Accurate Segmentation of Surgical Instruments via Spectral-Attentive Contextual Interaction Network

Surgical instrument segmentation is crucial for enhancing visual perception and enabling precise manipulation in robotic surgical systems. However, current segmentation models continue to face substantial challenges in terms of accuracy and robustness due to complex background interference, diverse instrument morphologies, and low contrast between instruments and surrounding tissues in surgical environments. Despite significant advances in deep learning-based approaches, existing models still fall short in capturing the fine edges and global contextual relationships of instruments. To address these issues, we propose a Spectral-attentive Contextual Interaction Network (SCI-Net) for surgical instrument segmentation. Specifically, we present a Global Context Aggregation Module (GCAM) to integrate high-level features, which is used to produce a global map for the coarse localization of the segmented target. Then, a Spectral-enhanced Feature Module (SFM) is proposed to enhance the expression of features in the form of frequency-domain attention by transforming features from the spatial domain to the frequency domain. In addition, we design the Scale-aware Dilation Module (SDM) in the decoder to further adaptively integrate the augmented features through multi-scale dilation convolution combined with a dynamic fusion mechanism, which improves the segmentation performance on the fine boundaries of instruments. We have extensively validated SCI-Net on multiple publicly available surgical instrument segmentation datasets, and the experimental results show that SCI-Net significantly outperforms other state-of-the-art segmentation methods.