CT-Based Automated Segmentation and Recurrence Prediction in Chronic Subdural Hematoma: A Dual-Label Multicenter Study.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiology, Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Department of Neurosurgery, Zhangzhou Affiliated Hospital of Fujian Medical University, Zhangzhou, China.
- Department of Medical Imaging, Shanghai General Hospital Jiuquan Hospital, The People's Hospital of Jiuquan, Jiuquan, China.
- Department of Medical Imaging, Ningde Municipal Hospital of Ningde Normal University, Dongqiao Economic and Technological Development Zone, Ningde, China.
- Department of Neurosurgery, First People's Hospital of Fuzhou, Fuzhou, China.
- Department of Radiology, Jintan Affiliated Hospital of Jiangsu University, Changzhou, China.
- School of Computer Engineering, Jiangsu Ocean University, Lianyungang, China.
- R&D Center of Medical Artificial Intelligence and Medical Engineering, Shanghai General Hospital, Shanghai, China.
Abstract
Chronic subdural hematoma (CSDH) frequently recurs after burr-hole surgery, yet most prior imaging studies have focused primarily on hematoma and have not addressed the biomechanical effects of brain compression, which may play an important role in recurrence. In addition, reliance on manual annotation and subjective feature selection limits reproducibility and hinders large-scale clinical translation. This multicenter study included 897 patients with CSDH from six medical centers and developed a fully automated dual-label framework, termed CSDH-Net, to simultaneously segment CSDH and compressed brain tissue using nnU-Net, characterize their interaction through radiomic, volumetric, topological, and intensity-based features, and provide objective recurrence prediction. Recurrence risk was modeled using LightGBM with cross-validation and externally validated in two independent centers, and model interpretability was assessed through shapley additive explanations (SHAP) analyses. The segmentation model achieved a Dice score of 0.953 in internal validation and scores of 0.875 (for CSDH) and 0.980 (for compressed brain tissue) in external testing. The recurrence prediction model yielded area under the receiver operating characteristic curves of 0.830 in training and 0.741 in external validation. At a clinically justified threshold prioritizing sensitivity, the model achieved 85% recall in the external test cohort with acceptable specificity. SHAP and feature importance analyses consistently identified gray-level dependence nonuniformity, hematoma surface area/axis length, mean curvature of compressed brain tissue, and a cross-label spatial descriptor reflecting hematoma thickness as biologically meaningful predictors across datasets. These findings demonstrate that CSDH-Net enables accurate dual-label segmentation and interpretable, imaging-based recurrence prediction across different centers, offering objective and reproducible risk stratification that may support preoperative counseling, personalized follow-up planning, and integration into routine neurosurgical workflows. This study was prospectively registered in the Chinese Clinical Trial Registry (ChiCTR2500110736).