Study tests 5 medical 3D AI models on MRI artifacts; 3DINO most robust
3DINO resists MRI corruption while BrainIAC collapses—raising clinical AI safety concerns.
A study evaluated five pretrained 3D medical foundation models—3DINO, BrainIAC, NeuroVFM, BrainFM, and Neuro-SimCLR—for robustness to MRI artifacts using BraTS-Africa cases with four MRI sequences. Across seven artifact types at five severity levels, robustness was strongly model- and artifact-dependent: 3DINO showed the most consistently stable representations, BrainIAC was highly sensitive to several corruptions, and the remaining models displayed intermediate but distinct profiles. The findings show that larger-scale or domain-specific pretraining alone does not guarantee artifact invariance and motivate explicit robustness evaluation before deploying 3D foundation models in heterogeneous MRI settings.
- Five 3D encoders (3DINO, BrainIAC, NeuroVFM, BrainFM, Neuro-SimCLR) were tested on BraTS-Africa MRI with 7 artifact types at 5 severity levels.
- 3DINO showed the most stable representations; BrainIAC was the most sensitive, and CKA dropped while RankMe stayed stable—indicating geometric distortion without collapse.
- Segmentation consistency degraded especially under ghosting and Rician noise, and larger-scale pretraining did not guarantee artifact invariance.
Why It Matters
Medical AI models can silently fail on corrupted MRI scans; robustness testing is essential before clinical deployment.