Multiclass Segmentation of Intracerebral Hemorrhage, Intraventricular Hemorrhage, and Perihematomal Edema: Public CT Dataset and Benchmark

Abstract Background Accurate mapping and volume assessment of spontaneous intracerebral hemorrhage (ICH) and its common complications, intraventricular hemorrhage (IVH) and perihematomal edema (PHE), on non-contrast computed tomography (NCCT) help identify patients at risk of poor outcomes or guide treatments but remain challenging. Conventional methods (e.g., ABC/2 formula) and manual segmentation are imprecise or labor-intensive. While deep learning (DL) models show promise, few simultaneously segment all three components, and no publicly available multiclass model or dataset exists. Purpose We aimed to develop and externally test a DL model for multiclass segmentation of ICH, IVH, and PHE on NCCT. Materials and Methods The 2019 Intracranial Hemorrhage Detection Challenge dataset was repurposed for segmentation, with neuroradiologist annotations for ICH, IVH, and PHE serving as the reference standard on 359 selected NCCTs. We trained a DL segmentation model using the nnU-Net framework with five-fold cross-validation (n = 341) and internal testing on 18 cases. External testing was conducted on 183 independent baseline NCCTs of patients with spontaneous ICH at our institution. Performance was evaluated using the Dice similarity coefficient (DSC) and volume agreement. Intra- and inter-annotator variability on 30 exams from the test set served as a benchmark. Results On external (local) testing, the model achieved the highest median DSC for ICH (0.90), with lower performance for IVH (0.67) and PHE (0.53). It did not differ from inter-annotator variability for ICH and inter-specialty variability for IVH, with small median volume differences (<2.5 mL for ICH, <2.0 mL for IVH). Intraclass correlation coefficients (ICCs) for volumetric agreement were 0.98 (ICH), 0.84 (IVH), and 0.80 (PHE). Conclusion We present an externally tested DL model for multiclass segmentation of ICH, IVH and PHE on NCCT. We aim to publicly release our dataset and model to improve reproducibility, support future applications, and provide a benchmark for performance.