Detection of Spinal Fracture Lesions based on Improved Yolov2
Gang Sha, Junsheng Wu, Bin Yu
Abstract
Yolo algorithm has a good detection effect in target detection. Because of its high detection accuracy and fast detection speed, it is widely used in practice. Because of the problem that the complexity of spine CT images, the irregular shape of vertebral boundary, meanwhile there are artificial deviations and low efficiencies in clinic, which needs doctors' prior knowledge and clinical experience to determine lesions location in CT images, so it can not meet the clinical real-time needs. In this paper, We use deep learning to process the CT images of spine, and to detect and locate lesion of (cervical fracture, cfracture), (thoracic fracture, tfracture), (lumbar fracture, lfracture) by the improved YOLOv2. Through using lesions bounding box dimensional cluster and multiscale transformation of input CT images to improve detection efficiency and accuracy. The experiment shows the results are more accurate, and mAP (mean average precision) of detection algorithm is 75.30%, detection rate is 0.027 seconds per detection, and IOU((Intersection-over-Union) is 77.3, which can basically meet the clinical real-time needs.