Towards Robust Polyp Detection: Self-Supervised Learning with Noise-Reduced Features and Enhanced Borders

Colorectal cancer remains one of the most lethal malignancies, underscoring the importance of automated polyp detection for early diagnosis. Although recent multi-scale deep learning models have demonstrated success by integrating local and global feature extraction, they face notable limitations: (1) noisy feature extraction caused by white light reflections and inconsistent feature scales, (2) insufficiently defined border information leading to higher miss rates for flat and irregularly shaped polyps, and (3) limited availability of labeled training data, which hinders generalization and convergence. To address these challenges, we propose a novel polyp segmentation network incorporating a noise mask-guided feature denoising module and a finite-differences-based border enhancement module to improve feature quality and border representation without excessive computational overhead. Furthermore, we introduce a selfsupervised pretext task that extracts meaningful representations from unlabeled colonoscopy images, enhancing model performance on the downstream polyp segmentation task. Our model, pre-trained on LDPolypVideo and fine-tuned on Kvasir-SEG and CVC-ClinicDB, demonstrates state-of-the-art performance across four public datasets (ETIS-LaribPolypDB and CVC-ColonDB), surpassing existing approaches in Intersection over Union (IoU), Dice Coefficient, and computational efficiency.