PolyDetect: A Quality-Aware Hybrid Deep Learning Framework for Robust Polyp Segmentation in Degraded Colonoscopy Images

<title>Abstract</title> <p>Colorectal cancer (CRC) is one of the most common cancer-related deaths, and colonoscopy is the preferred method of screening for CRC. In practice, the images captured by endoscopes are often corrupted by motion blur, noise, compression artifacts, low contrast and poor illumination, adversely affecting the existing segmentation models and only a few algorithms explicitly use the image quality signal within the segmentation pipeline. We suggest a hybrid scheme, PolyDetect, which combines five components: a five-type degradation simulator for robustness training, a no-reference image-quality assessment (IQA) module combining BRISQUE statistics with a lightweight CNN (~16K parameters) to form a per-image quality score, a quality-score-gated three-tier adaptive enhancement pipeline (Retinex, CLAHE, DnCNN), a PVT-v2 encoder with a CNN decoder and a dedicated boundary-refinement head, and a quality-weighted composite loss to down-weight degraded training samples. PolyDetect is trained on 900 Kvasir-SEG images and evaluated zero-shot on CVC-ColonDB (n=380) and ETIS (n=196) datasets, and achieves the highest Dice (0.8066) and IoU (0.7262) on all three tests, improvements statistically significant after multiple-comparison correction over CNN-based baselines. However, the ablation of the boundary-refinement head was found as the most critical component to remove (−6.47 pp Dice), while the ablation of the quality-gated enhancement pipeline results in an improvement of mean Dice by 0.94 pp in 91.8% of the frames with quality score q ≥ 0.5. The segmentation network has an average FPS (frames per second) of 53.62 on a Tesla T4 GPU (27.35M parameters, 12.47 GFLOPs), which is in the real time range for clinical application. PolyDetect is a feasible way to achieve a comprehensive quality assessment, quality-gated enhancement, and quality-weighted training in the same pipeline for polyp segmentation in degraded, real-world colonoscopy images.</p>