Author: Ruiyan Du, Mingzhu Li, Ying Li, Wei Liu, Shihuan Qin, Yiming Ren, Biao Tu, Hui Xu, Lian Zhang, Xiao Zhang, Zengren Zhao 👨🔬
Affiliation: Department of General Surgery, Hebei Key Laboratory of Colorectal Cancer Precision Diagnosis and Treatment, The First Hospital of Hebei Medical University, Medical AI Lab, The First Hospital of Hebei Medical University, Hebei Provincial Engineering Research Center for AI-Based Cancer Treatment Decision-Making, The First Hospital of Hebei Medical University, Cancer Center, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Department of Radiation Oncology, Mayo Clinic, Department of Oncology, The First Hospital of Hebei Medical University 🌍
Purpose: Fiducial tracking is widely used in CyberKnife to dynamically guide the gantry for moving target like liver cancer stereotactic body radiation therapy (SBRT). This study developed a robust framework based on an improved YOLO algorithm for real-time localization of fiducial markers in X-ray images to evaluate intrafraction motion.
Methods: The study included a dataset comprising 7033 X-ray images from 13 liver cancer patients underwent CyberKnife SBRT treatment fractions. 5626 X-ray images were split as training group, and 1407 for validation group. The deep learning based fiducial tracking framework consists of three modules: (1) A self-developed deep learning framework for fiducial tracking named HFC-YOLO, (2) 3D position reconstruction, and (3) online motion assessment. The framework introduces hierarchical context fusion network (HCF), and the performance of this approach was assessed using metrics such as accuracy, recall, mean average precision (mAP), and calculation time. 3D spatial fiducial positions were further reconstructed for rigid transformation-based motion evaluation.
Results: YOLOv10 and YOLOv11 showed good performance with accuracy 0.951 and 0.987 separately, and the self-developed HCF-YOLO algorithm further improved the accuracy to 0.991 (recall=0.994, mAP=0.992). The mean (±SD) discrepancy between the predicted and actual fiducial centers was 0.23 ± 0.44 pixels, showing its outstanding precision. For translational shift, the mean (±SD) values in the superior–inferior, left–right, and anterior–posterior directions were 12.9 ± 2.1 mm, 2.2 ± 0.5 mm, and 5.6 ± 1.5 mm, respectively. Furthermore, the calculation time is around 2.5 ms, clinically acceptable for real-time fiducial tracking.
Conclusion: This study presented a novel and high-performance real-time moving target tracking framework based on improved deep learning YOLO algorithm in CyberKnife SBRT, and can potentially be used in other moving target tracking workflow in radiation therapy.