Automated Intracranial Hemorrhage Segmentation in CT Images Using Attention U-Net with Deep Supervision and Test-Time Augmentation

Background/Objectives: Intracranial hemorrhage is a life-threatening condition that requires rapid and accurate diagnosis using computed tomography (CT). Manual segmentation is time-consuming and subject to interobserver variability, motivating the development of automated segmentation methods. This study proposes an enhanced Attention U-Net for intracranial hemorrhage segmentation using the RSNA Intracranial Hemorrhage Detection dataset, in which only 1082 of 20,154 images have pixel-level annotations. The proposed architecture integrates attention gates, deep supervision, cosine annealing, and test-time augmentation. An ablation study was also conducted to evaluate the contribution of each component. Methods: The proposed Attention U-Net and a baseline U-Net were trained and evaluated using identical experimental protocols. Performance was assessed on an independent test set using the Dice score, IoU, sensitivity, and precision. To evaluate the reproducibility of the performance advantage across datasets, both architectures were additionally trained and evaluated under identical protocols on the independent Brain Hemorrhage Segmentation Dataset (BHSD), comprising 2961 annotated axial slices derived from 192 CT volumes. The BHSD test set consisted of 621 slices. Results: On the independent RSNA test set, the proposed Attention U-Net achieved a Dice score of 0.7965 (±0.0814), outperforming the baseline U-Net (0.7683 ± 0.0906). The proposed model also achieved an IoU of 0.6689, sensitivity of 0.8094, and precision of 0.7966. On the BHSD test set, the Attention U-Net again outperformed the baseline U-Net, achieving a Dice score of 0.5676 versus 0.5063, sensitivity of 0.5350 versus 0.5138, and IoU of 0.4408 versus 0.3924. Thus, the proposed architecture demonstrated a consistent relative performance advantage across both independent datasets, although absolute performance was lower on BHSD than on RSNA. Conclusions: The findings demonstrate that incorporating attention mechanisms, multiscale supervision, and inference-time enhancement can improve intracranial hemorrhage segmentation performance. The consistent improvement over the baseline U-Net across two independent datasets further supports the effectiveness and cross-dataset robustness of the proposed architecture. The results also highlight the potential of semi-supervised learning to leverage the large volume of unlabeled CT data available in intracranial hemorrhage datasets.