Impact of Class Labeling on Myocardium Segmentation using Cascaded Deep Learning for Improved Myocardial Infarction Segmentation

Myocardial infarction is a leading cause of death and illness globally which necessitates early detection and characterization of the infarcted tissue for treatment planning and prognosis. Accurate and fast segmentation of this tissue is crucial for quantifying disease severity. CNNs have shown promise in this area. However, the non-differentiable contrast in Late Gadolinium Enhancement (LGE) MRI poses a challenge for infarcted tissue detection. Additionally, network architecture significantly impacts segmentation performance. Our study proposes a cascaded Deep Learning approach for automatic myocardial infarction segmentation in short-axis LGE cardiac magnetic resonance imaging (CMR). We investigate two cascaded network models (Q and R) that differ in class labeling approaches; single and multi-class labeling. These models specifically search for infarcted tissue within the myocardium, effectively restricting the search area. Notably, Model R achieves slightly better performance with median Dice Score different about 3.51% for infarcted tissue segmentation in the second stage compared to model Q.