Automated rating of Fazekas scale in fluid-attenuated inversion recovery MRI for ischemic stroke or transient ischemic attack using machine learning

Comprehensive Summary

White matter hyperintensities (WMH) are usually measured using the Fazekas scale, which relies on subjective visual grading. This study set out to build and test a deep learning pipeline that could automatically segment and grade WMH in stroke patients using T2-weighted FLAIR images. The model combined a residual U-Net for WMH segmentation with a 3D convolutional neural network for Fazekas score prediction. When tested on data from 471 stroke patients, the model showed strong accuracy in identifying WMH regions, close alignment with actual WMH volumes, and highly reliable agreement with expert-assigned Fazekas scores across both internal and external patient groups.

Outcomes and Implications

This work matters because WMH burden is an important factor in understanding and predicting vascular outcomes, yet current grading methods are subjective and inconsistent. By demonstrating that a deep learning pipeline can consistently and accurately automate WMH evaluation in stroke patients, the study provides a tool that could standardize assessment, improve efficiency, and ultimately support better prediction of future vascular events.