Improving Detection of Capsule Endoscopy Using YOLO

Document Type : Original Article

Authors

1 Faculty of Computer Engineering and Information Technology, Sadjad University, Mashhad, Iran

2 Sadjad University /member of computer faculty

Abstract

Capsule endoscopy (CE) technology is rapidly advancing due to its easy usability, long battery life, and exceptional image quality. However, the increasing clarity of image sequences captured by CE requires more time and effort to detect desired content. To address this issue, a new approach is presented in this paper using the popular YOLO v5 neural network architecture to detect the location and label of lesions in two public CE contents. A GPD neural network based on AlexNet is used as a rival classifier. The primary goal of this research is to reduce diagnostic time while maintaining accuracy using YOLO, and the results show a 6% increase in detection accuracy over the rival. Additionally, YOLO is 58% more time-efficient with an average prediction time of 5.39 milliseconds per frame. The scalability of YOLO is also analyzed, and results indicate a 6.95 times graceful degradation over Kvasir, proving YOLO's real-time applicability. Higher resolution inputs lead to better results with YOLO. Implementations and supplementary data are available on GitHub.

Keywords

TABRIZ JOURNAL OF ELECTRICAL ENGINEERING

Articles in Press, Corrected Proof
Available Online from 19 June 2024

Files

Share

How to cite

Statistics