Most existing intracranial hematoma segmentation models target acute hemorrhages and may not generalize to the heterogeneous morphology of chronic subdural hematomas (CSDH). We compared a model trained on an open-access acute intracranial hemorrhage dataset with a model trained in combination with a CSDH dataset and further evaluated the performance of the combined dataset model across Nakaguchi subtypes of CSDH (homogeneous, laminar, separated, and trabecular) and between isodense and non-isodense hematomas. We analyzed 377 patients with 512 CSDHs. A 3D nnU-Net model was initialized with the open-access Brain Hemorrhage Segmentation Dataset (BHSD). In the second stage, the BHSD and institutional CSDH data (75% training, 25% testing) were combined for retraining. Finally, a model trained exclusively on the CSDH dataset was developed for comparison with the combined dataset model. Segmentation performance was evaluated using the Dice similarity coefficient (DSC) stratified by subtype and volume. The combined dataset model outperformed the BHSD-only model (mean DSC 0.917 ± 0.099 vs. 0.425 ± 0.306; P 3: 0.786 ± 0.264 vs. > 25 cm3: 0.931 ± 0.046; P < .001). The proposed deep learning model enables more accurate and reliable segmentation of CSDH across hematoma subtypes and densities.
Reddy et al. (Tue,) studied this question.