Hybrid Vi+ECNN framework for advanced ADHD diagnostic accuracy in medical imaging

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Pediatric Neurology

Background:

• Attention-Deficit/Hyperactivity Disorder (ADHD) is a common neurodevelopmental disorder with significant personal and societal impacts.
• Current ADHD diagnostic methods rely on subjective behavioral assessments, highlighting the need for objective, automated tools.

Purpose of the Study:

• To develop an ADHD-specific, biologically informed deep learning framework for pediatric brain MRI classification.
• To integrate Vision Transformer (ViT) and Enhanced Convolutional Neural Network (ECNN) with multiple MRI representations for comprehensive neuroanatomical analysis.

Main Methods:

• A multi-stream deep learning architecture combining ViT and ECNN was employed.
• Input data included Raw MRI, Phase Spectrum Transform (PST) for cortical irregularities, and Quantile Histogram Equalization with Denoising (QHED) for gray-white matter contrast.
• The framework aimed to capture complementary global and local neuroanatomical features.

Main Results:

• The proposed ViT+ECNN model achieved exceptional classification performance on a pediatric MRI dataset.
• Achieved 99.4% accuracy, 99.3% precision, 99.5% recall, and a 0.99 F1-score.
• Outperformed standalone ViT and ECNN models, demonstrating the efficacy of the hybrid approach.

Conclusions:

• Hybrid transformer-convolutional models show significant potential for enhancing ADHD diagnostic accuracy.
• This deep learning framework offers a promising avenue for the early identification and intervention of ADHD in children.