Characterization of non-calcified predominant plaque using deep learning and radiomics analyses of coronary computed tomography angiography images

Abstract

Background

To use an automated system exploiting the advantages of both a neural network and radiomics for analysis of non-calcified predominant plaque (NCPP).

Methods

This study retrospectively included 234 patients. Using the workflow of the previous study, the coronary artery was first segmented, images containing plaques were then extracted, and a classifier was built to identify non-calcified predominant plaques. Radiomics feature analysis and a visualization tool were used to better distinguish NCPP from other plaques.

Results

Twenty-six representative radiomics features were selected. DenseNet achieved an area under the curve of 0.889, which was significantly larger (p = 0.001) than that obtained using a gradient-boosted decision tree (0.859). The feature variances and energy features in calcified predominant plaque were both different from those in NCPP.

Conclusions

Our automated system provided high-accuracy analysis of vulnerable plaques using a deep learning approach and predicted useful features of NCPP using a radiomics-based approach.