Machine learning methods for brain network classification: Application to autism diagnosis using cortical...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Autism Spectrum Disorder (ASD) impacts brain connectivity, posing challenges for diagnosis via magnetic resonance imaging (MRI) due to heterogeneity.
• Current machine learning (ML) diagnostic frameworks primarily focus on functional and structural connectivity, potentially missing morphological changes.
• Studies on ML for ASD diagnosis using morphological brain networks from T1-weighted MRI are limited.

Purpose of the Study:

• To address the gap in ML-based ASD diagnosis using morphological brain networks.
• To create a diverse set of ML pipelines for neurological disorder diagnosis, specifically applied to ASD.
• To leverage crowdsourcing via a Kaggle competition to achieve these goals.

Main Methods:

• A Kaggle competition was organized, providing participants with a training dataset of T1-weighted MRI scans.
• Participants developed machine learning pipelines to diagnose ASD using cortical morphological networks.
• Performance was evaluated on public and hidden test datasets using accuracy, sensitivity, and specificity metrics.

Main Results:

• The top-ranked team achieved 70% accuracy, 72.5% sensitivity, and 67.5% specificity.
• The second-ranked team achieved 63.8% accuracy, 62.5% sensitivity, and 65% specificity.
• Performance metrics were used to rank teams, with the final ranking based on the mean of all rankings.

Conclusions:

• A competitive ML setting facilitated the exploration and benchmarking of diverse ML methods for ASD diagnosis.
• The study successfully generated ML pipelines for diagnosing ASD using cortical morphological networks.
• This crowdsourcing approach demonstrated the potential of exploring a wide spectrum of ML methods for neurological disorder diagnosis.