LiteCardiacSAM: Efficient Adaptation of SAM for Cardiac MRI Segmentation

Background: Segmentation of cardiac magnetic resonance imaging (CMR) is essential for deriving clinical biomarkers such as ventricular volumes and ejection fraction. The Segment Anything Model (SAM) has shown strong zero-shot performance across general domains, but its behavior in specialized medical imaging, including CMR, is less understood. Key open questions include: (a) how well SAM and lightweight variants perform in strict zero-shot cardiac settings, (b) how parameter-efficient fine-tuning (e.g., Low-Rank Adaptation, LoRA) improves performance, (c) how many labeled examples are needed for peak accuracy, and (d) whether such gains transfer to unseen datasets.
We contribute: (1) LiteCardiacSAM, a lightweight, cardiac-specialized SAM checkpoint trained on ACDC; (2) a quantitative study on the minimum number of examples required for optimal performance; and (3) an evaluation of cross-dataset transfer of few-shot gains.

Methods: We compared three Tiny-ViT–based SAM derivatives: MobileSAM, LiteMedSAM, and our LiteCardiacSAM. Experiments used three datasets: ACDC (public) and M&Ms (public, multicentric) and SSC (private, Argentina). Two protocols were applied:
(a) Intra-dataset few-shot: LoRA fine-tuning with 1–100 labeled examples from the target dataset; Dice score measured on the same dataset.
(b) Cross-dataset transfer: fine-tuning on one dataset, evaluation on another. Transfer efficiency is defined as the ratio between the mean dice scores on the target and the source dataset.
All runs used identical preprocessing, augmentation, and training/inference pipelines.

Results: In intra-dataset few-shot learning, LiteCardiacSAM consistently reached peak Dice scores with fewer examples. On ACDC, SSC, and M&Ms, LiteCardiacSAM achieved close-to-maximum performance with ~10, ~15, and ~10 examples, respectively, versus 50 for MobileSAM and LiteMedSAM across the different datasets. In cross-dataset tests, LiteCardiacSAM maintained most of its few-shot gains and outperformed baselines in both Dice score and sample efficiency. Transfer efficiency for M&M is above 100% due to the multicentric nature of the dataset, allowing for a robust representation of different acquistion distributions. Dice vs. shots trends are shown in Figure 1.

Conclusion: LiteCardiacSAM surpassed MobileSAM and LiteMedSAM across datasets, achieving higher Dice scores with significantly fewer labeled examples—sometimes reaching near-max accuracy with only ~5 examples, versus 5–10× more for baselines. Few-shot improvements transferred well to unseen datasets, supporting the use of targeted LoRA fine-tuning on SAM derivatives for efficient, generalizable CMR segmentation. These results highlight the potential for specialized, lightweight SAM models in rapid clinical and research deployment.

Figure 1. Few-shot segmentation performance of LiteCardiacSAM (ours), MobileSAM, and LiteMedSAM on ACDC, M&Ms, and SSC datasets. Mean Dice score (myocardium) is shown as a function of the number of training examples (“shots”). LiteCardiacSAM achieves peak performance with fewer examples across all datasets. Bottom panels show cross-dataset transfer experiments, with LiteCardiacSAM maintaining high transfer efficiency (mean 91.1%) relative to same-dataset performance.