Explainable 3D deep learning from full-head MRI suggests scalp and skull involvement in Alzheimer's disease

BackgroundEmerging evidence suggests that extracranial tissues and immune-glymphatic interactions may contribute to neurodegenerative processes in Alzheimer's disease (AD). In this context, neuroimaging studies typically focus on intracranial brain structures, often excluding extracranial structures and adjacent meningeal regions through preprocessing steps.ObjectiveTo investigate whether full-head MRI contains potentially relevant information beyond the brain and whether explainable deep learning can identify spatial patterns associated with AD, mild cognitive impairment (MCI), and cognitively normal (CN) subjects.MethodsAn explainable full-head 3D deep learning framework based on DenseNet-121 with transfer learning from MedicalNet was proposed and applied to T1 MP-RAGE MRI data from the ADNI dataset. The model was trained on full-head MRI volumes without skull stripping using subject-level splits and evaluated across clinically relevant classification scenarios. Grad-CAM was used to localize regions contributing to model predictions.ResultsIn the three-class task (AD, CN, MCI), the model achieved 75.00% accuracy and 86.09% ROC-AUC. Performance was highest for distinguishing AD from CN (91.07% accuracy, 95.16% ROC-AUC) and remained strong for the combined (AD + MCI) versus CN (85.58% accuracy, 96.17% ROC-AUC). Explainability analysis revealed consistent saliency patterns not only within intracranial regions but also in peripheral regions, which may correspond to extracranial structures, such as scalp and skull.ConclusionsFull-head MRI may contain complementary, potentially diagnostically relevant information beyond the brain. Peripheral saliency patterns identified by our framework suggest a broader anatomical context for signals associated with AD. These findings are hypothesis-generating and require further validation across independent datasets and clinical studies, motivating targeted investigation.