Detection of Colon Polyps in Colonoscopy Images Using Deep Learning: Methodological Study

Objective: This study compares deep learning models, including the You Only Look Once (YOLO) family [v5, v8, v9, Neural Architecture Search (NAS)] and Roboflow 3.0, for automated polyp detection in colonoscopy images. We evaluate their accuracy and efficiency for real-time clinical application to improve early detection, a critical step in colorectal cancer prevention. Material and Methods: We retrospectively analyzed 342 images with 478 histopathologically confirmed polyps from 196 patients. After labeling by expert gastroenterologists, the dataset underwent extensive preprocessing and augmentation before being split into training (88%), validation (8%), and testing (4%) sets. Model performance was evaluated using recall, precision, and mean average precision (mAP). Results: The Roboflow 3.0 model demonstrated superior performance, achieving the highest recall (86.5%), precision (91.4%), and mAP (92.3%). In contrast, YOLOv5, despite being the fastest to train, yielded lower accuracy (80.1% recall, 58.4% precision). YOLO-NAS presented a balanced trade-off between speed and accuracy (65.6% recall, 73.4% precision), making it a practical choice for time-sensitive settings. Conclusion: This study validates the efficacy of modern deep learning models, particularly Roboflow 3.0, for automated polyp detection. The results support clinical integration to assist endoscopists, reduce miss rates, and enhance early diagnosis. Future research should validate these models on larger, diverse datasets and explore advanced data augmentation to improve robustness for widespread clinical adoption.