(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: In current image-guided radiosurgery clinical practice, the image registration-based method for patient setup and intra-fractional tracking is often inadequate, especially when the target involves multiple vertebral bodies. Here, we propose a fiducial-free targeting strategy using a deep learning approach to interpret routine live kV X-ray images for spinal radiosurgery.
Methods: We designed a patient-specific deep learning model for targeting the spine tumor for a given pair of input orthogonal live images acquired periodically throughout treatment. The model encompasses a target proposal network and a detection and refinement network. Both networks share the same features extracted using the VGG16 network. We trained this model using thousands of annotated digitally reconstructed radiographs (DRRs) which were generated using motion-incorporated planning CT images. To assess the accuracy of the approach, we retrospectively studied images from 12 treatment fractions of 7 patients who received spinal radiosurgery with CyberKnife. The tumor sites varied from C6 to S1. For each fraction, quantitative analysis on testing DRRs and real-time live X-ray images during treatment delivery was performed.
Results: The proposed model achieved a targeting error on testing DRRs range from -1.83 mm to 2.28 mm and from -1.45 mm to 1.23 mm for the two orthogonal X-ray systems. The overall mean absolute targeting errors for the two X-ray systems are 0.74 mm and 0.58 mm, respectively. For real-time tracking periodic live images, targeting results provided by the patient-specific model are consistent with the derived positions from the couch correction.
Conclusion: This study demonstrates that a deep learning model trained using motion incorporated DRRs can accurately targeting spinal tumors for radiosurgery without fiducials. The method may be useful for patient setup and real-time target tracking. It has the potential to provide a clinically valuable solution for routine spinal radiosurgery and could be adapted to other challenging treatment sites.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
