Comparative Evaluation of Transfer Learning Models on Dementia Prediction

Abstract

Deep learning algorithms show potential for predicting Alzheimer's disease (AD) or Dementia using Magnetic Resonance Imaging (MRI). Alzheimer's disease is more prevalent among the elderly population. The disease progresses to a severe stage before any symptoms occur, resulting in a brain dysfunction that cannot be treated medically. Thus, a timely diagnosis is essential for halting its development. Alzheimer disease detection and prevention is a contentious issue in the scientific community right now. This study demonstrates automatic feature extraction and categorization of brain Magnetic resonance imaging's (MRIs) associated with Alzheimer's disease using Convolutional Neural Network (CNN)-based transfer learning models. These models outperform on traditional techniques in differentiating between the four Alzheimer's disease phases from Kaggle dataset. Three transfer learning (TL) models, EfficientNetV2B1, InceptionResnetV2, and InceptionV3 have been employed in this study, each model's efficacy is also evaluated. Training accuracy of 87.12%, 99.23%, 98.40% is achieved on EfficientNetV2B1, InceptionResnetV2 and InceptionV3 respectively. InceptionV3 has achieved best accuracy on test (unseen) data.