RVDNet: A Reference-Guided Video Desmoking Network for Endoscopic Thyroid Surgery

Video desmoking in endoscopic thyroid surgery presents unique challenges due to confined anatomical structures and rapid accumulation of dense, spatially irregular smoke. While recent deep learning-based methods have shown promising results for smoke removal in natural and laparoscopic scenes, few address the distinctive smoke patterns and constrained anatomy characteristic of thyroid surgery. In this paper, we propose RVDNet, a reference-guided recurrent network tailored for thyroid surgical video desmoking. It leverages the first typically smoke-free frame of each video clip both as supervision to avoid reliance on scarce paired training data and as a reference input to guide the restoration of subsequent frames. To further improve desmoking effectiveness, we propose the CRAFF (Cross-frame Reference Attention-based Feature Fusion) module, which utilizes cross-frame attention to incorporate reliable structural information from the reference. Additionally, we construct EndoClear, the first real-world dataset for endoscopic thyroid video desmoking. Extensive experiments using five-fold cross-validation demonstrate that RVDNet outperforms state-of-the-art methods in both quantitative metrics and perceptual quality, removing smoke more effectively while preserving temporal coherence.