Comorbidity-aware transfer learning for neuro-developmental disorder diagnosis.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Neuroimaging-based diagnosis of neurodevelopmental disorders (NDDs) faces challenges due to complex spatiotemporal dynamics in fMRI data.
• Deep learning computer-aided diagnosis (CAD) systems are promising but struggle with noise and confounding factors in fMRI.
• NDDs like autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD) exhibit overlapping neural pathways, complicating diagnosis.

Purpose of the Study:

• To develop a novel framework, comorbidity-aware transfer learning (CATL), for enhanced NDD diagnosis using fMRI.
• To improve the accuracy and reliability of CAD systems for NDDs by addressing noise and confounding factors in fMRI data.
• To gain mechanistic insights into cross-disorder neural dynamics in NDD comorbidities.

Main Methods:

• Proposed the CATL framework integrating transfer learning and pseudo-labeling for enhanced representation generation from fMRI data.
• Employed an encoder-decoder architecture for feature reconstruction and a lightweight convolutional neural network (CNN) for classification.
• Utilized a unified semi-supervised transfer learning paradigm to model shared neurobiological pathways in NDD comorbidities.

Main Results:

• CATL achieved diagnostic accuracies of 77.34% for ASD and 73.28% for ADHD on benchmark datasets.
• Outperformed state-of-the-art transfer learning methods by 7.86% for ASD and 0.66% for ADHD.
• Demonstrated the ability to disentangle task-relevant temporal features from confounding patterns in fMRI data.

Conclusions:

• The CATL framework offers a robust approach for NDD diagnosis using fMRI, improving diagnostic accuracy.
• CATL provides valuable mechanistic insights into the shared neurobiological underpinnings of NDD comorbidities.
• This approach enhances the potential of AI-driven CAD systems in clinical neuroimaging for NDDs.