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

Background: The past decade has seen a reversal in the U.S long-term decline in age-adjusted mortality rate from stroke. Timely stroke detection can boost the patient’s chances for recovery by enabling life-saving treatment and informing the patient of their increased risk of successive infarcts. Since no single imaging modality can currently provide accurate and safe stroke detection at both acute and chronic stages, there is a need to develop novel imaging biomarkers with both diagnostic and prognostic value. Methods: We trained a U-shaped, nested hierarchical transformer model (UNesT) for T1-w white matter infarct segmentation using the ATLAS R2 dataset. Model reproducibility was independently evaluated on the Washington University (WU) stroke dataset. To boost T1-w UNesT stroke detection performance, automated machine learning techniques were used to extract 77 novel resting state fMRI (rs-fMRI) stroke biomarkers. Results: Stroke detection performance of the T1-w UNesT model degraded from Dice indices of 0.611 to 0.24 and 0.41 for the subacute and chronic timepoints respectively in the WU dataset. After UNesT re-optimization with the training portion of the WU dataset, the test set Dice index improved to 0.41–0.50. The spectral peak amplitude at the subacute timepoint increased the T1-w UNesT Dice index from 0.41 to 0.50 (p < 0.01) and correlated with language recovery. Conclusions: By training a UNesT model on the T1-w stroke data from one dataset and evaluating it on an independent dataset, we highlight the dataset drift concerns. Spectral peak amplitude is proposed as a novel rs-fMRI biomarker for improving stroke detection and predicting stroke recovery trajectory.