Objective:

To develop an objective framework for ADHD diagnosis using resting-state functional MRI (rs-fMRI) data that addresses limitations of current clinical assessments, including subjectivity and inter-rater variability.

Approach:

• Framework Development: VoxSTNet processes raw four-dimensional rs-fMRI BOLD volumes directly, employing a two-stage pipeline for moderate compression and subject-wise z-score normalization.
• Model Architecture: Utilizes a time-distributed 3D-CNN combined with a gated recurrent unit (GRU) for spatiotemporal representation, and HiResCAM for voxel-level interpretability.
• Data Analysis: Evaluated on the ADHD-200 dataset with 760 subjects using Leave-One-Site-Out (LOSO) and five-fold cross-validation methods.

Key Findings:

• Achieved 98.7% accuracy with five-fold cross-validation and 78.4% accuracy under LOSO protocol.
• Sensitivity of 98.2% and specificity of 99.1% were reported.
• Area under the receiver operating characteristic curve (AUC) of 99.4% (95% confidence interval [CI]: 97.9–99.5%).
• HiResCAM saliency maps highlighted the right caudate nucleus, consistent across validation subjects.

Interpretation:

VoxSTNet demonstrates effective voxel-level modeling of rs-fMRI data for ADHD diagnosis.

Limitations:

• Performance may vary across different scanner types and settings.
• Further work needed to enhance cross-site generalization.

Conclusion:

The study presents a promising approach for objective ADHD diagnosis using raw rs-fMRI data.