Precise detection of Alzheimer's disease (AD), especially at the early stages, i.e., early mild cognitive impairment (EMCI) and MCI, allows the physicians to promptly intervene to prevent the progression to advanced stages. However, identification of such stages using non-invasive brain imaging techniques like DWI, remains one of the most challenging tasks due to the subtle and mild changes in the brain structures of the subjects. Findings from previous studies suggested that topological organization alterations occur in the DTI-derived structural connectomes in MCI patients. Therefore, for improving diagnosis performance, we presented a connectome-based deep learning architecture based on BrainNet Convolutional neural network (CNN) model. The proposed model automatically extracts hidden topological features from structural networks using specially-designed convolutional filters. Experiments on 360 subjects, including 120 subjects with EMCI, 120 subjects with MCI and, 120 normal controls (NCs), with both T1-weighted MRI and DWI scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI), provided the highest binary classification accuracies of 0.96, 0.98, and 0.95 for NC/EMCI, NC/MCI and EMCI/MCI respectively.

In addition, we also investigated the effect of different atlas sizes and fiber descriptors as edge weights on the discriminative ability of the classification performance. Experimental results indicate that our approach exhibited superior performance to previous methods and performed effectively without any prior complex feature engineering and regardless the variability of imaging acquisition protocols and medical scanners.

Finally, we observed that DTI-based graph representation of brain regions connections preserve important but hidden connectivity pattern information to discriminate between clinical profiles, and our proposed approach could be easily extended to other neurodegenerative and neuropsychiatric diseases.