ESTRO 38 Abstract book

S214 ESTRO 38

Respiratory motion models can potentially estimate the motion of the full 3D anatomy using internal surrogate signals extracted from the real-time 2D images. In this study, we used 2D motion models to quantitatively evaluate potential surrogate signals with data from lung cancer patients. Material and Methods Surrogate cine-MR images from a fixed location and with sagittal orientation were used to generate surrogate signals by tracking the diaphragm motion, and performing principal component analysis (PCA) on image intensities, or PCA on deformation fields (DFs) resulting from deformable registration of the images. The surrogate images were interleaved in time with motion images from another fixed location. The motion images had sagittal orientation for 8 datasets and coronal orientation for 4 datasets. Deformable image registration was used to measure the motion in the motion images relative to an end-exhale reference image. The registration algorithm can account for sliding motion by taking a sliding interface (manually defined for each dataset) as input. The motion images were divided into a training set and a test set. Linear correspondence models were fit to the motion measurements using a variable number of training images and the corresponding interpolated surrogate signals. The models driven by the surrogate signals were used to estimate the motion for the test images. The different surrogate signals were evaluated by calculating the deformation field error (DFE) which is the difference between the DFs estimated by the model and the DFs from the sliding registration. Results The motion models are able to model the sliding motion. Figure 1 shows an example of the estimated DFs for a sagittal and a coronal dataset respectively. Figure 2 reports the mean of the L2 norm of the DFE against the number of training images for the different surrogate signals. For each dataset the mean DFE was computed over all pixels of all test images within the sliding interface (excluding the gastrointestinal organs) and averaged over all datasets for each slice orientation. The mean DFE when not using a motion model was equal to 3.41 mm for the sagittal datasets and 4.54 mm for the coronal datasets. The models for the different surrogate signals all give good results when using 8 or more training images, with mean DFEs below the in-plane pixel size (1.98 mm).

Conclusion The investigated surrogate signals are suitable to model the 2D motion, including sliding, in both sagittal and coronal planes, using few training images. Future work will extend the study to 3D motion models from multi-slice MR images. The 3D models can potentially estimate the motion of the tumour as well as the organs-at-risk during treatment on an MR-Linac from 2D cine surrogate images, and facilitate accurate dose calculations.