Development and validation of an automated MRI-based pipeline for temporal classification of intracerebral hemorrhage using U-Net segmentation and machine learning.

This study aimed to develop and validate an automated magnetic resonance imaging (MRI)-based pipeline for temporal classification of intracerebral hemorrhage (ICH) using clinically feasible multi-sequence MRI, 2.5D U-Net segmentation, and machine learning. This retrospective study included a development cohort of 56 patients and an independent validation cohort of 115 patients with ICH. Five MRI sequences-diffusion-weighted imaging, apparent diffusion coefficient map, fluid-attenuated inversion recovery (FLAIR), time-of-flight magnetic resonance angiography, and susceptibility-weighted angiography-were analyzed. A pontine reference region of interest was automatically placed using FLAIR-based brain segmentation. Hematoma regions were automatically segmented using a 2.5D U-Net trained in the development cohort, and the largest three-dimensional connected component was defined as the representative hematoma region of interest. Pontine ROI-referenced signal features and shape features were extracted, and Random Forest, XGBoost, and multinomial logistic regression classifiers were compared for five-stage temporal classification. The selected final model was applied to the validation cohort. In the validation cohort, automated hematoma segmentation achieved a Dice similarity coefficient of 0.794 ± 0.208, sensitivity of 0.799 ± 0.235, and precision of 0.842 ± 0.151. One case with automated hematoma ROI extraction failure was excluded from temporal classification analysis. In the remaining 114 cases, the final Random Forest model achieved an accuracy of 0.807, balanced accuracy of 0.763, macro F1-score of 0.749, and mean Average Precision of 0.851. The proposed automated MRI-based pipeline showed feasible performance for ICH temporal classification and may support objective MRI-based assessment of hemorrhage stage.