NeuroEarlyAI: An Explainable Deep Learning Framework with EarlyDetectNet for Alzheimer’s Disease Detection and Classification using MRI Scans

Alzheimer’s Disease (AD) is an ageing and progressive neurodegenerative disease, and one of the most prevalent dementias that affects millions of patients worldwide, including millions of people in Europe, and is a public health priority. Early and right diagnosis is of paramount importance for early therapeutic intervention as well as retardation of the disease. The visual search for small changes in the brain is rather time-consuming and subjective. For comparison, we chose two traditional diagnostic procedures: clinical evaluation and manual MRI analysis, where the clinical evaluation is more appropriate for the intended user in healthcare centres. The automatic detection of AD using recent advancements in deep learning are yet to lack a few drawbacks, such as class imbalance, overfitting and generalisation issue on external datasets, low transparency in the decision-making of the deep learning model, which makes it difficult for clinical applications. To fill this gap, in this paper, we propose NeuroEarlyAI, an end-to-end learning framework with EarlyDetectNet, a new model for early and interpretable detection of AD from the structural MRI scans. The framework also consists of sophisticated pre-processing strategies (denoising, data augmentation and skull stripping) to enhance the quality and diversity of the data. EarlyDetectNet has an attention mechanism which highlights clinically relevant anatomical regions. As a contrast, the visual and feature-wise explanation modules are based on the Grad-CAM and SHAP techniques, respectively, which project the model prediction onto the visual and feature-wise explanations. Extensive tests are performed on two publicly available datasets: Alzheimer's MRI Dataset and OASIS-1 dataset. The proposed system demonstrated improvements over the state of the art on reaching accuracy for binary and multi-class classification in a cross-dataset setting, with 96.7% and 93.8% accuracy, respectively. These results show the promise of the applications and broader use of NeuroEarlyAI, a tool that could help to diagnose AD earlier for better patient care.