Gastric Polyps Detection by Improved Faster R-CNN

Abstract

This paper presents the research results of detecting gastric polyps with deep learning object detection method in gastroscopic images. In this work, we use an improved Faster R-CNN network to detect the gastric polyps. We use the ROI align operation to replace ROI pooling operation, use the GIoU loss to replace the original smooth L1 loss and use the soft-NMS to replace the traditional NMS in the Faster R-CNN network. The ROI align operation can solve the problem of misalignment. This GIoU loss can take the IoU between the predicted value and the ground truth value into account to the greatest extent and improve the detection performance. This GIoU loss in the detection network will effectively improve the accuracy of the box regression. The soft-NMS can reduce the deletion of bounding boxes by mistake in the post processing stage. The Faster R-CNN network not only achieves good results in general image detection, but also in gastric polyps image detection. The improved Faster R-CNN can further improve the detection performance in the gastric polyps. Compared to the other polyps detection methods, precision, recall rate and F1 score of our network has been achieved higher values. The final detection results about precision, recall rate and F1-score of our work is 78.96%, 76.07%, 77.49%. In gastric polyps detection, this method can be of great help to doctors and patients.