Purpose: Extracting the phase pattern of respiration-induced tumor motion using cone-beam computed tomography (CBCT) projections is challenging since anatomic structure obstruction results in poor tumor visibility of projections. Predicting tumor phase pattern using external surrogates also has intrinsic difficulties, as the phase patterns between surrogates and tumors are not necessarily congruent. We developed an algorithm that is able to recover the primary oscillation components of tumor motion using the combined information from CBCT projection images and external surrogates.
Methods: The algorithm involves two steps. First, a preliminary tumor phase pattern is acquired by applying a Local Principal Component Analysis (LPCA) algorithm on Amsterdam Shroud images of tumor regions. Then, by performing a Multivariate Singular Spectrum Analysis (MSSA) on both the surrogate information and the preliminary pattern, primary components of phase oscillations can be recovered. A Quasar 4D phantom and optical tracking system (OTS) were employed to acquire projection images and the surrogate ground truth. The phase shifts or arbitrary amplitude changes was simulated to examine the robustness of the MSSA algorithm. Anatomic obstructions were imitated by attaching simulated objects to the phantom with patient breathing waveforms to mimic real clinical scenarios.
Results: The LPCAâ€“MSSA is able to remove the high-frequency noise components caused by anatomic structure obstructions, while preserving the oscillation signal of the primary components. With phase shifts and amplitude variations, the overall peak and valley accuracy was -0.009Â± 0.18 seconds and no time delay was found. For the anatomic obstruction experiments, the extracted LPCA â€“ MSSA signal had an expiration phase deviation of 1.59Â± 1.98 %, peak accuracy of -0.13Â±0.28 seconds, and valley accuracy of 0.0Â±0.16 seconds.
Conclusion: With the new LPCA-MSSA method, and the aid of the extra information of external surrogates, prediction accuracy was greatly improved as compared to using the LPCA algorithm alone.