Enhancing multi-class neurodegenerative disease classification using deep learning and explainable local interpretable model-agnostic explanations

Introduction Alzheimer's disease (AD) and Parkinson's disease (PD) are two of the most prevalent neurodegenerative disorders, necessitating accurate diagnostic approaches for early detection and effective management. Methods This study introduces two deep learning architectures, the Residual-based Attention Convolutional Neural Network (RbACNN) and the Inverted Residual-based Attention Convolutional Neural Network (IRbACNN), designed to enhance medical image classification for AD and PD diagnosis. By integrating self-attention mechanisms, these models improve feature extraction, enhance interpretability, and address the limitations of traditional deep learning methods. Additionally, explainable AI (XAI) techniques are incorporated to provide model transparency and improve clinical trust in automated diagnoses. Preprocessing steps such as histogram equalization and batch creation are applied to optimize image quality and balance the dataset. Results The proposed models achieved an outstanding classification accuracy of 99.92%. Discussion The results demonstrate that these architectures, in combination with XAI, facilitate early and precise diagnosis, thereby contributing to reducing the global burden of neurodegenerative diseases.