Advancing Neurological Disorder Diagnosis: Integrating Explainable and Generalizable Machine Learning Models

Abstract

Alzheimer’s disease (AD) and Parkinson’s disease (PD) are progressive neurodegenerative disorders that place a growing burden on patients, families and health systems around the world. MRI gives a non-invasive way to look at the structural brain changes tied to these conditions, but reading scans by eye is slow, subjective and hard to scale up. This paper brings together three stages of a doctoral research programme, each built on the last, aimed at an MRI-based diagnostic framework that is accurate, efficient and explainable. The first stage ran a baseline comparison, pitting a convolutional neural network (CNN) against four classical classifiers – Support Vector Machine, Random Forest, k-Nearest Neighbours and Logistic Regression – trained on PCA-reduced features, for four-class Alzheimer’s staging. The CNN came out ahead of every classical model, reaching 69% accuracy against Random Forest’s best classical result of 65.7%, which suggests that features learned automatically from the images carry more diagnostic signal than handcrafted, dimensionality-reduced ones. The second stage moved away from training a CNN from scratch and toward optimized transfer learning. Pre-trained MobileNetV2 and ResNet50 backbones were adapted to the Alzheimer’s Multiclass (Equal and Augmented) dataset using a two-stage freeze/fine-tune schedule, Adam optimization, data augmentation, dropout, early stopping and checkpointing. This optimization-focused approach pushed accuracy to 75% for MobileNetV2 and 67% for ResNet50, and pointed to a broader pattern: a lighter architecture, trained with a carefully designed schedule, can generalize better than a deeper one when medical-imaging data is limited. The third stage extended the framework in two directions at once – generalizing across diseases by adding an independent Parkinson’s disease MRI dataset, and adding interpretability through Grad-CAM and Grad-CAM++ visualizations layered onto the MobileNetV2/ResNet50 backbones. MobileNetV2 again came out on top, reaching 86.17% accuracy on the four-class Alzheimer’s dataset and 96.39% on the two-class Parkinson’s dataset, while ResNet50 struggled with unstable convergence and a bias toward the majority class on both tasks. Grad-CAM++ gave sharper, more clinically localized activation maps than standard Grad-CAM, and the model’s attention lined up with cortical and ventricular regions known to be involved in neurodegeneration. Together, the three stages trace a single line of argument: deep learning beats classical baselines, principled optimization of transfer learning lifts accuracy substantially further, and a lightweight, optimized backbone can end up both generalizable across diseases and explainable to clinicians. The combined results point to MobileNetV2 with Grad-CAM++ as a workable basis for computer-aided, interpretable diagnosis of neurological disease, and set up future work on more advanced architectures, additional XAI methods, and fusing MRI with PET, EEG and clinical data.