Optimizing Cardiac MRI Segmentation: An Ensemble Approach with U-Net Variants

Abstract

Segmentation of cardiac magnetic resonance images is a critical task in medical imaging, particularly to delineate the left and right ventricles and the myocardium. This study aims to improve segmentation performance using an ensemble approach with variants of the U-Net architecture, a widely adopted deep learning model for image segmentation. Multiple segmentation models were trained and optimized, and their outputs were combined using threshold-based binary conversion. Two ensemble strategies were evaluated: (1) Averaging, where the mean value of the binary masks at each pixel location was calculated to smooth discrepancies among model predictions, and (2) Voting, where majority voting determined the final pixel classification. The proposed ensemble approach demonstrates robustness to individual model errors and improves segmentation consistency.