Three-Stage Deep Learning Approach for Alzheimer’s and Parkinson’s Disease Detection

Abstract

Purpose: Diagnosis of Alzheimer's disease (AD) and Parkinson's disease (PD) remains challenging, and existing AI-based neuroimaging tools are typically applied in isolation for diagnosis, multimodal analysis, or progression tracking. This study aims to develop an integrated, three-stage deep learning framework that unifies these tasks into a single diagnostic and prognostic pipeline. Design/Methodology/Approach: The framework was developed and validated using the OASIS, ADNI, PPMI, and TCIA databases and comprises three sequential stages. The first stage, MHNFSP, applies fuzzy c-means segmentation to MRI/fMRI scans, followed by extraction of Fourier transform, entropy, and convolutional features; the Whale Optimisation Algorithm is then used for feature selection, and a hybrid neural-statistical classifier performs final prediction. The second stage conducts multimodal assessment by fusing MRI, CT, and X-ray data through a multi-head self-attention mechanism, using UNet++ for region-of-interest segmentation and Grad-CAM++ for visual explainability. The third stage, LMST-ADNet, focuses on longitudinal disease modelling by combining MRI, PET, and fMRI connectivity data with clinical-genetic information, employing modality-specific encoders, temporal modelling, and self-supervised pre-training. Research Limitation:  The study relies on retrospective, publicly available datasets that may not fully capture the demographic and clinical diversity of real-world patient populations. External, prospective, multi-site validation and testing on unseen clinical cohorts would be required to confirm generalisability before clinical deployment. Findings: The MHNFSP stage achieved 96.39% classification accuracy on MRI/fMRI data. The multimodal fusion stage achieved 95.3% accuracy, a mean Dice score of 0.92 and an Intersection over Union (IoU) of 88.0% agreement with expert-marked regions. The LMST-ADNet stage achieved 96.8% classification accuracy, an AUC of 0.982, and a C-index of 0.91 for progression risk prediction. Ablation studies confirmed that temporal modelling substantially contributes to performance, with additional gains obtained through self-supervised pre-training. Practical Implication: The proposed three-stage system offers clinicians and radiologists an integrated tool capable of supporting early diagnosis, multimodal cross-verification, and long-term monitoring of neurodegenerative disease progression, potentially reducing diagnostic delay and supporting more personalised treatment planning. Social Implication: Earlier and more reliable detection of AD and PD could improve patient quality of life, reduce caregiver burden, and lower long-term healthcare costs associated with delayed diagnosis and disease management, particularly benefiting ageing populations at higher risk. Originality/Value: Unlike prior approaches that address diagnosis, multimodal fusion, and progression prediction as separate problems, this study proposes a unified three-stage architecture that integrates segmentation-based classification, explainable multimodal fusion, and longitudinal transformer-based progression modelling within a single coherent framework, offering a more holistic assessment pathway for neurodegenerative disease research.