Several technological advances in radiotherapy have enabled the use of focused radiation to treat solid tumors within the thoracic cavity. Stereotactic Body Radiation Therapy (SBRT) offers a way to treat patients with high doses of radiation with just a few (three to five) treatments entirely non-invasively, providing excellent tumor control for both early stage non-small cell lung cancer and metastatic disease. However, respiratory motion causes the tumor and surrounding organs at risk to move. This movement is particularly concerning for thoracic SBRT, as radiation pneumonitis stems largely from an inability to visualize the tumor and lungs during treatment and thus requires larger margins. Respiratory motion has been characterized as irregular (can differ from breath to breath and minute to minute) and can induce motion of 5 cm or more, particularly at the diaphragm. Overall, there is a pressing need to measure and monitor thoracic motion while cancer patients are being treated with radiotherapy.

Our group at UCLA has previously created a 5DCT model to better represent a patient’s tumor motion. Patients are allowed to breathe freely while they undergo 25 fast, low-dose helical CT scans during simulation prior to radiation therapy. An over-determined linear model is fit to find the postion of any voxel based on the initial position and the current tidal volume and airflow. The corresponding imaged motion of the tumor and thoracic cavity can more accurately be measured with 5DCT compared to traditional 4DCT models.

We have taken this data (n = 91 patients) to begin fitting a motion model. Working with Varian, we have applied autosegmentation models to each of the 25 scans for each patient. We aim to build a motion model from CT Simulation data to be able to track 3D volumetric representations while a patient is on the treatment table undergoing radiotherapy. Working together with Varian and Siemens Healthineers, Ricky has extended a Conditional Variational Autoencoder (cVAE) model built for cardiac cine data and trained it to our 5DCT datasets. We now plan to extend this work by to be able to drive deformations with a surrogate marker.