Neuro-AI: An Integrated Deep Learning Framework for Enhanced MRI-Based Alzheimer’s Disease Classification

Abstract

Alzheimer’s Disease (AD) is a progressive neurodegenerative condition wherein the early identification of Cognitively Normal (CN), Mild Cognitive Impairment (MCI) and AD is of importance to researchers. The different medical equipment used and data collection settings bring about differences that do not pertain biologically when processing the data using automated analysis tools. This research work proposes, Neuro-AI, a deep learning-based pipeline for MRI image preprocessing, harmonization, super-resolution, explainable classification and automatic report generation. T1-weighted MRI images from 202 selected images of OASIS-3 dataset were used in this study and the obtained data are preprocessed by N4 bias correction, skull stripping and intensity normalization techniques. Then a multi-level enhancement process was employed through FusionNet, CycleGAN and SRGAN methods to maximize the imaging variation and spatial resolution from 128 × 128 pixels to 256 × 256 pixels. Then, a pretrained MobileNetV3-Small model was fine-tuned on classifying CN, MCI and AD images. The model achieved an accuracy of 99.95% and a weighted F1-score of 0.9995 on the balanced dataset. The image quality analysis indicated an increase in the average PSNR by 9.60 dB and the average SSIM by 0.0547 with SRGAN processing. GradCAM was used for generating visualization for model predictions, and TinyLLaMA for creating structured reports based on clinical data using Streamlit. While the results indicate the potential of the proposed system, its dependence on the subject level data and lack of multi-site validation prevent any claims about generalization.