An Interpretable Hybrid Multi-Channel Fusion Approach for Alzheimer’s Disease Diagnosis

Abstract Alzheimer disease (AD) is a common neurodegenerative disease in the elderly where a timely and proper diagnosis is required to reduce disease progression. The paper describes a deep learning method in classifying the MRI brain scans of Alzheimer disease (AD), Mild Cognitive impairment (MCI) and cognitively normal (CN) subjects in Alzheimer Disease Neuroimaging Initiative (ADNI) dataset. Several state-of-the-art convolutional neural network (CNN) models (DenseNet, ResNet, InceptionV3, Xception, and EfficientNet) were trained and compared to draw a baseline in terms of performance. To boost the accuracy of the diagnosis further, a Proposed Fusion Model which combines DenseNet169 and ResNet18 models was created by the hybrid strategy of fusion at both feature and decision levels. The proposed model was outperforming all the baseline models including EfficientNetB0 (accuracy 92.45%), which yields a superior performance (according to the accuracy) of 96.66, balanced accuracy 95.44%, precision 97.37%, recall 95.43%, F1-score 96.30%, AUC 99.42% and MCC 94.70%. These findings prove the multi-channel fusion method to be a successful method of exploiting complementary properties, thereby boosting generalization and resistance to imbalance among classes. The suggested framework presents a trustworthy and explicable computer-aided diagnostic system to detect early signs of Alzheimer and give a hopeful outlook on the subsequent clinical implementation of deep learning in the neurodegenerative disease investigation.