obtain K neighboring nodes of the pathological node includes: sequencing the node attention scores of the pathological nodes from large to small to obtain attention score orders of the pathological nodes; And determining k neighbor nodes of the pathological node according to the difference value between the attention score orders of any two pathological nodes.
5. The method of claim 1, wherein the target pathology classification data comprises a target pathology classification probability and a target classification heat map; The global pathology classification is carried out on the target pathology image according to the updated node characteristics of each pathology node to obtain target pathology classification data, and the method comprises the following steps: Performing node pathology calculation according to the updated node characteristics of each pathology node to obtain a pathology node classification probability; Performing global pathology calculation on the target pathology image according to the pathology node classification probability to obtain the target pathology classification probability; performing attention processing according to the updated node characteristics of each pathological node to obtain updated node attention scores; performing suspicious pathological scoring according to the updated node attention score to obtain node pathological suspicious score; And generating the target classification heat release map according to the node pathology suspicious score and the target pathology image.
6. The method according to claim 1, wherein before the feature extraction is performed on each node by the pre-trained target pathology image classification model, the method further comprises: The method comprises the steps of obtaining a training data set, wherein the training data set comprises a training pathological image and real pathological classification data corresponding to the training pathological image; Dividing the training pathological image to obtain a plurality of training pathological image blocks, and regarding each training pathological image block as a training node; extracting features of the training nodes through a preset original pathological image classification model to obtain training node features corresponding to the training nodes; Performing pathological classification on the characteristics of each training node through a lightweight branch network to obtain the prediction preliminary node classification probability of each training node; performing pathological classification on the training pathological image according to the training preliminary node classification probability through a lightweight branch network to obtain a predicted preliminary pathological classification probability; Node screening is carried out on each training node according to the predicted preliminary node classification probability, and training pathological nodes are obtained; Performing attention processing on the node characteristics of the training pathological node through a weight branch network to obtain the attention score of the training node; Performing neighbor searching on each training pathological node according to the attention scores of the training nodes through a weight branch network to obtain K training neighbor nodes of the training pathological node, creating training edges between the training neighbor nodes and the training pathological node, and generating weights of the training edges according to the node characteristics of the training neighbor nodes and the node characteristics of the pathological node to construct and obtain a training target pathological node diagram; performing perception aggregation on the training target pathological node diagram through a weight branch network to obtain training updated node characteristics of the training pathological node; performing global pathology classification on the training pathology image according to the training update node characteristics of each training pathology node to obtain predicted pathology classification data; calculating the predicted preliminary node classification probability, the predicted preliminary pathology classification probability, the predicted pathology classification data and the target loss value of the real pathology classification data according to a preset loss function; training the original pathological image classification model according to the target loss value and a preset loss condition to obtain the pre-trained target pathological image classification model.
7. The method of claim 6, wherein the predicted pathology classification data comprises a predicted pathology node classification probability, a predicted global classification probability, and a predicted attention classification probability; The global pathology classification is performed on the training pathology image according to the training update node characteristics of each training pathology node to obtain predicted pathology classification data, including: Performing node pathology calculation according to the training updated node characteristics of each training pathology node to obtain the classification probability of the predicted pathology node; performing global pathology calculation on the training pathology image according to the predicted pathology node classification probability to obtain the predicted global classification probability; Performing attention processing according to the training update node characteristics of each training pathological node to obtain the attention score of the training update node; and carrying out attention classification according to the attention score of the training update node to obtain the prediction attention classification probability.
8. The method of claim 7, wherein the true pathology classification data comprises a true preliminary pathology classification label, a true global classification label, and a true attention classification label, wherein the penalty functions comprise a preliminary classification penalty function, a global classification penalty function, an attention global classification penalty function, and a local node penalty function; The calculating the predicted preliminary node classification probability, the predicted preliminary pathology classification probability, the predicted pathology classification data and the target loss value of the real pathology classification data according to a preset loss function comprises the following steps: calculating a preliminary classification loss value between the predicted preliminary node classification probability and the real preliminary pathology classification label according to the preliminary classification loss function; calculating a global classification loss value between the predicted global classification probability and the real global classification label according to the global classification loss function; Calculating an attention classification loss value between the predicted attention classification probability and the true attention classification label according to the attention classification loss function; calculating an attention global classification loss value between the predicted global classification probability and the predicted attention classification probability according to the attention global classification loss function; Calculating a local node loss value between the predicted preliminary node classification probability and the predicted pathological node classification probability according to the local node loss function; And integrating the preliminary classification loss value, the global classification loss value, the attention global classification loss value and the local node loss value to obtain the target loss value.
