Study Reveals MRI Artifact Sensitivity in 3D Medical Foundation Models
A recent investigation published on arXiv (2608.06613) thoroughly examines the resilience of five self-supervised 3D medical foundation models against MRI artifacts. The models analyzed include 3DINO, BrainIAC, NeuroVFM, BrainFM, and Neuro-SimCLR, utilizing BraTS-Africa cases with four MRI sequences and subjected to seven types of frequency- and image-domain artifacts across five specific corruption settings. The evaluation of robustness employed methods such as linear centered kernel alignment (CKA), RankMe, UMAP, and an independent segmentation-consistency assessment. Findings indicate that robustness varies significantly by model and artifact: 3DINO shows the most stable representations, whereas BrainIAC is particularly vulnerable to various corruptions. NeuroVFM, BrainFM, and Neuro-SimCLR present intermediate, yet distinct, responses to specific artifacts, underscoring the importance of careful model selection in clinical scenarios involving MRI artifacts.
Key facts
- Study evaluates five 3D medical foundation models: 3DINO, BrainIAC, NeuroVFM, BrainFM, Neuro-SimCLR
- Uses BraTS-Africa cases with four MRI sequences
- Generates seven frequency- and image-domain artifacts at five predefined corruption settings
- Robustness assessed using linear CKA, RankMe, UMAP, and segmentation-consistency analysis
- 3DINO shows most stable representations
- BrainIAC is highly sensitive to several corruptions
- NeuroVFM, BrainFM, and Neuro-SimCLR show intermediate but distinct artifact-specific profiles
- Study published on arXiv with identifier 2608.06613
Entities
Institutions
- arXiv