Date of Award
Santa Clara : Santa Clara University, 2019
Bioengineering; Computer Science and Engineering
In the treatment of cancer using ionizing radiation, it is important to design a treatment plan such that dose to normal, healthy organs is sufficiently low. Today, segmentation requires a trained human to carefully outline, or segment, organs on each slice of a treatment planning computed tomography (CT) scan but it is laborious, time-consuming, and contains intra- and inter-rater variability. Currently, existing clinical automation technology relies on atlas-based automation, which has limited segmentation accuracy. Thus the auto-segmentations require post process editing by an expert. In this paper, we propose a machine learning solution that shortens the segmentation time of organs-at-risk (OARs) in the thoracic cavity. The overall system will include preprocessing, model processing, and postprocessing steps to make the system easily integratable into the radiotherapy planning process. For our model, we chose to use a 3D deep convolutional neural network with a U-net based architecture because this machine learning strategy takes into account local spatial relationships, will restore the original image resolution and has been utilized in image segmentation, especially in medical image analysis. Training and testing were done with a 60 patient dataset of thoracic CT scans from the AAPM 2017 Grand Challenge. To assess and improve our system we calculated accuracy metrics (Dice similarity coefficient (DSC), mean surface distance (MSD)) and compared our model’s segmentation performance to that of an expert and the top two performing machine learning methods of the challenge. We explored using preprocessing steps such as cropping and image enhancement to improve the model segmentation accuracy. Our final model was able to segment the lungs as accurately as a dosimetrist and the heart and spinal cord within acceptable DSC ranges. All DSC values of the OARs from our method were as accurate as other machine learning methods. The DSC for the esophagus was below tolerable error for radiotherapy planning, but our mean surface distance was superior to other auto-segmentation methods. We were successful in significantly reducing manual segmentation time by developing a machine learning system. Though our approach still necessitates a single preparatory step of manually cropping anatomical regions to isolate segmentation volume, a general hospital technician could complete this task which removes the need of an expert for one time-consuming step of radiotherapy planning. Implementation of our methods to provide radiotherapy in lower-middle income countries brings us closer to accessibility of treatment for a wider population.
Goo, Brie; May, Katrina; Zhang, Haobo; and Olivas, James, "Machine Learning Solution to Organ-At-Risk Segmentation for Radiation Treatment Planning" (2019). Interdisciplinary Design Senior Theses. 55.