9. An electronic device comprising a memory storing a computer program and a processor implementing the pathological image classification method according to any one of claims 1 to 8 when the processor executes the computer program.
10. A computer-readable storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the pathology image classification method according to any one of claims 1 to 8.

Description

Pathological image classification method, electronic device, and readable storage medium Technical Field The present application relates to the field of medical image processing and artificial intelligence, and in particular, to a pathological image classification method, an electronic device, and a readable storage medium. Background At present, the traditional pathological image classification method is generally realized based on a multiple example learning method (such as a CLAM model) of attention, specifically, firstly, a pathological full-slice image is divided into a plurality of non-overlapping image blocks, a pre-training convolutional neural network is used for extracting the characteristics of each image block, then, a leavable attention weight is distributed to the characteristics of each image block through an attention network, the weight reflects the importance of the image block to the final pathological image classification, and finally, the weighted characteristics of all the image blocks are aggregated to obtain the characteristic representation of the whole slice, and a pathological classifier is input to carry out pathological image classification. But this approach focuses mainly on the contribution of individual image blocks, ignoring spatial and semantic associations between image blocks, and has difficulty capturing global context information. Therefore, how to capture global context information of pathological images to improve pathological image classification accuracy becomes a challenging research problem. In order to solve the problem, in the prior art, pathological image classification is realized by a multi-example learning (such as WiKG model) method based on a graph neural network, specifically, each pathological image block is firstly regarded as a node in a pathological image, characteristics of the pathological image block are extracted to serve as node initial representation, then, association strength between every two nodes is calculated based on the node characteristics, connection of node edges is dynamically determined, and then, the graph neural network is used for node information propagation and aggregation, and finally, characteristic graph representation is obtained through graph pooling for pathological classification. However, since the prior art method needs to calculate the association between all node pairs (i.e., each image block), the prior art method needs huge memory and calculation overhead, and it is difficult to complete pathological classification within a clinically acceptable time. Therefore, how to reduce the calculation overhead without losing the classification precision of the pathological image is a urgent problem to be solved. Disclosure of Invention The embodiment of the application mainly aims to provide a pathological image classification method, electronic equipment and a readable storage medium, aiming at reducing the calculation cost on the premise of not losing the pathological image classification precision so as to improve the pathological image classification efficiency. To achieve the above object, a first aspect of an embodiment of the present application proposes a pathological image classification method, including: acquiring a target pathological image, dividing the target pathological image to obtain a plurality of pathological image blocks, and regarding each pathological image block as a node; Extracting the characteristics of each node through a pre-trained target pathological image classification model to obtain the node characteristics corresponding to each node; Performing pathological classification on the node characteristics through a lightweight branch network to obtain preliminary node classification probability of each node; node screening is carried out on each node according to the preliminary node classification probability, and pathological nodes are obtained; Performing attention processing on node characteristics of the pathological node through a weight branch network to obtain a node attention score; Performing neighbor searching on each pathological node according to the node attention score through a weight branch network to obtain K neighbor nodes of the pathological node, creating edges between the neighbor nodes and the pathological node, and generating weights of the edges according to node characteristics of the neighbor nodes and node characteristics of the pathological node to construct a target pathological node diagram; Performing perception aggregation on the target pathological node diagram through a weight branch network to obtain updated node characteristics of the pathological nodes; And carrying out global pathology classification on the target pathology image according to the updated node characteristics of each pathology node to obtain target pathology classification data. In some embodiments, the node screening is performed on each node according to the preliminary node classificatio