A device for computer-aided pulmonary nodule classification based on transferable multi-model ensemble

Abstract

The invention discloses a computer-assisted pulmonary nodule classification device based on transferable multi-model integration, which is used to solve the technical problem of poor classification accuracy of the existing pulmonary nodule. The technical solution is to train three pre-trained deep convolutional neural networks (Pre-trained DCNN), respectively describe the heterogeneity of lung nodule texture, shape heterogeneity and global features, and perform the results of the deep convolutional neural network Weighted average, the weight of each network is adaptively learned through the error backpropagation mechanism, thereby improving the accuracy of pulmonary nodule classification, in which the Pre-trained DCNN combines the good images of the deep convolutional neural network on large data sets Representation ability transfer pulmonary nodule classification task effectively improves the training performance of deep convolutional neural network on small pulmonary nodule data. The invention overcomes the technical problem of low classification accuracy based on single information, and the accuracy reaches 93%.

Description

technical field [0001] The invention relates to a pulmonary nodule classification device, in particular to a computer-aided pulmonary nodule classification device based on transferable multi-model integration. Background technique [0002] The traditional benign and malignant classification technology of pulmonary nodules based on CT images can generally be divided into three parts: pulmonary nodule segmentation, feature extraction and benign and malignant classification of pulmonary nodules. The above methods rely on the pre-segmentation of pulmonary nodules. Many current segmentation methods rely on the initialization of algorithms, such as region growing algorithms, level set algorithms, etc. Different initializations will have different effects on the final segmentation results, therefore, the features obtained by using such segmentation results are usually inaccurate. [0003] The document "Application Publication No. CN 104700118 A Chinese Invention Patent" discloses...

AI Technical Summary

Problems solved by technology

[0005] In order to overcome the deficiency of poor accuracy rate of existing pulmonary nodule classification methods, the present invention provides a computer-aided pulmonary nodule classification device based on transferable multi-model integration

Images