(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Adaptive imaging and gated radiotherapy treatment involves making decisions in response to patient respiratory phase. Therefore, the ability to accurately determine phase is paramount, but is currently constrained by the respiratory prediction horizon that contains fiducial markers. A variety of prediction methods exist however, they struggle to predict beyond a 500 ms horizon and are unable to predict breathing turning points. The aim of this work was to investigate if a 5 second respiratory horizon can be used to extract the real-time respiratory phase.
Methods: A 24 lung cancer patient study with ~22 hours of RPM data and CAPNOBASE (42 patients, ~6 hours) datasets were fused, resulting in a dataset containing ~2.3 million training/validation examples. A machine learning (UNET) architecture was trained using reinforcement learning to predict a 5-second respiratory window. The model performance was tested on a 4DCBCT dataset consisting of 10 participants, 2 fractions each. We assessed the performance of the UNET in terms of RMSE and Pearson correlation to the actual respiratory displacement and phase. For comparison we used a state-of-the-art phase prediction method.
Results: We achieved a 3x prediction horizon improvement compared to state-of-the-art (1500ms vs 500ms) and Pearson correlation r>0.6. The UNET had a good (r>0.6) predictive ability up to 1.5 seconds. For a five second prediction window, the RMSE between actual and derived respiration displacement was 0.29, r=0.58. However, in terms of phase prediction, the UNET did not offer any benefit compared to state-of-the-art method with RMSE of 0.084 vs 0.019.
Conclusion: This is the first time a UNET has been used to predict long-range respiratory displacement from which phase was extracted. Combining a current state-of-the-art phase predictor with the UNET may offer the ability to mitigate against latencies found in adaptive imaging technologies while maintaining a high degree of phase accuracy.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by grant 1123068 awarded through the Priority driven Collaborative Cancer Research Scheme and funded by Cancer Australia. RO would like to acknowledge the support of a Cancer Institute of NSW Career Development Fellowship
Keywords
Respiration, Cone-beam CT, Computer Software
Taxonomy
IM- Dataset Analysis/Biomathematics: Machine learning
Contact Email
