Data-Efficient Lung Segmentation Using Foundational Models: A Comparative Study of SAM and CNNs for Hyperpolarized Gas MRI

Motivation: Improving hyperpolarized gas MRI segmentation is crucial, especially with limited annotated data, as current methods are labor-intensive and slow. Goal(s): This study aims to investigate whether foundational models like SAM can perform high-quality lung segmentation using smaller, task-specific datasets, improving data efficiency and segmentation accuracy. Approach: We conducted experiments using SAM model fine-tuned on 25% of the training dataset, comparing their performance against traditional CNN-based models on proton and hyperpolarized gas MRI. Results: SAM outperformed traditional CNN models with Dice Similarity Coefficients of 0.97 on proton MRI and 0.97/0.96 on hyperpolarized gas MRI, proving its robustness with 25% data. Impact: This study demonstrates the potential of foundational models like Segment Anything Model (SAM) to significantly improve lung segmentation accuracy using less data, opening new avenues for medical imaging in clinical settings where acquiring large, annotated datasets is challenging.