Purpose: Spine implants has been shown to decrease local control for spinal sarcoma and chordoma patients because of artifact-induced dose calculation inaccuracies. Artificial intelligence (AI)-based methods have shown advantages for metal artifact reduction (MAR) by assuming that prior knowledge of implants is available. We propose a novel component method to accurately delineate implants from CT images to train AI-based models that potentially can increase image accuracy for proton Monte Carlo dose calculation.
Methods: We propose a novel method to identify implant components (tulip, screw & rod), utilizing various levels of prior knowledge such as CT scanned implant components and implant volumes and screw lengths from medical records. The tulips from two major vendors are manufactured with different materials such as titanium and Chromalloy. With the use of extended Hounsfield units (HU), and a fine resolution reconstruction algorithm, the implant component method was developed and demonstrated for various levels of prior knowledge both in a phantom and in vivo for one patient.
Results: Implant systems from two major vendors were analyzed. It was found that mischaracterization of 8 typical implants in a patient can cause under coverage of a clinical target volume (CTV) around 20 cm³, consistent with excessive proton range uncertainties of 10 mm reported in the literature. Using our method, tulip, the most important implant component, can be characterized with proper material and volume as well as other components. Intended CTV coverage can be maintained within tolerance.
Conclusion: Our component method provides accurate implant characterization for proton dosimetry and can potentially enhance the accuracy of AI-based MAR models. The current study primarily focuses on the implant characterization for lumbar spine. Future work will include cervical spine and dental implants for head-and-neck patients where tighter margins are required due to proximity of organs-at-risk.