Comparing foundation models and nnU-Net for segmentation of primary brain lymphoma on clinical routine post-contrast T1-weighted MRI
Résumé
Primary Central Nervous System (PCNS) lymphoma is an aggressive brain tumor with variable response rates to current therapeutic strategies. Segmentation of PCNS lymphoma on magnetic resonance imaging (MRI) is important to enhance diagnosis and follow-up. However, it is a difficult task due to the highly variable presentation of lymphoma lesions and the heterogeneity of acquisitions in clinical routine. The recent rise of foundation models (FM) for segmentation, such as Segment Anything Model (SAM), offers a potential alternative to classical supervised deep learning approaches. Nevertheless, their value on challenging tasks such as lymphoma segmentation on clinical routine data has not been thoroughly studied. In this paper, we assessed the performance of several FMs (SAM, MedSAM, UniverSeg) for lymphoma segmentation and compared them to a classical supervised learning with nnU-net. In addition, we performed experiments on the public dataset MSD-BraTS so that others can reproduce our findings. We found that nnU-net outperformed FMs by vast and statistically significant margins. For the lymphoma clinical routine dataset, the difference between nnU-Net and the best FM was about 19 percent points of Dice. For MSD-BraTS, it was about 10 percent points. Our findings suggest that current FMs fall short in handling complex segmentation tasks in particular on clinical routine data and that 3D supervised learning (e.g. with nnU-Net) is essential at present. Trained models for both tasks, code and box prompts (for MSD-BraTS) are available at: https://github.com/GuanghuiFU/medical_cv_foundation_eval.
Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
licence |