Enhancing Alzheimer’s disease Classification from MRI Scans Using Deep Learning techniques

Abstract

Detecting Alzheimer’s disease as early as possible is important in slowing cognitive decline and enhancing patient outcomes. In this work, we introduce a deep-learning framework for an automatic categorization of brain MRI scans into non-demented, very mild demented, mild demented, and moderate demented. The proposed model incorporated a deep learning algorithm, EfficientNet-B3, using transfer learning and fine-tuning with the inclusion of an original dataset with modified and augmented MRIs to provide a better model generalization and robustness. The data was partitioned with 70% designated for training 15% designated for validating, and 15% designated for testing the augmented MRI data. Data augmentation techniques were implemented to limit overfitting and reduce class imbalances. A build-out Pytorch data flow pipeline was employed, which included weighted sampling and implementation of an adaptive learning rate schedule optimized using cross-entropy loss functions for optimization. The model achieved validation accuracy of 99.4%, with high precision and F1-scores across all dementia stages. The results indicated successful application of transfer learning and network augmentation within the context of medical imaging studies. The implications of this work highlight that lightweight Maria models, such as EfficientNet-B3, with slight adaptations to the architecture and transfer learning can be a powerful diagnostic assistive-screening tool for clinical and diagnostic decision support for Alzheimer’s disease.