Author ORCID Identifier
0000-0002-7683-5093
Date of Graduation
5-2023
Document Type
Dissertation (PhD)
Program Affiliation
Medical Physics
Degree Name
Doctor of Philosophy (PhD)
Advisor/Committee Chair
Laurence Court
Committee Member
Sam Beddar
Committee Member
Tina Marie Briere
Committee Member
Carlos Eduardo Cardenas
Committee Member
Prajnan Das
Committee Member
David Fuentes
Abstract
Background
Rectal cancer is a common type of cancer. There is an acute health disparity across the globe where a significant population of the world lack adequate access to radiotherapy treatments which is a part of the standard of care for rectal cancers. Safe radiotherapy treatments require specialized planning expertise and are time-consuming and labor-intensive to produce.
Purpose:
To alleviate the health disparity and promote the safe and quality use of radiotherapy in treating rectal cancers, the entire treatment planning process needs to be automated. The purpose of this project is to develop automated solutions for the treatment planning process of rectal cancers that would produce clinically acceptable and high-quality plans. To achieve this goal, we first automated two common existing treatment techniques, 3DCRT and VMAT, for rectal cancers, and then explored an alternative method for creating a treatment plan using deep learning.
Methods:
To automate the 3DCRT treatment technique, we used deep learning to predict the shapes of field apertures for primary and boost fields based on CT and location and the shapes of GTV and involved lymph nodes. The results of the predicted apertures were evaluated by a GI radiation oncologist. We then designed an algorithm to automate the forward-planning process with the capacity of adding fields to homogenize the dose at the target volumes using the field-in-field technique. The algorithm was validated on the clinical apertures and the plans produced were scored by a radiation oncologist. The field aperture prediction and the algorithm were combined into an end-to-end process and were tested on a separate set of patients. The resulting final plans were scored by a GI radiation oncologist for their clinical acceptability.
To automate of VMAT treatment technique, we used deep learning models to segment CTV and OARs and automated the inverse planning process, based on a RapidPlan model. The end-to-end process requires only the GTV contour and a CT scan as inputs. Specifically, the segmentation models could auto-segment CTV, bowel bag, large bowel, small bowel, total bowel, femurs, bladder, bone marrow, and female and male genitalia. All the OARs were contoured under the guidance of and reviewed by a GI radiation oncologist. For auto-planning, the RapidPlan model was designed for VMAT delivery with 3 arcs and validated separately by two GI radiation oncologists. Finally, the end-to-end pipeline was evaluated on a separate set of testing patients, and the resulting plans were scored by two GI radiation oncologists.
Existing inverse planning methods rely on 1D information from DVH values,2D information from DVH lines,or 3D dose distributions using machine learning for plan optimizations. The project explored the possibility of using deep learning to create 3D dose distributions directly for VMAT treatment plans. The training data consisted of patients treated by the VMAT treatment technique in the short-course fractionation scheme that uses 5 Gy per fraction for 5 fractions. Two deep learning architectures were investigated for their ability to emulate clinical dose distributions: 3D DDUNet and 2D cGAN. The top-performing model for each architecture was identified based on the difference in DVH values, DVH lines, and dose distribution between the predicted dose and the corresponding clinical plans.
Results:
For 3DCRT automation, the predicted apertures were 100%, 95%, and 87.5% clinically acceptable for the posterior-anterior, laterals, and boost apertures, respectively. The forward planning algorithm created wedged plans that were 85% clinically acceptable with clinical apertures. The end-to-end workflow generated 97% clinically acceptable plans for the separate test patients.
For the VMAT automation, CTV contours were 89% clinically acceptable without necessary modifications and all the OAR contours were clinically acceptable without edits except for large and small bowels. The RaidPlan model was evaluated to produce 100% and 91% of clinically acceptable plans per two GI radiation oncologists. For the testing of end-to-end workflow, 88% and 62% of the final plans were accepted by two GI radiation oncologists.
For the evaluation of deep learning architectures, the top-performing model of the DDUNet architecture used the medium patch size and inputs of CT, PTV times prescription dose mask, CTV, PTV 10 mm expansion, and the external body structure. The model with inputs CT, PTV, and CTV masks performed the best for the cGAN architecture. Both the DDUNet and cGAN architectures could predict 3D dose distributions that had DVH values that were statistically the same as the clinical plans.
Conclusions:
We have successfully automated the clinical workflow for generating either 3DCRT or VMAT radiotherapy plans for rectal cancer for our institution. This project showed that the existing treatment planning techniques for rectal cancer can be automated to generate clinically acceptable and safe plans with minimal inputs and no human intervention for most patients. The project also showed that deep learning architectures can be used for predicting dose distributions.
Keywords
Rectal Cancer; Radiotherapy; Treatment Planning; Medical Physics