Room: Track 2
Purpose: When testing new radiation therapy techniques, it can be extremely difficult to locate multiple patient scans that match the desired characteristics for the treatment. Using machine learning, we provide a tool that is capable of editing pre-existing CT scans by inserting realistic looking tumours in hand-chose locations, bypassing the difficulty of finding actual patient data which matches the treatment requirements.
Methods: A Generative Adversarial Network (GAN) was used to edit individual volumes of interest (VOIs) in pre-existing CT scans. In our structure, the Editor network translated features of the healthy VOIs into features of cancerous volumes. The Discriminator network was used to compare actual cancerous VOIs to the edited segments, acting as a guide to help the Editor improve its ability to produce realistic looking tumours over time. A dataset of 460 diagnostic and lung cancer screening CT scans (from the LIDC-IDRI database) was used to compile both the healthy and cancerous VOIs, ranging in size from 15mm³ to 31mm³. The new CT data is being used to investigate FLASH therapy planning for a hypothetical 6 MV x-ray machine with 16 waveguides.
Results: Through a graphical user interface (GUI), the Editor network was successfully able to map healthy CT segments to realistic looking cancerous volumes. Furthermore, due to the architecture of the two networks, the Editor was found to be able to extrapolate well beyond the upper size limit of the domain of the training set, providing much more interesting and useful capabilities. The FLASH planning results are upcoming.
Conclusion: This work provides a valuable tool for the field of Radiology Oncology, which can be used to develop an abundance of patient data. While this application is focused towards lung cancer patients, the method can be extended to a variety of cancer types if given an appropriate baseline dataset.