ADHD prediction from individual-space T1 images using a Vision Transformer with a gross-region grid framework

Area of Science:

• Neuroimaging and Artificial Intelligence
• Computational Neuroscience
• Psychiatric Diagnostics

Background:

• Predicting attention-deficit/hyperactivity disorder (ADHD) using neuroimaging is challenging due to significant individual differences in brain structure.
• Existing methods often rely on predefined regions of interest (ROIs) or convolutional neural networks (CNNs), which may not fully capture complex morphological variations.

Purpose of the Study:

• To develop and evaluate an end-to-end framework using Vision Transformer (ViT) models for predicting ADHD directly from T1-weighted MRI data.
• To assess the impact of anatomical data reduction on diagnostic performance by comparing whole-brain (WB) versus a reduced set of 11 representative slices (R11).

Main Methods:

• An end-to-end framework employing ViT models was proposed to learn discriminative features directly from individual-space T1-weighted MRIs.
• Two anatomical coverage patterns were evaluated: whole-brain (WB) axial slices and a reduced set of 11 representative slices (R11).
• Performance was benchmarked against baseline CNN and conventional ROI-based approaches, and interpretability was assessed using SHAP analysis.

Main Results:

• The ViT model achieved the highest numerical Area Under the Curve (AUC), outperforming baseline CNN and ROI-based methods, and performing comparably to ResNet.
• Reducing anatomical coverage from WB to R11 slices did not result in a statistically significant degradation of diagnostic performance (AUC 0.75, p=0.19).
• SHAP analysis identified the precentral gyrus and occipital regions as critical neuroanatomical substrates for ADHD classification within the R11 configuration.

Conclusions:

• Transformer-based self-attention models can effectively integrate distributed morphological variations for ADHD diagnosis from T1-weighted MRI.
• High diagnostic integrity for ADHD can be maintained even with substantial reduction in anatomical data, suggesting efficient feature learning by ViT.
• The findings highlight the potential of ViT models for developing anatomically coherent and data-efficient approaches to ADHD diagnosis.