DE-UNeXt: Dual encoder UNeXt for intracranial hemorrhage segmentation on a novel HBU CH dataset

Intracranial hemorrhage(ICH) is one of the highly fatal diseases. Clinically, CT scans are key for doctors to assess the extent of hemorrhage and develop treatment plans. However, there are currently few publicly available pixel-level ICH datasets, which limits the improvement of CT image segmentation performance for intracranial hemorrhage. To address this problem, this paper proposes a new intracranial hemorrhage image dataset, HBU CH, which aims to provide rich, diverse, and realistic intracranial hemorrhage cases. Meanwhile, an end-to-end intracranial hemorrhage segmentation network, DE-UNeXt, is proposed to improve the segmentation accuracy of the lesion parts. First, a Dual Encoder Structure (DE) is proposed, where both the original image and its inverted counterpart are input into two parallel encoders. This approach captures object features from two complementary perspectives, enabling the learning of a richer feature representation. Next, a Dual Feature Compensation (DFC) Module is proposed, which combines traditional convolutional methods with contextual Transformers to process and fuse the features extracted from the DE structure in parallel. The DFC module accounts for both local spatial information and global semantic context, thereby enhancing segmentation accuracy and refining lesion boundaries. Experiments conducted on the proposed HBU CH dataset, as well as two widely used datasets – BCIHM and BHSD – demonstrate that the proposed DE-UNeXt, built on a lightweight network, outperforms the baseline method. Specifically, the Intersection over Union (IoU) and F1 score of DE-UNeXt show improvements of approximately 3.06% and 2.01%, respectively. The code and dataset are available at https://github.com/davidsmithwj/DE-UNeXt for further research. • A intracranial hemorrhage segmentation dataset named HBU CH is proposed. • Light weighted DE-UNeXt proposed for real-time intracranial hemorrhage segmentation. • Dual encoder structure proposed to learn more comprehensive image features. • Dual Feature Compensation proposed to better fuse global and local features.