Parameter-Efficient Adaptation of SAM3 for Left Atrium Segmentation from MR Images for Improved Planning and Guidance of Cardiac Ablation Interventions
Foundation models such as Segment Anything Model 3 (SAM3) have demonstrated strong generalization in natural images through prompt-driven segmentation. However, their applicability to medical imaging remains unclear due to the domain-specific visual characteristics and task requirements. In this work, we investigate the domain adaptability of SAM3 for left atrium (LA) segmentation from cardiac MRI, a challenging, yet clinically critical task in the context of image-guided cardiac interventions, specifically aimed at improving planning and guidance of catheter-based cardiac ablation therapy in arrhythmia patients. We first evaluate SAM3 in a zero-shot setting using text prompts and show that performance remains limited despite descriptive prompt formulations, highlighting a fundamental mismatch between semantic prompting and the visual cues required for medical image segmentation. To address this limitation, we study two parameter-efficient fine-tuning (PEFT) strategies, including LoRA and a prompt-free adapter-based framework. Our comprehensive experimental results demonstrate that LoRA achieves stronger performance (Dice 90.6% and 91.7%) over zero-shot baselines and cross-dataset generalization. LoRA leverages the semantic priors, while the adapter provides stable predictions with improved computational efficiency. In particular, adapter training reduces per-epoch training time by over 7× compared to LoRA, while maintaining comparable segmentation performance. These findings reveal a key trade-off between generalization and efficiency and show that semantic guidance from text prompts and feature-level adaptation to modality-specific visual cues offer two viable, complementary pathways for adapting SAM3 to cardiac MRI. Our code is available here.
Authors
- Bipasha Kundu
- Cristian A. Linte (ORCID: https://orcid.org/0000-0001-7602-7937)
- Richard M Simon (ORCID: https://orcid.org/0000-0003-3558-4598)
- Nakul Poudel (ORCID: https://orcid.org/0009-0000-8957-9730)
Institutions
- Rochester Institute of Technology (US)
Publication Details
- Journal
- Authorea
- Published
- 2026-11-21
- DOI
- https://doi.org/10.22541/authorea.15012559/v1
- Primary Topic
- Medical Image Segmentation Techniques
- Type
- preprint