Abstract 4143921: Adaptation of Prompt-enabled Segment-Anything-Model Enhance the Accuracy and Generalizability of Cine Cardiac Magnetic Resonance Segmentation

Introduction: Accurate segmentation of cine cardiac magnetic resonance (CMR) throughout the cardiac cycle is essential for comprehensive cardiac functional analysis. However, current deep-learning (DL) approaches often suffer from reduced accuracy on unseen datasets due to generalizability issues. The Segment-Anything Model (SAM) is a new prompt-enabled segmentation foundation model trained on one billion natural images, known for its generalizability and user-defined prompts. Hypothesis: Adapting SAM for cine CMR segmentation can improve accuracy and generalizability on previously unseen CMR data. Methods: We adapted SAM for cine CMR segmentation by incorporating a temporal attention mechanism to maintain temporal consistency across the cardiac cycle and parameter-efficient transfer learning to segment the left ventricle (LV) and myocardium (MYO). User-defined bounding box prompts were used around MYO at end-diastole and end-systole phases to guide the segmentation region. The model can output segmentations for all phases throughout one cardiac cycle at once. We evaluated the model using a multi-center, multi-vendor (M&M) international dataset of 136 cine CMR cases, which was unseen during model training.The trained model was applied to this dataset without fine-tuning to assess generalizability. We compared SAM with three state-of-the-art (SOTA) cine CMR DL segmenters. Segmentation accuracy was evaluated using the Dice coefficient. Clinical parameters, including LV end-diastolic volume (LVEDV) and left ventricular ejection fraction (LVEF), were compared between manual and automatic measurements. Results: The adapted SAM with box prompts demonstrated superior generalization compared to SOTA methods (p<0.01 by one-sided Wilcoxon test). The Dice coefficient was 0.937±0.024 for LV and 0.870±0.029 for MYO, with no significant differences across different centers and vendors, except for one center. The agreement for LVEDV and EF was high, with r = 0.995 and bias = 4.21±6.75 ml for LVEDV and r = 0.97 and bias = -0.50±3.75% for EF. Using bounding box prompts around the MYO region enhanced its segmentation accuracy from 0.853±0.032 to 0.870±0.029 (p<0.05) compared to no prompts. Conclusion: Adapting SAM with box prompts enhances accuracy and generalizability in cine CMR segmentation..