Real-time colonoscopic detection and precise segmentation of colorectal polyps via PESNet

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colorectal cancer (CRC) prevention relies on detecting and removing precancerous polyps during endoscopy.
• Current endoscopic procedures lack real-time tools for simultaneous polyp diagnosis and segmentation, impacting adenoma detection and resection precision.

Purpose of the Study:

• To introduce PESNet, a real-time AI framework for standard endoscopy workstations.
• To enable simultaneous frame-level polyp diagnosis and pixel-level segmentation with minimal latency.

Main Methods:

• PESNet integrates polyp diagnosis and segmentation into a single framework, operating at 225 FPS.
• It uses a dynamic "presence of polyp" prompt and an adaptive module to refine boundaries and compensate for visual variations.
• Evaluations were performed on PolypDiag and CVC-12K datasets, measuring latency with TensorRT FP16 on an RTX 6000 Ada GPU.

Main Results:

• PESNet improved diagnostic F1 score from 95.0% to 97.2% and segmentation Dice score from 85.4% to 89.1% on benchmark datasets.
• Clinical impact included a 26% reduction in missed flat polyps and a 15% reduction in residual tumor margins post-resection.
• Achieved end-to-end latency of 12.6 ± 0.3 ms per frame, well within the 40 ms clinical budget.

Conclusions:

• PESNet significantly enhances adenoma detection rates and resection precision in endoscopic procedures.
• The framework's real-time performance and hardware compatibility facilitate integration into standard workflows.
• This technology has the potential to reduce colorectal cancer incidence through improved polyp management.