A Novel Approach Based on Spatio-temporal Features and Random Forest for Scar Detection Using Cine Cardiac Magnetic Resonance Images

Abstract

Aim. To identify the presence of scar tissue in the left ventricle from Gadolinium (Gd)-free magnetic resonance cine sequences using a learning-based approach relying on spatiotemporal features. Methods. The spatial and temporal features were extracted using local binary patterns from (i) cine end-diastolic frame and (ii) two parametric images of amplitude and phase wall motion, respectively, and classified with Random Forest. Results. When tested on 328 cine sequences from 40 patients, a recall of 70% was achieved, improving significantly the classification resulting from spatial and temporal features processed separately. Conclusions. The proposed approach showed promising results, paving the way for scar identification from Gd-free images.