Comparative Analysis of Machine Learning Models for Colorectal Polyp Detection

Biomedical Engineering

This comparative analysis explores the efficacy of four prominent machine learning models—VGG16, FCN (Fully Convolutional Network), DUCK-Net, and YOLO (You Only Look Once)—for the detection of colorectal polyps using the CVC-Clinic DB dataset. Colorectal polyps that are greater than 1 cm are more likely to cause colorectal cancer. Early detection is crucial for effective treatment and patient outcomes. The study evaluates each model's ability to accurately identify polyps from medical images, considering metrics such as precision, recall, F1 score, dice coefficient, mIoU and computational efficiency. VGG16, which is well-known for its deep architecture and heavy reliance on convolutional layers, is excellent at extracting features but may have issues with processing power because of its high number of parameters. Specifically engineered for semantic segmentation tasks, FCN provides accurate polyp localization at the pixel level in pictures, potentially yielding higher spatial accuracy than previous models. With its focus on robust feature extraction and disease-specific pattern recognition, DUCK-Net—a medical image analysis specialist—may be better able to identify minute polyp features in the CVC-Clinic DB dataset. With its real-time object detection capabilities, YOLO puts efficiency and speed first, which is essential for swiftly processing a lot of medical photos in a clinical context. This work attempts to shed light on the advantages and disadvantages of these models for colorectal polyp diagnosis by means of a thorough assessment and comparison. The research intends to advise healthcare practitioners and researchers on the best machine learning frameworks to choose for improving automated diagnostic systems by examining performance indicators and computing needs. In the end, enhancing polyp detection efficiency and accuracy can help advance colorectal screening initiatives and enhance patient outcomes.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.