Optimized Computer Vision Model for Accurate Polyp Detection in Endoscopic Procedures

Abstract

Colorectal Cancer (CRC) is the main reason for cancer-linked morbidity and death globally, and early recognition has an important responsibility in enhancing patient endurance rates. Detecting polyp’s precursors to CRC significantly reduces mortality when identified in the early stages. The data is gathered from endoscopic video data from publicly available datasets. The preprocessing pipeline includes Contrast-Limited Adaptive Histogram Equalization (CLAHE) to enhance image contrast, followed by Histogram of Gradients (HOG) for feature extraction. This research introduces a framework for concurrent polyp detection in endoscopic videos utilizing advanced computer vision techniques, specifically the Adaptive Masked Cuttlefish Region Convolve NeuroNet (AMC-RCN). This hybrid model integrates the strengths of Mask Region Convolve NeuroNet (R-CNN) and Adaptive Cuttlefish Optimization (ACFO) to achieve precise and efficient polyp detection. The Mask R-CNN component utilizes Region Proposal Networks (RPN) to accurately locate polyps, generating bounding boxes and pixel-wise segmentation masks. The ACFO algorithm further refines the model by optimizing hyper-parameters, improving segmentation boundaries, and selecting the most relevant features from the endoscopic frames, ensuring optimal performance. The AMC-RCN framework effectively handles small and irregular polyps, ensuring high segmentation (98.02%), precision (97.91), F1-score (96.97%), and recall (97.07%) even in complex and challenging scenarios. The model is evaluated on prominent video datasets, providing a comprehensive set of endoscopic video footage for rigorous testing. The framework demonstrates superior detection accuracy, faster training convergence, and robust performance in clinical applications.