African Brain Scans Reveal Vulnerabilities in 3D AI Medical Models to MRI Artifacts

New research investigates the resilience of 3D medical foundation models when confronted with common MRI scanning artifacts. These AI models are increasingly vital for medical image analysis, acting as feature extractors, but their reliability in real-world clinical settings, particularly with imperfect scan data, has been a significant unknown. The study systematically evaluates five different pretrained 3D encoders, assessing how various frequency and image-domain artifacts impact their internal data representations.

The researchers utilized a dataset of brain scans from BraTS-Africa, a crucial resource for brain tumor research in an African context, applying seven types of artifacts at varying intensities across four MRI sequences. Robustness was measured using several metrics, including linear centered kernel alignment (CKA), RankMe, and UMAP, alongside an independent analysis of segmentation consistency. This comprehensive approach allowed for a detailed understanding of how different AI models react to corrupted input.

The findings reveal that the robustness of these AI models is highly variable, depending significantly on both the specific model architecture and the type of artifact. For instance, 3DINO demonstrated the most consistent stability in its representations, whereas BrainIAC proved highly susceptible to several forms of corruption. Other models like NeuroVFM, BrainFM, and Neuro-SimCLR showed intermediate but distinct vulnerabilities. Importantly, the study highlights that simply using larger datasets or domain-specific pretraining does not automatically guarantee artifact immunity.

This research carries significant implications for the deployment of AI in medical imaging across Africa. Given the diverse healthcare infrastructure and potential for varied MRI scan quality, understanding and mitigating the impact of artifacts is paramount. The study underscores the necessity for explicit robustness evaluations before integrating 3D foundation models into clinical workflows, especially in regions where scan conditions might be less than ideal. Ensuring the reliability of these AI tools is critical for accurate diagnoses and effective patient care.