Intracerebral hemorrhage (ICH) and perihematomal edema (PHE) are key imaging markers of primary and secondary brain injury in hemorrhagic stroke. Accurate segmentation and quantification of ICH and PHE can help with prognostication and guide treatment planning. In this study, we combined Swin-Unet Transformers with nnU-NETv2 convolutional network for segmentation of ICH and PHE on non-contrast head CTs. We also applied test-time data augmentations to assess individual-level prediction uncertainty, ensuring high confidence in prediction. The model was trained on 1782 CT scans from a multicentric trial and tested in two independent datasets from Yale (n = 396) and University of Berlin Charité Hospital and University Medical Center Hamburg-Eppendorf (n = 943). Model performance was evaluated with the Dice coefficient and Volume Similarity (VS). Our dual Swin-nnUNET model achieved a median (95% confidence interval) Dice = 0.93 (0.90–0.95) and VS = 0.97 (0.95–0.98) for ICH, and Dice = 0.70 (0.64–0.75) and VS = 0.87 (0.80–0.93) for PHE segmentation in the Yale cohort. Dice = 0.86 (0.80–0.90) and VS = 0.91 (0.85–0.95) for ICH and Dice = 0.65 (0.56–0.70) and VS = 0.86 (0.77–0.93) for PHE segmentation in the Berlin/Hamburg-Eppendorf cohort. Prediction uncertainty was associated with lower segmentation accuracy, smaller ICH/PHE volumes, and infratentorial location. Our results highlight the benefits of a dual transformer-convolutional neural network architecture for ICH/PHE segmentation and test-time augmentation for uncertainty quantification.

Hematoma expansion (HE) is an independent predictor of poor outcomes and a modifiable treatment target in intracerebral hemorrhage (ICH). Evaluating HE in large datasets requires segmentation of hematomas on admission and follow-up CT scans, a process that is time-consuming and labor-intensive in large-scale studies. Automated segmentation of hematomas can expedite this process; however, cumulative errors from segmentation on admission and follow-up scans can hamper accurate HE classification. In this study, we combined a tandem deep-learning classification model with automated segmentation to generate probability measures for false HE classifications. With this strategy, we can limit expert review of automated hematoma segmentations to a subset of the dataset, tailored to the research team’s preferred sensitivity or specificity thresholds and their tolerance for false-positive versus false-negative results. We utilized three separate multicentric cohorts for cross-validation/training, internal testing, and external validation (n = 2261) to develop and test a pipeline for automated hematoma segmentation and to generate ground truth binary HE annotations (≥3, ≥6, ≥9, and ≥12.5 mL). Applying a 95% sensitivity threshold for HE classification showed a practical and efficient strategy for HE annotation in large ICH datasets. This threshold excluded 47–88% of test-negative predictions from expert review of automated segmentations for different HE definitions, with less than 2% false-negative misclassification in both internal and external validation cohorts. Our pipeline offers a time-efficient and optimizable method for generating ground truth HE classifications in large ICH datasets, reducing the burden of expert review of automated hematoma segmentations while minimizing misclassification rate.