Answers are grounded in this paper and cite their sources.
Accurate polyp size measurement during colonoscopy is essential for colorectal cancer diagnosis; however, monocular depth estimation is significantly affected by complex lighting variations, including mucus reflections, exposure inconsistencies, and synthetic-to-real domain gaps. To address these challenges, we propose LAM-Soft, a semi-supervised framework that integrates a dynamic Lighting Adjustment Model (LAM) with confidence-aware learning. The proposed LAM converts rigid pseudo-labels into soft constraints by modeling near-field light attenuation and surface reflection, facilitating the alignment between synthetic per-pixel shading features and real clinical grayscale distributions. A teacher–student architecture is adopted to jointly exploit hard labels from synthetic datasets and LAM-refined soft labels from clinical data, while a confidence decoder suppresses low-quality frames caused by motion blur or contamination. Experimental results on the C3VD dataset demonstrate that LAM-Soft achieves state-of-the-art performance with an L1 error of 0.1789 mm and a δ-accuracy of 57.19%. For polyp size estimation on the SUN database, the proposed method reduces the mean absolute error to 0.819 mm. Ablation studies verify the effectiveness of the proposed components. In addition, the lightweight architecture supports real-time inference at 71 FPS, indicating its potential for practical clinical deployment.
No takes yet. Share an insight, caveat, or question.
Wu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: