BG-3DM2F: Bidirectional gated 3D multi-scale feature fusion for Alzheimer's disease diagnosis

A computer-aided diagnosis system is one of the crucial decision support tools under the medical imaging scope. It has recently emerged as a powerful way to diagnose Alzheimer's Disease (AD) from structural magnetic resonance imaging scans. However, due to the deficit of recognition memory in the Mild Cognitive Impairment (MCI) stage, semantic feature ambiguity, and high inter-class visual similarities problems, computer-aided diagnosis of AD remains challenging. To bridge these gaps, this paper proposed a hippocampus analysis method based on a novel 3D convolutional neural network fusion strategy, called Bidirectional Gated 3D Multi-scale Feature Fusion (BG-3DM2F). The suggested BG-3DM2F framework consists of two modules: 3D Multi-Scale Chained Network (3DMS-ChaineNet) and Bidirectional Gated Recurrent Fusion Unit (Bi-GRFU). The 3DMS-ChaineNet architecture is introduced to design the subtle features and capture the variations in hippocampal atrophy, while the Bi-GRFU scheme is investigated to store 3DMS-ChaineNet levels in the forward and backward fashion and retain them in the decision-making process. For validation, our solution is completely evaluated on the public Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset. Practically, we conducted empirical evaluations to verify the effect of BG-3DM2F components. In comparison with the current state-of-the-art methods, the experiments show that our proposed approach provides efficient results, achieving the accuracies of 98.12%, 95.26%, and 96.97% for binary classification of Normal Control (NC) versus AD, AD versus MCI, and NC versus MCI, respectively. Therefore, we can conclude that our proposed BG-3DM2F system has the potential to dramatically improve the conventional classification methods for assisting clinical decision-making.