Automated rating of Fazekas scale in fluid-attenuated inversion recovery (FLAIR) MRI of ischemic stroke patients using machine learning

초록

Background: While the Fazeka scale poses limitations in white matter hyperintensities (WMH) assessment, deep learning offer promising methods for WMH segmentation in ischemic stroke. Methods: This study is a substudy of multi-center randomized double-blind controlled trial, involving patients within 6 months of experiencing stroke. Additionally, patients meeting the criteria from a stroke registry at a tertiary hospital were included. We developed a two-step deep learning pipeline for automatic Fazeka scale prediction from FLAIR MRI images: (1) WMH segmentation using a U-net-based residual network (uResNet), and (2) Fazeka scale rating using a 3-dimensional convolutional neural network (3D CNN). The inferred raw probability volume of WMHs obtained through uResNet was used to train a 3D CNN. Results: A total of 421 patients were included, and the number of patients with Fazeka scores 0, 1, 2, and 3 were 97 (23.0%), 102 (24.2%), 119 (28.3%), and 103 (24.5%), respectively. Overall, uResNet achieved an AUPRC of 0.81 (95% CI, 0.55–0.95) and a Dice score of 0.73 (95% CI, 0.49–0.87) in the test set. The absolute error between true and predicted WMH volumes was 3.1 ml (95% CI, 0.0–15.9 ml), with no significant differences observed within each Fazeka scale subgroup. In the agreement analysis, the Rsquared value was 0.94, the concordance correlation coefficient was 0.97, and the systematic difference was 0.4 ml. The 3D CNN achieved a quadratic weighted kappa value of 0.951 for Fazeka score prediction. Conclusions: Our deep learning pipeline demonstrated accurate automatic WMH segmentation and Fazekas scale grading in stroke patients.