Introduction: Frontotemporal dementia and Alzheimer's disease exhibit numerous overlapping symptoms, making differential diagnosis challenging. While magnetic resonance imaging (MRI) and diffusion tensor imaging provide valuable information about the structure of the brain, the integration of different features from these data and applying a deep learning model for classifying frontotemporal dementia and Alzheimer's disease remains relatively unexplored. We aimed to construct a structural connectome based on integrated data using anatomically constrained tractography and evaluate a convolutional neural network's performance dedicated to structural connectomes in classifying Frontotemporal dementia, Alzheimer's disease, and healthy individuals.

Methods: Neuroimaging data included a total of 360 subjects, including Frontotemporal dementia, Alzheimer's disease, and normal controls, from the Neuroimaging Institute of Frontotemporal Degeneration and the Alzheimer's Disease Neuroimaging Initiative datasets. By merging T1-weight and DTI scans, a structural connectivity matrix was constructed based on anatomically constrained tractography. To enable an accurate diagnosis using those connectomes, we trained a modified deep neural network on connectome datasets, which effectively extracted topological features from the structural connections. Finally, the performance of the proposed DL model was assessed via cross-validation.

Results: this approach recognized the discriminative features from the structural network of the brain in Frontotemporal dementia, Alzheimer's disease, and healthy individuals achieving the accuracy levels of 0.96, 0.94, 0.92, and 0.93 for ternary and binary classification between Frontotemporal dementia/Alzheimer's disease, Frontotemporal dementia/healthy individuals, and Alzheimer's disease/healthy individuals, respectively.

Conclusion: the results show that the key features emerging from structural connections can be potentially discriminative in disease-vulnerable regions in Frontotemporal dementia, Alzheimer's disease, and control groups, and the developed model outperform other conventional neural networks and traditional machine learning used in previous studies.