Multi-view fusion network with test time augmentation for cardiac image segmentation and ejection fraction estimation

Accurate ejection fraction (EF) estimation is critical for diagnosing and managing heart failure. EF is typically calculated using the biplane method of disks, which relies on precise segmentation of the left ventricle from 2-chamber and 4-chamber views at the same cardiac phase. However, despite the strong correlation between these views, variations in cardiac motion can lead to inconsistent segmentation performance when models are trained separately on each view. This results in suboptimal data utilization, limiting the performance and robustness of automated methods. To address this, we propose Multi-View XNet, a dual-headed encoder-decoder architecture that fuses information from both views using our novel Cross-View Fusion Module. The module regularizes mutual information in the latent space by simultaneously processing both views through separate encoder heads. The fused information is then passed to two decoder heads to output segmented results for each respective view, which are used to compute EF. To further enhance performance, we introduce test-time augmentation (TTA) to improve robustness against scanning variations. Our method outperforms state-of-the-art approaches in segmentation (DICE scores of 0.947 and 0.961 for endocardium and epicardium in end-diastole, and 0.932 and 0.957 in end-systole) and in clinical parameters (correlations of 0.980, 0.982, and 0.916 for EDV, ESV, and EF, respectively, with a mean absolute EF error of 4.7 percent). We also demonstrate generalization on an external dataset, achieving DICE scores of 0.921 and 0.933 for the endocardium and epicardium in ED, and 0.844 and 0.895 in ES. To support reproducibility and future research, our code is publicly available at: https://github.com/waqasanwaar/MV-XNet . • A novel multi-headed model (MV-XNet) is presented for multi-view data. • Cross-view Fusion module (CVFM) is defined to fuse the information from multiple views. • Test time augmentation is introduced to combat the variation during data acquisition. • Validated efficacy for clinical metric estimation.