Deep learning-based segmentation of aneurysmal subarachnoid hemorrhage: toward accurate and scalable prognostic imaging.

BACKGROUND: Accurate segmentation of aneurysmal cerebral hemorrhages, including subarachnoid hemorrhage (SAH), intraparenchymal hemorrhage (IPH), and intraventricular hemorrhage (IVH), is essential for clinical decision-making. Manual segmentation, however, is time-consuming and prone to interobserver variability. This study aimed to develop and evaluate deep learning models for automatic aneurysmal hemorrhage segmentation, compare them against state-of-the-art methods, and assess their impact on six-month Glasgow Outcome Scale (GOS). METHODS: This retrospective study analyzed CT scans from in-house datasets of aneurysmal SAH, IPH, and IVH, complemented with public datasets covering broader hemorrhage types and etiologies. Two training experiments were conducted: Experiment 1 used SAH, IPH, and IVH cases, while Experiment 2 incorporated trauma-related datasets to increase hemorrhage variability. Both employed nnU-Net for model training. Segmentation performance was evaluated quantitatively against manual segmentation and other AI algorithms, and qualitatively through blinded assessments by seven neurosurgeons. Logistic regression analyses assessed the predictive value of lesion volume measured by manual and automated segmentations, as well as the modified Fisher CT grade, for six-month functional outcome. RESULTS: Experiment 1 trained on 356 patients and Experiment 2 on 530; both tested on the same cohort of 89 hemorrhagic cases. Experiment 2 achieved the best performance, with a median Dice Score of 0.81, recall of 0.82, intraclass correlation coefficient type 3 of 0.92, and a median volume difference of 1.40 mL. The model outperformed previous AI algorithms and reduced processing time by 97% compared to manual annotations. In outcome modeling, both manual and automated lesion volumes showed similar predictive ability for six-month GOS, with comparable AUCs and discrimination metrics. Volume-based models also outperformed the modified Fisher scale. CONCLUSIONS: This work proposes a fully automated approach for segmenting intracranial hemorrhages, focused on aneurysmal SAH. The model outperformed existing methods in accuracy and drastically reduced annotation time. Automated volume estimations matched manual annotations in predicting long-term outcomes and surpassed the modified Fisher scale. These results support integrating AI-based segmentation tools into clinical workflows for outcome modeling in aneurysmal hemorrhage.