Fully Automated Regional Analysis of Myocardial T2* Values for Iron Quantification Using Deep Learning

Abstract

Cardiovascular magnetic resonance (CMR) T2* mapping is the gold standard technique for the assessment of iron overload in the heart. The quantitative analysis of T2* values requires the manual segmentation of T2* images, which is a time-consuming and operator-dependent procedure. This study describes a fully-automated method for the regional analysis of myocardial T2* distribution using a deep convolutional neural network (CNN). A CNN with U-Net architecture was trained to segment multi-echo T2*-weighted images in 16 sectors in accordance with the American Heart Association (AHA) model. We used images from 210 patients (three slices, 10 multi-echo images) with iron overload diseases to train and test the CNN. The performance of the proposed method was quantitatively evaluated on an independent holdout test set by comparing the segmentation accuracy of the CNN and the T2* values obtained by the automated method against ground-truth labels provided by two experts. Segmentation metrics and global and regional T2* values assessed by the proposed DL method closely matched those obtained by experts with excellent intraclass correlation in all myocardial sectors of the AHA model (ICC range [0.944, 0.996]). This method could be effectively adopted in the clinical setting for fast and accurate analysis of myocardial T2*.