Efficiency Meets Accuracy: Benchmarking Object Detection Models for Pathology Detection in Wireless Capsule Endoscopy

Wireless capsule endoscopy (WCE) is a huge step forward in diagnosing problems in the digestive system because it can take pictures without hurting the patient. As part of this study, we thoroughly test anomaly detection models that were created with the specific goal of finding WCE disease. The Single Shot Multibox Detector (SSD), EfficientDet, Faster RCNN, RetinaNet, yolov3, Real-Time Multi-Detector (RTMDet), and semi-supervised learning detection models were tested on a translated MS COCO dataset from the Kvasir-Capsule repository. The dataset has a variety of gastrointestinal abnormalities. We used three main evaluation methods: normalized confusion matrices for class-specific accuracy, macro and weighted averages(precision, recall, and F-score). We compared normal and semi-supervised versions of models like Faster RCNN to find out what the benefits of semi-supervised methods are. The models that were tested showed that RTMDet-S and RTMDet-Tiny were the most accurate in their class, finding complex diseases with over 99% accuracy. With an precision, recall, and F-score of 99.67%, SSD300 also had the best total result among statistics measures. Along with that, YOLOv3 was the model that worked the best in terms of operating economy. It achieved a frame rate of 92.1 frames per second while using only 376MiB of memory. There were big improvements in sensitivity and specificity to 98.70% and 99.84% for semi-supervised models like Soft-teacher (Faster RCNN + ResNet).The purpose of this study is to assist researchers and clinicians in improving their diagnostic abilities by providing highly informative information on the most effective object detection models that can be used for the identification of WCE pathology.