Comparing foundation models and nnU-Net for segmentation of primary brain lymphoma on clinical routine post-contrast T1-weighted MRI

Primary Central Nervous System (PCNS) lymphoma is an aggressive brain tumor, and MRI segmentation is crucial for diagnosis and follow-up. Segmentation foundation models (FM) offer an alternative to classical supervised deep learning. This paper evaluates the performance of segmentation foundation models (FMs), including Segment Anything Model (SAM), MedSAM, and UniverSeg, against traditional supervised learning (with nnU-net) for lymphoma segmentation using clinical routine data and for glioma segmentation using the public dataset MSD-BraTS. Results showed that supervised learning outperformed FMs significantly by vast margins on both datasets. Task-specific models like nnU-net remain essential for complex segmentation tasks.