CSFF-Net: Cross-Scale Feature Fusion and Multi-Scale Context Attention-Based Approach for Robust Polyp Segmentation

Engineering Research · Springer Nature · 2026-11-19 · DOI: 10.1007/s13369-026-11781-2

Аннотация

Colorectal cancer is a leading cause of cancer-related mortality worldwide, where early detection and removal of precancerous polyps during colonoscopy significantly reduce disease incidence. However, automated polyp segmentation remains challenging due to substantial variations in polyp size, morphology, texture, illumination, and imaging conditions. In this study, we propose CSFF-Net, an efficient transformer-based segmentation framework designed to enhance multi-scale contextual representation and cross-dataset generalization. The proposed architecture employs a hierarchical SegFormer MiT-B2 encoder to extract rich multi-scale features, while convolutional block attention modules are integrated within skip connections to reinforce discriminative feature learning. A cross-scale feature fusion module and a multi-scale context attention module are introduced to effectively aggregate contextual information across resolutions. Furthermore, a dilated decoder with progressively increasing dilation rates is utilized to enlarge the receptive field and improve boundary delineation. The proposed method is evaluated on five publicly available benchmark datasets, including Kvasir-SEG, CVC-ClinicDB, CVC-ColonDB, ETIS-LaribPolypDB, and CVC-300. Experimental results demonstrate strong and competitive performance across all datasets, achieving mDice/mIoU scores of 0.945/0.897 on Kvasir-SEG and 0.954/0.914 on CVC-ColonDB, establishing state-of-the-art results on the latter. On the challenging ETIS-LaribPolypDB dataset, the model attains an mDice of 0.949 and an mIoU of 0.902, indicating robustness to small and low-contrast polyps. Cross-dataset evaluation further confirms generalization capability, achieving an mDice of 0.847 when trained on Kvasir-SEG and tested on CVC-ClinicDB, highlighting stability under domain shift. The proposed framework provides an effective balance between segmentation accuracy, computational efficiency, and generalization, demonstrating strong potential for computer-aided colorectal cancer screening systems.

Читать у издателя →

Похожие статьи