Development of a deep learning-based system for Alzheimer’s disease diagnosis

Ling, Sheng Han (2026) Development of a deep learning-based system for Alzheimer’s disease diagnosis. Final Year Project, UTAR.

PDF Download (3021Kb)

Abstract

Alzheimer’s Disease (AD) is an irreversible neurodegenerative condition that requires early and precise diagnosis for successful treatment. Current diagnosis methods using manual MRI analysis are labor-intensive and subjective. Although deep learning models automate diagnosis, they can be difficult to interpret as being “black-box” and computationally expensive, making them challenging to use in practice. This research aims to overcome these issues by creating a lightweight and interpretable deep learning-based desktop AD diagnostic system. The application uses a lightweight convolutional neural network (CNN) with less than five million parameters to classify T1-weighted structural magnetic resonance imaging (sMRI) into Cognitively Normal (CN), Mild Cognitive Impairment (MCI) and AD classes. For enhanced interpretability, the system was associated with Gradient-weighted Class Activation Mapping (Grad-CAM), which creates heatmaps that display the disease-affected brain region. The system was developed and evaluated solely on Alzheimer’s Disease Neuroimaging Initiative (ADNI-1) data. By carefully tuning hyperparameters and applying factor-based structured pruning, the final model has only 2.997 million trainable parameters, while maintaining a high test accuracy of 97.05%. Additionally, this study delivered a desktop application using CustomTkinter consisting of an interactive MRI viewer, real-time inference with heatmap visualization, session management, and PDF report generation. This research successfully implemented a transparent, efficient, and actionable AI-assisted diagnostic system, bridging the gap between powerful deep learning models and clinical applications.

Actions (login required)