A Comparative Analysis of Pre-Trained Models for Identifying Alzheimer's Disease

Neurodegenerative diseases like Alzheimer's disease (AD) remain a critical global healthcare challenge, highlighting the importance of early and accurate detection. Deep learning models, especially transfer learning (TL), leverage pre-trained weights, enhancing model performance with limited data. These models are trained on various images, and each task requires fine-tuning. This study evaluates three prominent pre-trained architectures, ResNet50, EfficientNetB0, and InceptionV3, to compare their precision and accuracy in image classification. These three models are fine-tuned to diagnose AD efficiently. The four-class unbalanced dataset used for our task is collected from the Open Access Series of Imaging Studies-1 (OASIS-1). Synthetic Minority Oversampling Technique (SMOTE) is performed to convert the dataset into a balanced one. This dataset was then fed to the fine-tuned models, and accuracy was calculated for each model. The results were also compared with one another and it was found that EfficientNetB0 outperformed the others in all aspects, having a validation accuracy and loss of 99.65 % and 1.18 %, respectively. Its performance was found superior to some of the state-of-the-art models also tried by other prominent researchers.