Leveraging multi-modal foundation model image encoders to enhance brain MRI-based headache classification

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Headaches are common and typically diagnosed via symptoms.
• Advanced imaging and AI offer new diagnostic possibilities.
• Existing AI models often need large datasets, limiting their use.

Purpose of the Study:

• To fine-tune a pre-trained AI model (BioMedCLIP) for headache classification using brain MRI data.
• To detect potential headache biomarkers from structural MRI.
• To overcome data limitations in AI model training for headache disorders.

Main Methods:

• Leveraged BioMedCLIP, a multimodal foundation model with Vision Transformer (ViT) and PubMedBERT.
• Fine-tuned the ViT model on a dataset of 721 individuals (297 local participants + 424 healthy controls from IXI).
• Classified migraines, acute post-traumatic headache (APTH), and persistent post-traumatic headache (PPTH) against healthy controls using five-fold cross-validation.

Main Results:

• Achieved high accuracy: 89.96% for migraine vs. HC, 88.13% for APTH vs. HC, and 83.13% for PPTH vs. HC.
• Identified key brain regions for classification using Gradient-weighted Class Activation Mapping (Grad-CAM).
• Migraine classification implicated postcentral cortex, supramarginal gyrus, superior temporal cortex, and precuneus cortex.

Conclusions:

• This study is the first to use a multimodal biomedical foundation model for headache classification and biomarker detection from structural MRI.
• The AI model demonstrates high accuracy in differentiating headache types from healthy controls.
• Findings provide novel insights into the neuroanatomical correlates of headache disorders.