Related Experiment Video
Updated: Aug 5, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
Cross-Subject Registration-Based Augmentation: Alleviating Anatomical Misalignment in Trauma CT for Robust Hemorrhage
Sujong Shin1, Jaewoo Chung2, Jungchan Cho3
1Department of AI-Based Convergence, Dankook University, Yongin 16890, Gyeonggi-do, Republic of Korea.
Diagnostics (Basel, Switzerland)
|July 28, 2026
Summary
A new data augmentation strategy improves deep learning models for brain hematoma segmentation in emergency CT scans. This method enhances anatomical alignment, leading to more accurate results, especially for boundary agreement.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neurosurgery
Background:
- Emergency brain CT scans often show anatomical misalignment due to patient positioning difficulties.
- This misalignment negatively impacts the accuracy of deep learning models for 3D hematoma segmentation.
- Accurate segmentation is crucial for diagnosing and managing traumatic brain injury.
Purpose of the Study:
- To develop a novel data augmentation strategy to improve the robustness of deep learning-based hematoma segmentation.
- To address the challenge of anatomical inconsistencies in emergency CT imaging.
- To enhance the performance of segmentation models without altering their architecture.
Main Methods:
- Proposed a framework named cross-subject registration-based augmentation (CSRA).
- Utilized anatomical landmarks for rigid registration of patient CT volumes to reference volumes.
- Generated anatomically aligned CT-label pairs for model training.
Main Results:
- CSRA-3s demonstrated superior performance, achieving the highest mean Dice and IoU scores.
- CSRA significantly reduced the mean Hausdorff distance at the 95th percentile (HD95), indicating improved boundary agreement.
- Paired analysis confirmed a statistically significant benefit in boundary agreement compared to conventional augmentation.
Conclusions:
- The CSRA strategy effectively mitigates performance degradation caused by anatomical position discrepancies.
- This model-agnostic approach enhances segmentation robustness during training.
- Further multicenter external validation is recommended before clinical implementation.
