Explainable 3D deep learning on full-head MRI suggests extracranial involvement in Alzheimer's disease

Deep learning on full-head T1 MRI distinguished Alzheimer's from normal cognition with 91.07% accuracy, 95.16% AUC; saliency maps highlighted scalp and skull, suggesting extracranial signals are relevant. Hypothesis-generating, requires validation.

• Three-class AD/MCI/cognitively normal accuracy was 75.00% with 86.09% ROC-AUC.
• Grad-CAM saliency consistently involved peripheral regions beyond the brain (scalp, skull).
• These preliminary findings require independent external validation and are hypothesis-generating only.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.