Cross-Subject Registration-Based Augmentation: Alleviating Anatomical Misalignment in Trauma CT for Robust Hemorrhage Segmentation.
Authors
Affiliations (4)
Affiliations (4)
- Department of AI-Based Convergence, Dankook University, Yongin 16890, Gyeonggi-do, Republic of Korea.
- Department of Neurosurgery, Dankook University College of Medicine, Cheonan 31116, Chungcheongnam-do, Republic of Korea.
- Department of Computing, Gachon University, Seongnam 13120, Gyeonggi-do, Republic of Korea.
- Department of Computer Engineering, Dankook University, Yongin 16890, Gyeonggi-do, Republic of Korea.
Abstract
<b>Background/Objectives:</b> In emergency settings, it is often infeasible to place patients in an anatomical position for CT scanning. Consequently, emergency brain CT scans of patients with traumatic brain injury frequently demonstrate considerable anatomical misalignment. These inconsistencies compromise the performance of deep learning-based 3D hematoma segmentation. This study aims to enhance segmentation robustness by proposing a registration-based data augmentation strategy utilizing anatomical landmarks. <b>Methods:</b> We propose a framework termed cross-subject registration-based augmentation (CSRA), which uses anatomical landmarks to rigidly register patient CT volumes to selected reference CT volumes and generate anatomically aligned CT-label pairs for training. <b>Results:</b> A total of 339 patients who underwent brain CT imaging were enrolled from a Level 1 trauma center. CSRA-3s achieved the highest mean Dice and IoU and the lowest mean HD95 among the evaluated augmentation strategies. Patient-level paired analysis showed that the most robust statistically supported benefit was a significant reduction in HD95 compared with a conventional geometric augmentation baseline, indicating improved boundary agreement. <b>Conclusions:</b> The proposed augmentation strategy mitigated the effect of anatomic position discrepancies on segmentation performance, particularly in terms of boundary agreement, without modifying existing model architectures. CSRA may serve as a model-agnostic training-time augmentation strategy for improving segmentation robustness in anatomically inconsistent emergency CT imaging, although multicenter external validation is required before clinical deployment.