Evaluating a Deep Learning Foundation Model for Neuroimaging Segmentation in the Data-Rich and Data-Constrained Settings

Motivation: Deep learning models for segmentation require large datasets, limiting their use in clinical settings. Foundation models, which learn from non-medical images before fine-tuning for specific tasks, have emerged as a possible solution. Goal(s): We aimed to adapt and evaluate a recent foundation model, the Medical Segment Anything model (MedSAM), for neuroanatomy segmentation. Approach: Using the Human Connectome Project dataset, we trained MedSAM to segment 102 regions-of-interest and compared its accuracy with a baseline UNet model. Results: UNet outperformed MedSAM in almost all regions and dataset sizes, but MedSAM showed potential when training with very few MRIs. Impact: While foundation models such as MedSAM have potential for medical segmentation, they currently may not surpass traditional models when using sufficient data. In the data-limited setting, however, they can be useful when extremely little labeled data is available.