CNN-based Alzheimer's disease classification using fusion of multiple 3D angular orientations

Abstract

Convolutional neural networks (CNN) can extract the features necessary for the recognition and classification of several diseases. Yet, the intricate symptoms encompassing changes in brain anatomy pose challenges for CNN training. While an ideal scenario would leverage a patient's entire magnetic resonance imaging (MRI) data with minimal preprocessing and human involvement, it does not always yield optimal results. To improve the performance of CNNs, researchers utilize much larger and more complex networks, which does not guarantee improvement. In this paper, we propose an innovative way to increase performance, manifested through utilizing multiple distinct 3D orientations of the data, coupled with a multi-classifier framework. The method consists of predictions from networks trained on unique angular orientations of the same data set that combine to offer a unified prediction. The results obtained from the proposed method underscore that these minimalistic, computationally frugal alterations can propel average accuracy rates from 89.84% to a commendable 94.37%, signifying a near 5% performance surge.