Evaluating Foundational 'Segment Anything' (Med-SAM1, Med-SAM2) Deep Learning Models for Left Atrial Segmentation in 3D LGE CMR (RF_FR_376)

Background: Atrial fibrillation (AF), the most common cardiac arrhythmia, is associated with heart failure and stroke. Accurate left atrium (LA) segmentation in 3D late gadolinium-enhanced (LGE) CMR is crucial for evaluating AF, as fibrotic remodeling of the LA myocardium sustains arrhythmia and guides treatment. However, manual LA segmentation is time-consuming and challenging. Recently, foundational deep learning models like the Segment Anything Model (SAM)^1,2, pre-trained on millions of diverse datasets and adaptable to various tasks with minimal input, have shown potential in generic segmentation tasks. MedSAM^3,4, a fine-tuned version of SAM for medical tasks, offers efficient, zero-shot segmentation without domain-specific training. However, these models have not yet evaluated for the challenging LA segmentation in 3D LGE-CMR. This study aimed to (1) assess the performance of MedSAM in automating LA segmentation and (2) compare performance of MedSAM2 3D model, using a single prompt with automated tracking, with MedSAM1 2D model, which requires a manual prompt for each slice

 


Methods:

The cohort included 44 pre-ablation AF patients (age 67±9 years, 34% female) from the public CARMA 3D LGE CMR database with a total of 1227 2D slices ⁵, which includes ground truth segmentations of both the LA wall and endocardium (blood pool). We tested MedSAM1 and MedSAM2 foundational models for semi -automated, prompt-based LA segmentation following the analysis pipeline in Fig. 1. MedSAM1 performs 2D segmentation, requiring a user-defined box prompt (ROI) for each LA slice (Fig.1a). MedSAM2, on the other hand, is capable of 3D segmentation using only a single freehand (scribble) prompt drawn on one slice⁶, followed by automated tracking across the subsequent slices (Fig.1b).  The accuracy of LA segmentation was assessed using image-based Dice scores (median [IQR]), with the ground truth defined as a single combined ROI that includes both the LA wall and endocardium (blood pool).
 


Results: Both 2D-based MedSAM1 (per slice prompt) and 3D-based MedSAM2 (single prompt per scan) demonstrated good comparable performance for LA segmentation, with MedSAM1 Dice = 0.84 [0.73, 0.89]  and MedSAM2 Dice = 0.84 [0.64,0.91]  (average of 30 LA slices per scan ) (Fig. 2). However, both models exhibited limitations, in areas with multiple enhanced objects. MedSAM2 frequently produced erroneous segmentation of the aorta in the final slices (Fig. 3a) and sometimes skips the first few LA slices where LA size is small with poor contrast. Yet, MedSAM2 excelled in mid-slice LA segmentation, delivering near-perfect results (Fig. 3b). In contrast, MedSAM1 struggles when pulmonary veins are visible, as its single box prompt per slice approach seem to cause confusion. Additionally, in MedSAM1, the user manually creates a box prompt for each slice, taking approximately 10 seconds per slice (5 minutes for a case with 30 slices), while MedSAM2 requires only a single scribble, taking 20 seconds.

Conclusion: This study demonstrates the potential of foundational models like MedSAM1 and MedSAM2 for automating LA segmentation in 3D LGE-CMR scans of AF patients. MedSAM2 offers similar accuracy with greater efficiency, needing only a single prompt per scan, unlike MedSAM1's per-slice prompts. While both struggled in complex regions, their key strength lies in their adaptability without pre-scan tuning. This generalizability could reduce manual effort, enhance consistency, and streamline cardiac imaging workflows.