White Matter Infarct Detection with Transformer and Auto-ML-Derived Models

Area of Science:

• Neuroimaging
• Medical Artificial Intelligence
• Stroke Medicine

Background:

• Stroke mortality rates are rising after a decline.
• Early stroke detection is crucial for treatment and preventing further infarcts.
• Current imaging lacks comprehensive acute and chronic stroke detection capabilities.

Purpose of the Study:

• To develop novel imaging biomarkers for stroke detection and prognosis.
• To evaluate a U-shaped, nested hierarchical transformer model (UNesT) for stroke segmentation.
• To identify new biomarkers from resting-state fMRI (rs-fMRI) to improve stroke detection.

Main Methods:

• Trained a UNesT model for T1-weighted white matter infarct segmentation on the ATLAS R2 dataset.
• Evaluated model reproducibility on the independent Washington University (WU) stroke dataset.
• Utilized automated machine learning to extract 77 rs-fMRI biomarkers to enhance UNesT performance.

Main Results:

• T1-w UNesT model performance decreased significantly on the WU dataset (Dice indices 0.24-0.41).
• Re-optimization on the WU dataset improved test set Dice index to 0.41-0.50.
• Spectral peak amplitude from rs-fMRI improved T1-w UNesT Dice index (0.41 to 0.50, p < 0.01) and correlated with language recovery.

Conclusions:

• Model performance is sensitive to dataset drift, necessitating careful validation.
• Spectral peak amplitude emerges as a promising rs-fMRI biomarker for stroke detection.
• This rs-fMRI biomarker aids in predicting stroke recovery trajectories.