2361 - Deep Dive into the Performance of the 5D Breathing Motion Model

D. Low¹, D. OConnell², C. Miller³, M. Lauria¹, R. R. Savjani¹, P. Boyle³, L. Naumann³, R. Andosca³, J. P. Neylon³, A. Lee¹, and D. Moghanaki⁴; ¹Department of Radiation Oncology, University of California, Los Angeles, Los Angeles, CA, ²University of California, Los Angeles, Los Angeles, CA, ³UCLA, Los Angeles, CA, ⁴VA Greater Los Angeles Healthcare System, Los Angeles, CA

Purpose/Objective(s): To conduct a deep-dive into the performance of the 5DCT breathing motion model from a clinical dataset Materials/

Methods: The 5DCT free-breathing CT simulation technique has been conducted on 295 patients. In a recent summary of the clinical 5DCT utilization, it was shown that 80% of the acquired 5DCT simulations were used for treatment planning. Some of the 20% of those unused simulations failed during part of their processing. 34 patients from the whole dataset were selected for in-depth analysis, including those with regular and irregular breathing. The scans were processed using a workflow that analyzes the 25 free-breathing CT scans and simultaneous surrogate, a sealed hollow accordion-shaped bellows wrapped around the abdomen. The workflow a) Analyzed the surrogate using the imaged diaphragm dome to calibrate the surrogate in terms of diaphragm position and speed b) Conducted deformable image registration using the current version of DEEDS, c) and Fit the 5D motion model to each lung voxel. The calibrated breathing waveforms were analyzed to determine the overall breathing amplitudes, the breath-to-breath breathing amplitudes, and breathing irregularity (defined as the ratio of the standard deviation of breath-to-breath breathing amplitudes to the mean amplitude), breathing rate variation, and the lung-specific motion model root-mean squared (RMS) error residuals. Each lung voxel had an independent model residual measurement and the mean and 90th percentile errors were investigated, as well as the relative model errors by dividing the RMS model residual by the craniocaudal voxel motion and limiting the reported relative errors to voxels that moved more than 7.5 mm.
Results: The mean breathing amplitude and irregularity were 22.5 mm ± 7.7 mm (12.7 mm-42.4 mm) and 0.2 ± 0.12 (0.05 - 0.44, lower corresponding to more regular), respectively. The average period was 5.9 s ± 1.4 s (4.0 s-10.9 s) and irregularity of 0.26 ± 0.09 (0.10 - 0.41). The median RMS motion model residual error was 1.2 mm ± 0.5 mm (0.5 mm - 2.4 mm) and the worst 90th percentile by volume RMS error was 1.8 mm ± 0.7 mm (0.9 mm - 3.8mm). The 90th percentile relative model errors were 15.0% ± 4.0% (0.1% - 23.9%). The 90th percentile RMS was correlated against the breathing period and breathing amplitude irregularity, with correlation coefficients of 0.41 (p=0.015) and 0.69 (p<0.01), respectively.
Conclusion: The 5DCT relative and absolute errors were relatively small but were well correlated with breathing period and amplitude irregularity. The correlation between model error and breathing irregularity were unsurprising and indicate that the motion model has room for improvement. Additionally, some of the worst performing regions were near the heart due to the uncorrelated cardiac motion, and the addition of an ECG could aid in those regions. We will expand this deep dive to all 295 patients and add tumor-specific statistics to determine the overall failure rate and attempt to create the next-generation motion model.