A Novel Markerless Target Tracking Pipeline Enhanced By Computer Vision Algorithms For Lung Cancer Radiotherapy

During the process of radiation treatment for lung cancer,technologies have been developed to track the intrafraction movements of GTV target due to respiration,via the k V on-board imaging system mounted on commercial Linac.One of which is the Markerless Tumor Tracking(MTT)methods that seek ways to directly locating the tumor on the x-ray fluoroscopic images using the template matching algorithm and have already achieved the accuracy required in the clinical environment.But the need to scan a single fluoroscopic image with multiscale sliding windows leads to high computational complexity.Moreover,to solve the major tumor-bone overlapping problem with MTT,additional image processing is introduced to suppress the bone tissue background signals,making those methods less capable of being real-time.In recent years,the rapid development of deep learning techniques has brought new blood into the object detection task of computer vision,leading to remarkable breakthroughs on the speed and precision of locating objects on natural images.However,these object detection algorithms have not been tested on x-ray fluoroscopic images for lung target tracking.In this study,we aimed at evaluating both the accuracy and speed of a novel MTT pipeline enhanced by computer vision algorithms(Faster R-CNN and Yolo V3)for lung cancer radiotherapy and verifying the feasibility for object detection frameworks to be applied in real-time target locating.To characterize the performance of the Faster R-CNN framework,we initially use a CIRS dynamic thorax phantom to emulate relatively simple cases.Three simulated training sets with the amount of 25,000 have been generated to train the model,in which the bone tissue impact level(Index B)was varied.For all the successful tracking cases,the lowest locating I-S errors for three sizes of tumor targets were 0.92mm,0.85mm,and 0.43mm,corresponding to 12mm,18mm,and 27mm,respectively,when the tumor is free from being blocked.Overall end-to-end processing time of a single fluoroscopic image was around 220ms(～4.5FPS),which is 5 times less than that of traditional MTT workflow.Despite the decent results,apparent locating failure did happen.Increasing Index B could reduce the failure rate to less than 10%.More sophisticated experiments were done on 3D printed anthropomorphical phantom,using the Yolo V3 framework.The amount of training data was augmented to 100,000.For the target 1 in the middle part of the right lung,the distribution of the deviations from locating positions to manually annotated ones was fit by 2D Gaussian,of which the results were μu=0.58mm,μv=-0.35mm;σu=1.7mm,σu=2.4mm.For the target 2 in the cranial part of the right lung,the same type of results were μu=0.74mm,μv=-1.32mm;σu=1.9mm,σu=1.47mm.furthermore,the estimated 3D position of targets has an average error of 1.22mm.Noticeably,tracking failure almost disappeared(<1%)in the experiments’ configuration.Overall end-to-end processing time was further reduced to 60ms(～16.7FPS),which is 15 times less than that of traditional MTT workflow.In conclusion,the object detection algorithms in computer vision have a bright prospect in real-time lung target tracking tasks for its unparalleled speed and balanced accuracy.