Purpose: Real-time tumor tracking is a key feature of MRgRT. However, uncertainty in real-time tumor contours from low-resolution MRI acquired on a 0.35T MR-Linac system can lead to inaccurate beam delivery. To assist tumor tracking capabilities, we developed intra- and inter-fractional models for tumor position prediction from anatomical surrogates on 0.35T cine MR images.
Methods: Cine MR images were acquired for real-time tumor tracking in standard MRgRT clinical workflow. Images included the tumor contour produced by the MRgRT system. Surrogates were automatically contoured using in-house software. Models were developed using linear regression and principal component analysis (PCA) to predict tumor motion from diaphragm and abdominal contours. A total of 1,069,071 images from 53 fractions for 10 lung cancer patients were analyzed retrospectively. The model for intra-fractional prediction was trained and tested on each fraction. The inter-fractional prediction was tested on each fraction using the patient’s remaining fractions for training. The mean square error (MSE) and mean absolute error (MAE) between predicted and true position of the tumor were used for evaluation.
Results: The regression model predicted the superior-inferior and anterior-posterior position of the tumor center from the diaphragm contour and abdomen contour with an average MSE of 0.57 and 0.54 mm2 and an average MAE of 0.54 and 0.81mm, respectively. The PCA model predicted the tumor center with an average MSE of 1.15 and 1.70 mm2 and an average MAE 0.80 and 1.24 mm, respectively. These errors were within two pixels on sagittal images.
Conclusion: We developed models to predict tumor motion from anatomical surrogates using sagittal cine MRI. In this study, the models can predict the tumor motion using a linear model within a couple of millimeters. This study is on-going to apply the models to other treatment sites and additional imaging modalities, for example, on-board x-ray imaging.