Developing a Data-Driven System for the Early Identification of Alzheimer’s Disease through MRI Analysis

Abstract

Alzheimer’s Disease is one of the most common neurological disorders and currently affects more than 55 million people worldwide. Detecting the disease in its early stages is important because early treatment can help slow down the progression of symptoms. In most hospitals, MRI brain scans are examined manually by radiologists, which can take time and may sometimes lead to inconsistent interpretations.

This project presents a deep learning-based system designed to classify MRI brain images into different stages of Alzheimer’s disease using the OASIS dataset. One of the main difficulties encountered during this work was the severe imbalance between classes. The Moderate Dementia category represents only a very small portion of the dataset compared to the healthy cases. To address this issue, techniques such as Weighted Random Sampling and Weighted Cross-Entropy Loss were applied during model training.

Several neural network architectures were evaluated, including a basic Convolutional Neural Network (CNN), EfficientNet-B0, and a fine-tuned ResNet-18 model. Among these, ResNet-18 produced the best performance, achieving a test accuracy of 96.82% and a Macro F1-score of 0.96. The model also achieved 100% recall for the Moderate Dementia class, meaning that all severe cases were successfully detected during evaluation.

To demonstrate practical usability, the model was integrated into a web-based system using FastAPI for the backend and Next.js for the frontend, allowing MRI images to be analyzed quickly on consumer-level hardware.