Advancing Alzheimer's Diagnosis: Performance Analysis of Deep Learning Ensembles in Medical Imaging

Neuroimaging and deep learning have become the focus of much research for diagnosing Alzheimer's disease (AD) in recent years. Nevertheless, the limited availability of neuroimaging training data has resulted in significant overfitting issues for several deep learning models. However, the dataset was very imbalanced. To address this issue, we employed several augmentation techniques such as flipping, rotating, and zooming to balance the dataset. The ADNI dataset consisted of four classes: mild demented, moderate demented, very mild demented, and non-demented. In this work, we have proposed an ensemble of EfficientNetB2, DenseNet121, VGG16, and Xception for the diagnosis of Alzheimer at early stage. Experiment results shows that EfficientNetB2 ensemble with Xception and VGG16 performed extremely well with 99.1% and 99.4% respectively.