Real-Time Colonoscopic Detection and Precise Segmentation of Colorectal Polyps via PESNet

Jing Yu¹Jianchun Zhu²Qi Gu³Yuhan Sun³Qin Wang⁴Pengcheng Sun^2*Liugen Gu¹

• ¹Department of Gastroenterology, Nantong First People’s Hospital, Nantong, China
• ²Suzhou Xiangcheng People's Hospital, Suzhou, China
• ³Nantong University School of Medicine, Nantong, China
• ⁴The Third People's Hospital of Nantong, Nantong, China

Precise and timely visual assistance is crucial for detecting and completely removing colorectal cancer precursor polyps, which is a key step in preventing interval cancer and reducing patient morbidity. PESNet is a real-time assistance framework that can simultaneously perform frame-level diagnosis and pixel-level polyp outlining at 225 FPS on standard endoscopy workstations, with minimal additional latency and no specialized hardware required. This method dynamically injects a simple "presence of polyp" prompt into the segmentation stream, real-time refinement of lesion boundaries, and automatically compensates for changes in lighting and mucosal texture through a lightweight adaptive module. On the PolypDiag and CVC-12K benchmark datasets and replay resection scenarios, PESNet improved diagnostic F1 from 95.0% to 97.2% and segmentation Dice from 85.4% to 89.1%, equivalent to a 26% reduction in missed flat polyps, and a 15% reduction in residual tumor margins after cold snare resection. End-to-end latency (1080p) is 12.6 ± 0.3 ms per frame (TensorRT FP16, RTX 6000 Ada), of which segmentation accounts for 4.4 ms, prompt fusion 0.6 ms, and prototype lookup < 0.2 ms, satisfying a 40 ms clinical budget with > 3× headroom. These clinically significant improvements demonstrate that PESNet has the potential to enhance adenoma detection rates, support cleaner resection margins, and ultimately help reduce the incidence of colorectal cancer during routine endoscopic examinations.