DCG-ANet: A Dual-Stream Causal Graph Adversarial Network for Early Detection of Alzheimer’s Disease from Multimodal Brain Imaging

Resumo

Alzheimer disease (AD) is a rising health hazard that creates complicated mental loss and memory loss. Diagnosis within a reasonable time and at the right time is significant in retarding the occurrence or development of the disease and recovery of the patient. Although neuroimaging has increased our knowledge of changes in the brain as is witnessed in AD, the issue of accurate mapping of abnormal neural circuitry at both structural and functional levels is a complex one. To remedy this, a new Dual-Stream Causal Graph Adversarial Network (DCG-ANet) is proposed in this work, that takes advantage of the DTI data and rs-fMRI data to identify the strongest patterns relating to AD. Perceiving the architecture as a whole, parallel graph encoders with vision-transformer get to learn whole-brain representations, whereas the model embraces a temporal attention system that regulates changes in the functional connectivity as time progresses. Causal subgraph selection module identifies disease-relevant neural circuits and adversarial training improves the attention of the network to the discriminative features. Using the ADNI dataset, the proposed DCG-ANet achieved 94.3% accuracy, 90.6% F1-score, 91.2% recall, 89.9% precision, and 0.92 MCC, with stable training and validation loss convergence and superior ROC curve performance. The model performs exceptionally well in detecting early Alzheimer, and it is even better than the most sophisticated base techniques. It uses circuit level knowledge to provide an impressive performance. This implies that it has greater predictive power, strength and resilience to all the significant evaluation measures.