Fully Convolutional One-Stage Circular Object Detector on Medical Images

In an endoscopic thyroidectomy, parathyroid injury can lead to hypocalcemia for life. It is of great help to assist surgeons to complete endoscopic thyroidectomy without damaging the parathyroid. Considering that the shape of the parathyroid is usually round, we replaced the rectangle bounding box with the circular bounding box, and propose a fully convolutional one-stage circular object detector (FCOSC) based on FCOS. The loss and optimization are improved to adapt to the calculation of the circular bounding box. FCOSC achieved 37.9% in AP with single-model testing on the parathyroid dataset, which was 5% higher than FCOS. Considering that our data is strictly confidential, we also evaluated the accuracy of our method in the BCCD dataset and achieved 70.4% in AP, comparing to 68.4% with a rectangle bounding box. Experiments show that the proposed FCOSC framework can obtain expected results on parathyroid detection and be used in another metical circular object detection task.