Prog Med Phys.  2020 Sep;31(S1):S61.

Retrospective Analysis of Lung Tumor Position Prediction Using MR Image for X-ray Image-Guided Adaptive Radiotherapy

• Shin J
• Ji Z
• Huang Y
• Benjamin L
• Guta A
• Oh S
• Mazur T
• Kim T
• Kim JS

Abstract

Purpose
To provide improved tumor positional information from on-board x-ray images, we developed a statistical model to predict tumor motion from surrogates visible on x-ray imaging, for example diaphragm, from MR images acquired on an MRgRT system, which provide excellent soft tissue contrast and real-time tumor tracking.
Methods
Cine MRI was acquired during treatment for real time tumor tracking in the standard MRgRT clinical workflow. Auto-segmented target contours for tracking and gating were generated by the system on sagittal cine MRI images. Inhouse software was developed to generate surrogate contours on MR images using a template-based auto-segmentation method. Two approaches for modeling the correlation between the position of the tumor and surrogate were considered. The first approach utilized traditional regression analysis between the position of the tumor center and the dome of the diaphragm and linear regression was used to develop a simple linear model. A second approach used principle component analysis (PCA) to apply multiple input data sets, such as multiple time points or surrogates, for model generation and all diaphragm dome positions from a single fraction were used for PCA, including those that occurred in later cine MRI frames. In total 85,214 cine MRI images from 53 fractions for 3 lung cancer patients were analyzed retrospectively. These models were made fraction by fraction for the intrafractional motion prediction. These models were trained and tested using 70% and 30% of cine MRI images, respectively.
Results
The first approach, the traditional linear regression model utilizing ordinary least squares (OLS), predicted the y-position of the tumor center from the y-position of the diaphragm with a mean square error (MSE) of 1.73 mm. The second model from PCA predicted the tumor center with the maximum average error of 1.35 mm with standard deviation 0.72 within each fraction. In our study, analysis focused on the motion of tumor and diaphragm in the superior-inferior direction due to negligible variation, within a subpixel of the image, in the anterior-posterior position of the tumor and diaphragm on sagittal images.
Conclusions
We developed a model to predict tumor motion from an anatomical surrogate, visible on MR imaging, using sagittal cine MRI. In the preliminary result, the tumor motion can be predicted using a simple linear model within a couple of mm subpixel difference and accuracy can be further improved with a more complex modeling method. This study is still on-going to develop a more accurate model with an increased number of patients and the inclusion of additional imaging modalities, for example on-board X-ray imaging.

Keyword

MR-guided radiotherapy (MRgRT); Cine MRI; Surrogate; Linear regression; Principle component analysis (PCA)