GLocalSeg: A global–local collaborative segmentation network for rectal cancer segmentation

Accurate segmentation of normal rectal and tumor regions from CT images is essential for clinical management of rectal cancer. However, existing methods still face significant challenges. On the one hand, low contrast, blurred boundaries, and high morphological variability make the segmentation task inherently difficult. On the other hand, current methods struggle to effectively extract and fuse multi-scale global and local features simultaneously. In this paper, a global–local collaborative segmentation network named GLocalSeg is proposed to address the aforementioned challenges and improve the segmentation accuracy of normal rectum and rectal tumors. A dual-parallel encoder composed of a Hybrid Attention CNNs encoder and a Vision Transformer (ViT) encoder is first constructed to jointly extract fine-grained local details and long-range global context. Building upon these complementary representations, we further design a HybridFusionCDG module that integrates edge-guided structural enhancement, semantic-difference modeling, and gated bidirectional feature interaction, enabling deeper and more coherent coordination between local detailed features and global contextual information. Experimental results demonstrate that our method achieves state-of-the-art performance compared with existing approaches. On CARE dataset, it attains a Mean Dice of 67.27%, Mean IoU of 51.95%, Mean HD95 of 11.3775 mm, and Mean ASD of 3.5519 mm for normal rectum and rectal tumor segmentation. And on TeddyCup dataset, our method achieves Dice, IoU, HD95, and ASD scores of 67.00%, 51.69%, 8.7844 mm, and 2.3415 mm, respectively, for rectal tumor segmentation.