Headache classification and automatic biomarker extraction from structural MRIs using deep learning.

Area of Science:

• Neuroimaging
• Machine Learning
• Neurology

Background:

• Traditional machine learning for neurological disease classification requires manual feature extraction and domain expertise.
• Deep learning offers an automated approach, identifying relevant features directly from neuroimaging data.

Purpose of the Study:

• To develop and validate a deep learning pipeline for classifying brain MRIs of individuals with migraine, acute post-traumatic headache, and persistent post-traumatic headache against healthy controls.
• To identify brain regions serving as biomarkers for each headache type using the developed deep learning model.

Main Methods:

• A deep learning pipeline using a 3D ResNet-18 model was developed for binary classification tasks.
• The pipeline incorporated data preprocessing, classification, and biomarker extraction.
• Techniques including incorporating a public dataset and resampling were used to address challenges of limited and imbalanced data.

Main Results:

• The pipeline achieved high classification accuracies: 75% for migraine, 75% for acute post-traumatic headache, and 91.7% for persistent post-traumatic headache versus healthy controls.
• Significant biomarkers were identified for each headache type, including specific regions in the caudate, thalamus, occipital, temporal, and parietal lobes.
• The model demonstrated high sensitivity and specificity in distinguishing headache types from healthy controls.

Conclusions:

• Deep learning methods can effectively classify different headache types from brain MRIs without requiring prior domain knowledge.
• The study highlights the potential of deep learning for automated biomarker discovery in neuroimaging for headache disorders.
• This approach significantly reduces manual effort and enhances the efficiency of diagnosing neurological conditions.