Goal: Automated ADHD classification from neuroimaging is limited by cross-site acquisition heterogeneity, high inter-subject variability, and reliance on region-of-interest (ROI)-averaged connectivity that discards voxel-level detail. Methods: We propose a four-branch 3D convolutional neural network (3D-CNN) operating directly on voxel-resolution Athena-derived maps, fALFF and ReHo from resting-state fMRI, and Grey Matter (GM) and White Matter (WM) probability maps from structural MRI, without ROI summarisation. Each modality is processed by a dedicated branch, and the four 32-d embeddings are combined through two-stage hierarchical fusion: functional and structural pairs first, then across modalities, before binary classification. Results: On 768 ADHD-200 subjects, 5-fold cross-validation gave 74.80% mean accuracy, a 5.65 percentage-point gain over the strongest voxel-level baseline. Leave-One-Site-Out validation on five mixed-diagnosis sites gave 73.40-79.74% accuracy (mean 76.54%,±,2.41%), matching or exceeding within-distribution performance. Conclusions: The four-branch 3D-CNN generalises across unseen sites, supporting voxel-resolution multimodal fusion as a robust approach for ADHD classification.