Abstract Number: Esoc2026a1400 Architectural Influences on Automated Detection and Localization of Intracranial Hemorrhage Subtypes on Noncontrast Head Ct: A Comparative Deep Learning Study

Abstract Background and aims To develop a deep learning framework with optimized training and input strategy for automated detection and localization of five acute intracranial hemorrhage subtypes on brain CT scans. Methods This retrospective study utilized three publicly available datasets: Brain Hemorrhage Extended (BHX, 15,979 thin-slice images), CQ500+ (2,586 regular-slice images), and RSNA+ 2019 (90,982 training and 14,155 testing images) including 12,868 patients with five hemorrhage subtypes: intraventricular (47,239 annotations), intraparenchymal (52,240), subarachnoid (54,667), subdural (46,997), and epidural (2,220). Twelve object detection architectures were evaluated including YOLOv8 variants, RetinaNet, Faster R-CNN with multiple backbones, RT-DETR, and CO-DETR. Training strategies compared single-model multi-class versus single-model single-class approaches. Performance was assessed using mean average precision, Dice similarity coefficient, and intersection over union. Statistical significance was determined using appropriate parametric or non-parametric tests based on normality assessment, with Bonferroni correction applied for multiple comparisons (p 0.05). Results Hybrid Faster R-CNN with Swin Transformer achieved superior performance with mean average precision of 0.90~0.96 and Dice similarity coefficients (DSC) of 0.88~0.93 on internal validation. Single-model single-class training significantly outperformed multi-class training on external validation (DSC: 0.25~0.45 versus 0.14~0.27, p 0.05). Despite substantial performance degradation for most subtypes, epidural hemorrhage maintained robust detection with only 2~7% reduction in non-zero mean average precision (0.95~0.97 to 0.89~0.94) across top architectures. Conclusions Hybrid convolutional neural network-transformer architectures with single-model single-class training achieve optimal intracranial hemorrhage detection with superior external generalization. Conflict of interest Chiao-Hua Lee, Hikam Muzakky, Cheng-En Juan, Chia-Ching Chang, Ya-Hui Li, Tung-Yang Lee, Cheng-Hsuan Juan, Ming-Ting Tsai, Yui-Jui Liu, Chun-Jung Juan: nothing to disclose