Abstract
Purpose/Objective(s): Accurate treatment verification is of particular importance when delivering high doses per fraction, as in lung SBRT. In this study, we characterize and validate the performance of a real-time markerless lung tumor-tracking algorithm, using continuous MV portal imaging. MV portal imaging is advantageous, as beam’s-eye-view, images allow motion capture in both directions of steep dose fall-off, and the patient is not exposed to additional imaging dose. Furthermore, a markerless tracking system avoids the risk associated with implanting fiducial markers. Materials/Methods: We retrospectively analyzed the delivery of 9 lung SBRT treatments carried out at our institution. Radiation therapy was delivered with a 6 MV beam in 3 fractions to 54 Gy or in 5 fractions to 60 Gy, using 9-11 beams, including non-coplanar geometries. Treatments were monitored by continuous acquisition of portal images with a frame rate of 2 Hz. Soft tissue motion was tracked on the portal images with a multi-region algorithm developed for this purpose. For each field and each fraction a set of landmark candidates was automatically defined on the first image. The tracker then identified this set on each subsequent image by means of a similarity measure and calculated a centroid trajectory. In this way, the images were analyzed in a prospective manner to simulate in-treatment real-time analysis. The first 40 images of 54 fractions from 34 beam angles were used for comparison with expert human examiners. To minimize bias, the images were presented to the examiner in random sequence and each image was shown 3 times. Tracking error in a realistic phantom was determined as a baseline. Computation time was measured for real-time processing of the images. Results: In-treatment EPID images were successfully acquired for 9 lung SBRT patients with 3 or 5 fractions each (19,942 portal images), including non-coplanar geometries. By comparing with the manual tracking, the RMS tracking accuracy of the automatic marker-less algorithm was found to be (2.1 1.7) mm. Excluding the worst 9 series (out of 54) reduced the RMSE to (1.5 1.1) mm. If we define tracking success as having a RMSE <3 mm, the success rate in our data was 86%. Without any tracking, the error would have been (3.9 2.7) mm. For the phantom study, the tracking error was <1 mm at all times, with a RMSE of (0.8 0.2) mm. The latency of the current implementation is 70-80 ms. Conclusions: In order to verify geometric fidelity during lung SBRT treatments at our institution, we have developed and tested a robust, prospective, real-time algorithm to track lung tumors on portal image sequences without the use of fiducial markers or user intervention. This algorithm can be implemented as a clinical tool for treatment verification, to derive delivered dose, or to drive a real-time motion compensation protocol. Author Disclosure: J. Rottmann: None. Y. Yue: None. A. Chen: None. D. Kozono: None. R. Mak: None. F. Hacker: None. R.I. Berbeco: None.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have