DB-Net: A Dual-Branch Hybrid Network for Stroke Lesion Segmentation on Non-Contrast CT Images

Lesion segmentation in acute ischemic stroke (AIS) is critical for accurate diagnosis, with non-contrast CT (NCCT) serving as the primary imaging modality. However, the nature of NCCT often results in low contrast and blurred lesion boundaries, creating significant challenges for accurate segmentation. These limitations can lead to omissions, misdiagnoses, or inaccurate segmentations, directly impacting clinical assessment and timely intervention. To address these challenges, this study proposes a two-branch hybrid network combining a convolutional neural network (CNN) with a Transformer framework. The proposed architecture features the aforementioned dual-branch encoder, as well as a spatial channel difference learning module and parallel attention module. The CNN branch of the encoder performs local feature extraction whereas the Transformer branch captures global dependencies, overcoming limitations of CNN in modeling long-range relationships. The spatial channel difference learning module with multiscale parallel subnetworks enhances feature interaction between branches. Additionally, the parallel attention module in the decoder improves the detection of low-contrast lesion boundaries. Experiments on publicly available AIS datasets demonstrate the effectiveness of the proposed method, achieving a Dice score of 59.11%, with improvements of 4.2% in Dice, 10.6% in Precision, and 5.0% in Jaccard over the best-performing state-of-the-art method. The proposed approach overcomes segmentation challenges in noisy NCCT images and enhances automated methods for AIS diagnosis.