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

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Left Ventricular Non-compaction (LVNC) is a rare cardiomyopathy with limited representation in public cardiac datasets.
• This scarcity impedes the development of specialized deep learning segmentation models for LVNC.

Purpose of the Study:

• To evaluate the generalization capability of state-of-the-art deep learning models on an LVNC cohort.
• To investigate the impact of training data size and pathology composition on model performance for rare diseases.

Main Methods:

• Benchmarked CNN-based, Transformer-based, Mamba-based, and pre-trained foundation models.
• Trained models on the M&M dataset and tested on an independent LVNC dataset.
• Evaluated performance using Dice score, mIoU, ASD, 95% HD, and EF estimation error.

Main Results:

• STU-Net, nnU-Net, and U-Mamba-Bot demonstrated strong generalization to LVNC.
• Achieved LV Dice scores up to 91%, MYO Dice scores up to 80%, and EF estimation errors as low as 4.6% MAE.
• U-Mamba-Bot showed superior robustness with limited data and unbalanced pathologies; diverse training data generally improved performance, but excess HCM cases reduced accuracy.

Conclusions:

• State-of-the-art segmentation models can generalize to rare diseases like LVNC with sufficient multi-pathology training data.
• Careful selection and balancing of pathology types are crucial for robust model performance in data-scarce, rare disease scenarios.