A lesion classification system based on deep learning and probabilistic radiomics

Abstract

The invention relates to a lesion classification system based on deep learning and probabilistic radiomics, belonging to the technical field of medical image classification. Aiming at the problems of ambiguity and low classification accuracy caused by the ambiguity of the existing lesion classification system, the present invention uses the deep convolutional neural network as the backbone, and proposes a non-local shape analysis module to extract the feature cloud of the lesion on the medical image. , remove the interference of the surrounding pixels of the lesion on the classification judgment, and obtain the essential representation of the lesion; at the same time, in order to capture the ambiguity of the label, a fuzzy prior network is proposed to simulate the fuzzy distribution of different expert labels, showing that the ambiguity of the expert label is modeled, The classification results of model training are more robust, and a new lesion classification system is constructed by combining fuzzy prior samples with lesion representation, which is controllable and probabilistic, compared with traditional convolutional neural networks. The network can better solve the problem of classification ambiguity and obtain higher classification accuracy.

Description

technical field [0001] The invention relates to the classification technology of CT images of lesions, in particular to a lesion classification system based on deep learning and probabilistic radiomics, and belongs to the technical field of medical image classification. Background technique [0002] At present, the morbidity and mortality of various cancers in China rank first among all kinds of diseases, among which the morbidity and mortality of lung cancer rank first among various malignant tumors, posing a huge threat to people's health. Pulmonary nodule screening is an important means to achieve early diagnosis and treatment of lung cancer. The radiomics analysis method uses doctors to manually outline regions of interest, image processing, feature extraction, and feature screening, and then combines machine learning algorithms to predict target variables and assist doctors to analyze lesions. It has powerful functions and has been widely used. Compared with traditiona...

AI Technical Summary

Problems solved by technology

However, first of all, the method adopted by this patent is based on a 2D convolutional neural network, which makes this patent not capable of capturing three-dimensional spatial features

Secondly, the patent uses traditional black-box deep learning, which makes the patented technology not controllable and transparent

There is also the problem of ambiguity caused by classification fuzziness.

[0010] Therefore, it is urgent to propose a new classification system to solve the problem of ambiguity and insufficient classification accuracy caused by the ambiguity of classification of lesions in the existing classification technology.

Images