Explainable Deep Morphological Learning of Medial Temporal Lobe Atrophy for Alzheimer’s Diagnosis and Cognitive Scoring

Alzheimer’s disease (AD) is characterized by cognitive decline and hippocampal atrophy. Conventional diagnosis relies on clinical evaluations, biomarkers, and neuroimaging, but these methods are time-intensive and subject to variability. Deep learning (DL) offers automated solutions. However, current research emphasizes classification accuracy, leaving interpretability challenge largely unresolved. This paper presents an explainable DL framework for assessing localized medial temporal lobe atrophy (MTA) and its association with deteriorated cognitive performance. Eighty-four participants underwent T1-weighted magnetic resonance imaging (MRI) and Thai version of the Mini-Mental State Examination (TMSE) assessments. MTA severity was visually rated by three experts. A modified ResNet-18 network trained on hippocampal images for two tasks: (1) MTA rating classification and (2) TMSE-based cognitive group prediction. Explainability was achieved through entropically fused class activation mapping (CAM). MTA and TMSE are strongly correlated. Unlike most AI‑driven classification, our method separates MTA across extended visual scales and examines its morphological links to disease progression. The mean absolute errors (MAE) of the left and right MTA models were 0.604 ± 0.193 and 0.216 ± 0.084, respectively. Normal and severe cases were reliably detected, whilst intermediate results showed deviations within a score unit at the 95% CI, despite inherent ground-truth variability and heterogeneous appearances. The TMSE model achieved a balanced accuracy (BA) of 0.891 ± 0.065. CAM delivered interpretable and relevant evidence of model’s reasoning. It empirically attributed atrophic and impairment patterns to localized structures, in line with clinical diagnostic strategies and established neurophysiological principles. The proposed framework gives effective interpretation of hippocampal morphology for diagnostic and cognitive prediction tasks. Although it provides complementary value to existing clinical approaches, additional comprehensive and multidisciplinary investigations remain essential.