Automated image segmentation and accelerated dose calculation for proton therapy using deep convolutional neural networks

Radiation therapy (RT) is a cancer treatment modality that uses ionizing radiation to target cancer cells. Conventionally, treatment planning includes delineation of anatomical structures on a Computed Tomography (CT) image, followed by a dose calculation. Manual delineation is still a common clinical practice, which is both tedious and time-consuming. Also, it suffers from observer variability, manual delineation being possibly inaccurate. Further in the workflow, quality of a dose calculation engine depends on accurate modeling of the underlying physics that is typically computationally expensive. This thesis mainly focuses on the aforementioned challenges in RT treatment planning. It investigates deep convolutional neural networks (CNNs) to address such challenges for improved treatment. This work can be split into three parts. Part 1 deals with data generation using Generative Adversarial Networks. Synthetic CT images were generated to augment the limited training data for robust network learning. Part 2 explores multi-organ segmentation using dilated UNet (an encoder-decoder styled CNN architecture) to automate the delineation task in chest CT images. In addition, delineation variability (inter- or intra-observer) was modeled using a proposed framework (contour augmentation) in pelvic CT images, which uses a manual delineation to generate multiple contours by applying a local deformation to the given delineation. Contour augmentation can enhance the discriminative capabilities of the network, thus offering increased segmentation performance compared to standard network training. Finally, Part 3 addresses accelerating Monte Carlo (MC) dose calculation for proton therapy. The long computation times of MC simulations were reduced by assuming that faster simulations with fewer particles can still compare to slower ones in accuracy and that denoising with dilated UNet can compensate for the lower precision. Tests were conducted in multiple tumor sites. The reduced inference time of the proposed denoising framework can enable its integration into the clinical workflows when speed matters, like building large dictionaries of elementary beams (beamlets) prior to treatment plan optimization or robustness evaluation with many scenarios.