The Evaluation of Bone Fracture Detection of YOLO Series
Pongsakorn Samothai, Parinya Sanguansat, Adisorn Kheaksong, Kanabadee Srisomboon, Wilaiporn Lee
Abstract
Convolutional Neural Network (CNN) becomes the most powerful tool for object detection application and is mostly used in several filed such as security and medical. Bone fracture detection is one of the most medical application which will be useful in the medical treatment since there are number of misdetections according to human errors and small fracture locations. YOLO is the advanced CNN model that is interesting for researchers to be implemented with several applications. The YOLO-X and YOLO-R are the current models which are implemented in many applications. However, these models have not been evaluated on the low feature images such as bone X-ray images. Therefore, in this paper, we comprehensively study the performance of YOLO-X and YOLO-R under fracture bone X-ray images to locate the fracture locations. As a result, YOLO-X presents faster convergence time with much higher accuracy than YOLO-R since the YOLO-X model locates the fracture area with alternative processing such as detection head decoupling, anchors-free and augmentation strategies. Then, it can detect the fracture locations even if the features of X-ray images are low.