ADVANCED DEEP LEARNING FRAMEWORKS FOR ALZHEIMER’S DISEASE DETECTION: A COMPREHENSIVE MULTI-ARCHITECTURE APPROACH INTEGRATING LIGHTWEIGHT 3D CNN AND EFFICIENTNETV2B3 TRANSFER LEARNING WITH EXPLAINABLE AI

Abstract

Background Alzheimer’s disease (AD) is a progressive neurodegenerative disorder characterized by gradual cognitive decline and structural brain changes. Early and objective detection of AD is critical for timely clinical intervention and improved patient management. However, conventional diagnostic approaches rely heavily on neuropsychological assessments and expert interpretation, which may be subjective and insufficiently sensitive to subtle early-stage neuroanatomical alterations visible in magnetic resonance imaging (MRI). Objective This study aims to develop a computationally efficient and interpretable deep learning framework for automated Alzheimer’s disease detection and staging using MRI data. The objective is to achieve high diagnostic accuracy while enhancing clinical transparency through explainable artificial intelligence (XAI) techniques. Methods An integrated deep learning framework was proposed consisting of two complementary pipelines. For binary classification (cognitively normal vs. Alzheimer’s disease), a lightweight three-dimensional convolutional neural network (3D-CNN) augmented with efficient channel attention was employed to directly analyze volumetric MRI data. For multi-class dementia staging (non-demented, very mild, mild, and moderate dementia), a transfer learning approach based on EfficientNetV2B3 was utilized, operating on two-dimensional slices extracted from 3D MRI volumes. Five central slices per subject were selected, and patient-level predictions were generated using majority voting across slice-level predictions. To enhance interpretability, Grad-CAM was applied to the EfficientNetV2B3 model, while volumetric occlusion sensitivity analysis was used for the 3D-CNN. Experimental evaluation was conducted on a balanced ADNI-derived dataset with a 70/15/15 train/validation/test split for the binary task, and on the Kaggle Alzheimer’s Disease Classification dataset using stratified 10-fold cross-validation for multi-class staging. Additionally, hierarchical regression analysis was performed to examine the relationship between model predictions and Mini-Mental State Examination (MMSE) scores. Results The lightweight 3D-CNN achieved an accuracy of 90.6% for binary classification of cognitively normal individuals versus Alzheimer’s disease patients. For four-class dementia staging, the EfficientNetV2B3 model achieved an accuracy of 99.45%. Explainability analyses highlighted neuroanatomically relevant regions consistent with known AD pathology. Hierarchical regression results demonstrated that deep learning model outputs explained a significant proportion of variance in MMSE scores, supporting the clinical relevance of the proposed framework. Conclusion The proposed integrated deep learning framework delivers high diagnostic and staging performance while maintaining computational efficiency and interpretability. By combining volumetric and slice-based analysis with explainable AI techniques, the system provides clinically meaningful insights aligned with neuroanatomical expectations. This approach has the potential to support objective, early-stage Alzheimer’s disease detection and improve decision-making in clinical practice.