Generalization of Left Ventricular Segmentation Models to LVNC Patients: A Comparative Study

Left ventricular non-compaction (LVNC) is a rare cardiomyopathy with distinctive myocardial morphology. Due to its low prevalence, LVNC cases are rarely included in public cardiac datasets, hindering the development of specialized deep learning segmentation models. This study aims to systematically evaluate the generalization capability of state-of-the-art models, trained on public, multi-disease cardiac datasets, to the challenging LVNC cohort. Beyond performance benchmarking, we investigate the impact of critical data factors, including training set size and pathology composition, to derive actionable insights for building effective models in data-scarce, rare disease scenarios. We benchmarked state-of-the-art segmentation models from four architectural categories (CNN-based, Transformer-based, Mamba-based, and pre-trained foundation models). Trained on the M&M dataset, the models were tested on an independent LVNC dataset, with performance evaluated using Dice score, mIoU, ASD, 95% HD, and EF estimation error. Additional experiments assessed the effects of training data size and pathology composition. STU-Net, nnU-Net, and U-Mamba-Bot showed strong generalization to the LVNC dataset, achieving LV Dice scores of  91%, MYO Dice scores of  80%, and EF estimation errors as low as 4.6% MAE and 6.7% RMSE. Notably, U-Mamba-Bot exhibited the best robustness under limited data and unbalanced pathology conditions. While diverse disease types in training data generally improved performance, an excessive proportion of HCM cases reduced segmentation accuracy. With sufficient multi-pathology training data, state-of-the-art segmentation models can effectively generalize to rare diseases like LVNC. Under limited-data scenarios, careful selection and balance of pathology types are critical to ensuring robust model performance.