Medical image classification method based on deep multi-instance learning and self-attention

Abstract

The invention discloses a medical image classification method based on deep multi-instance learning and self-attention, and relates to the medical image processing technology, and the method comprisesthe steps of carrying out the preprocessing of a medical image, and generating an instance package in multi-instance learning; extracting example features in the packet by using a convolutional neural network; constructing a feature extraction module based on a self-attention mechanism, and learning a dependency relationship between examples; aggregating the example features in the packet by using a feature pooling module to obtain a packet-level feature vector; and taking the packet-level feature vector as the input of a classifier, and generating a prediction mark of the input image. According to the invention, the local detail features of the to-be-recognized image are captured through the convolutional network, the global structure features of the to-be-recognized image are learned through the self-attention mechanism, the two features supplement each other, the classification performance and stability of the whole network are improved, and in addition, by introducing the trainable pooling operator, the interpretability of the network is further enhanced.

Description

Technical field[0001]The present invention relates to medical image processing techniques, and more particularly to a medical image classification method based on depth multi-example learning and self-focus.Background technique[0002]In recent years, deep learning methods have far beyond the traditional shallow machine learning methods in many artificial intelligence, and have a wide and profound impact on academic and industrial communities. In the field classification, deep supervision learning methods have achieved unprecedented success, exhibiting unprecedented identification and classification capabilities in many large image classification tasks, one of the key drivers is a large number of sample data with exactly labeled. However, in practical applications, a large amount of detailed labeling is often extremely difficult, for example, in medical image analysis, the specific location of the lesion is often cost-effective, and only the overall annotation information of the image...

AI Technical Summary

Problems solved by technology

[0004] The present invention proposes a medical image classification method based on deep multi-instance learning and self-attention, which solves the problem of low classification performance caused by ignoring the global structure information of the image in the existing method, and provides a new method for medical image classification A solution based on deep multiple-instance learning networks

Images