YOLO networks for polyp detection: A human-in-the-loop training approach

Abstract

Introduction: Early detection of adenomas and polyps is one central goal of colonoscopic screening programs. As the adenoma detection rate (ADR) depends on the experience of the endoscopist, AI-based polyp detection systems can be used for real-time assistance. Hence, to support the physicians such AI-based systems using deep-convolutional neural networks (DCNNS) have been introduced in the past years. One disadvantage of these techniques is the need of a huge amount of labeled training data. Method: We investigate a "human-in-the-loop approach" to minimize the required time to generate labeled training data. The approach is evaluated within the training a YOLOv4 neural network to detect polyps in colonoscopic image data. The performance metrics of the neural network are evaluated on three public datasets. Results: The performance of the YOLO network increased from a precision of 0.88, recall of 0.83, F1 score of 0.86, and a F2 score of 0.86 to a precision of 0.91, recall of 0.87, F1 = of 0.89, and F2 = of 0.88. The interactive labelling of 1,000 images only takes one hours. Conclusion: The proposed "human-in-the-loop approach" is capable of generating labelled image data in a minimum of time while increasing the performance metrics as well. For higher performance increase more data can now be labeled within this new approach.