Stage-Aware Swin-Enhanced nnU-Net v2 for Robust Polyp Segmentation in Collaborative Endoscopic Visual Sensing

Automatic colon polyp segmentation is important for reliable endoscopic visual sensing. However, segmentation performance can be degraded by ambiguous lesion boundaries, heterogeneous appearance, and image quality variations during acquisition or transmission. This study proposes a stage-aware Swin-enhanced nnU-Net v2 framework, where Swin Transformer blocks are inserted into selected encoder stages while preserving the original nnU-Net v2 pipeline. Different insertion strategies were evaluated on an independent Kvasir-SEG test set, and robustness was further assessed under six synthetic corruption types with three severity levels. The results show that the insertion stage strongly influences the effectiveness of Swin enhancement. Among the evaluated variants, Stage5-Swin achieved the best overall trade-off between segmentation performance, robustness, and computational cost. It achieved the highest Dice scores across all corruption–severity combinations while maintaining comparable external-domain performance on CVC-ClinicDB. Additional FedAvg experiments demonstrated the compatibility of the proposed architecture with collaborative training workflows. Resource analysis further quantified the computational and communication overhead. The findings indicate that middle-to-deep encoder insertion provides a favorable balance between contextual modeling, spatial representation, and efficiency for robust endoscopic segmentation. However, improvements were metric- and corruption-dependent, and moderate overexposure revealed a Precision–Recall trade-off; practical deployment also remains to be validated.