Abstract
BACKGROUND: We propose an approach to adapt a segmentation foundation model, segment-anything-model (SAM), for cine cardiovascular magnetic resonance (CMR) segmentation and evaluate its generalization performance on unseen datasets. METHODS: We present our model, cineCMR-SAM, which introduces a temporal-spatial attention mechanism to produce segmentation across one cardiac cycle. We freeze the pre-trained SAM's weights to leverage SAM's generalizability while fine-tuning the rest of the model on two public cine CMR datasets. Our model also enables text prompts to specify the view type (short-axis or long-axis) of the input slices and box prompts to guide the segmentation region. We evaluated our model's generalization performance on three external testing datasets including a public multi-center, multi-vendor testing dataset of 136 cases and 2 retrospectively collected in-house datasets from 2 different centers with specific pathologies: aortic stenosis (40 cases) and heart failure with preserved ejection fraction (HFpEF) (53 cases). RESULTS: Our approach achieved superior generalization in both the public testing dataset (Dice for LV=0.94 and for myocardium=0.86) and two in-house datasets (Dice ≥0.90 for LV and ≥0.82 for myocardium) compared to existing CMR deep learning segmentation methods. Clinical parameters derived from automatic and manual segmentations showed a strong correlation (r ≥0.90). The use of both text prompts and box prompts enhanced the segmentation accuracy. CONCLUSION: cineCMR-SAM effectively adapts SAM for cine CMR segmentation, achieving high generalizability and superior accuracy on unseen datasets.