Assessment of "Zero-Shot" General Purpose Segmentation Models: An Analysis of the Meta "Segment Anything Model" on...

Newly developed zero-shot segmentation algorithms, like "Segment Anything Model 2" (SAM2) of Meta, have the potential to automate segmentation processes more efficiently than existing solutions. The goal of this study was to assess the ability of SAM2 to segment meningioma MRIs and suggest paradigms to enhance and assess performance.

We used SAM2 to produce segmentation masks using T1-weighted MRIs within the 2023 BraTS Preoperative Meningioma Data Set. We also proposed interactive click-based and contour-based augmentation strategies to simulate a neuroradiologist's workflow, alongside a novel ensembling method. Analyses evaluated the performance across model iterations both overall and within clinical subgroups of interest using standard statistical techniques and measures.

Our cohort included a total of 690 meningiomas, most being World Health Organization grade 1 (75%). SAM2 achieved an overall zero-shot segmentation average Dice score of 0.785. Both click-based and contour-based augmentation strategies provided significant model improvement (0.876 and 0.872, respectively, P < .001). Layering a directional consensus approach on top of the contour-based model further enhanced performance (0.921, P < .001). Across all model iterations, smaller tumor volumes and tumors without peritumoral edema proved more difficult for SAM2 to segment (P < .001).

SAM2 demonstrated reasonable zero-shot segmentation performance on meningioma MRIs, with observable improvements seen with contour-based prompting and directional ensembling. These results suggest that zero-shot segmentation models, with some degree of radiologist assistance or intervention, are promising tools for aiding in image segmentation for meningiomas. Future work can investigate methods to improve segmentation performance for small tumor volumes and tumors without peritumoral edema.