Access Type

Open Access Dissertation

Date of Award

January 2020

Degree Type


Degree Name



Medical Physics

First Advisor

Carri K. Glide-Hurst


Growing evidence suggests that radiation therapy (RT) doses to the heart and cardiac substructures (CS) are strongly linked to cardiac toxicities, though only the heart is considered clinically. This work aimed to utilize the superior soft-tissue contrast of magnetic resonance (MR) to segment CS, quantify uncertainties in their position, assess their effect on treatment planning and an MR-guided environment.

Automatic substructure segmentation of 12 CS was completed using a novel hybrid MR/computed tomography (CT) atlas method and was improved upon using a 3-dimensional neural network (U-Net) from deep learning. Intra-fraction motion due to respiration was then quantified. The inter-fraction setup uncertainties utilizing a novel MR-linear accelerator were also quantified. Treatment planning comparisons were performed with and without substructure inclusions and methods to reduce radiation dose to sensitive CS were evaluated. Lastly, these described technologies (deep learning U-Net) were translated to an MR-linear accelerator and a segmentation pipeline was created.

Automatic segmentations from the hybrid MR/CT atlas was able to generate accurate segmentations for the chambers and great vessels (Dice similarity coefficient (DSC) > 0.75) but coronary artery segmentations were unsuccessful (DSC<0.3). After implementing deep learning, DSC for the chambers and great vessels was ≥0.85 along with an improvement in the coronary arteries (DSC>0.5). Similar accuracy was achieved when implementing deep learning for MR-guided RT. On average, automatic segmentations required ~10 minutes to generate per patient and deep learning only required 14 seconds. The inclusion of CS in the treatment planning process did not yield statistically significant changes in plan complexity, PTV, or OAR dose.

Automatic segmentation results from deep learning pose major efficiency and accuracy gains for CS segmentation offering high potential for rapid implementation into radiation therapy planning for improved cardiac sparing. Introducing CS into RT planning for MR-guided RT presented an opportunity for more effective sparing with limited increase in plan complexity.