Author ORCID Identifier

0000-0003-1934-8157

Date of Graduation

5-2023

Document Type

Dissertation (PhD)

Program Affiliation

Medical Physics

Degree Name

Doctor of Philosophy (PhD)

Advisor/Committee Chair

Laurence Court, PhD

Committee Member

Carlos Cardenas, PhD

Committee Member

David Fuentes, PhD

Committee Member

Rebecca Howell, PhD

Committee Member

Arnold C. Paulino, MD

Committee Member

Julianne Pollard-Larkin, PhD

Abstract

Over the past 50 years, pediatric cancer 5-year survival rates increased from 20% to 80% in high-income countries, however, these trends have not been mirrored in low-and-middle-income countries (LMICs). This is due in part to delayed diagnosis, higher rates of advanced disease at presentation and a growing lack of access to high quality medical personnel and technology necessary to deliver complex treatments.

The long-term goal of this study was to alleviate demanding workflows and increase global access to high-quality pediatric radiation therapy by harnessing the power of artificial intelligence to automate the radiation therapy treatment planning process for pediatric patients with medulloblastoma. Radiation therapy for medulloblastoma consists of radiation to the craniospinal axis (CSI) and a boost of radiation to the post-operative tumor resection cavity. In this study we automated the treatment planning process for the primary course and boost treatment using deep learning and other automation approaches for autocontouring and autoplanning.

First, we developed and tested a 3D conformal CSI autoplanning tool for varying patient sizes based on the recommendations from the International Society of Pediatric Oncology (SIOP). The autocontoured structures’ average Dice similarity coefficient (DSC) ranged from 0.65-0.98. Of the 18 plans tested, all were scored as clinically acceptable as-is or clinically acceptable with minor, time-efficient edits preferred or required. No plans were scored as clinically unacceptable.

Next, we tested the autocontouring and autoplanning tools on 51 CSI CT scans provided from St. Jude Children’s Research Hospital to generate 15 autocontours and a composite CSI treatment plan. Three pediatric radiation oncologists from 3 institutions reviewed and scored each autocontour and plan. Of the 795 autocontours reviewed by 3 physicians, 97% of the autocontours were scored as clinically acceptable, with 92% of them requiring no edits. The clinically acceptability of the autoplans was divided by treatment field (brain and spine). For the brain field dose distributions, 100% were clinically acceptable. For the spine dose distributions, 92% of single field, 100% of extended field, and 68% of multiple field cases were scored as clinically acceptable. Most unacceptable cases were from the multiple field configuration, which is the most complex spine field configuration to plan. In all cases (major or minor edits), the physicians noted that they would rather edit the autoplan rather than create a new plan.

In the second aim of the experiment, we set to automate the treatment planning process for the resection cavity boost which included automatically contouring the post-operative gross tumor volume (GTV) resection cavity and generating a 3D conformal treatment plan. To automatically contour the GTV, we trained a CT-based, MRI-based, and multi-modality based autocontouring model. DSC (Mean±1σ) scores were 0.75±0.16 for CT-only, 0.77±0.15 for MRI-only, and 0.80±0.12 for multi-modality models. Hausdorff distances for the MRI-only and multi-modality models were significantly lower than for the CT-only model (p<0.001 and p=0.013, respectively). In clinical review, the MRI-only model achieved the best boundary detection.

Finally, using the automatically contoured GTV volumes from each respective imaging modality, we designed a script to automatically generate 3D conformal boost treatment plans. We investigated the impact of adjusting planning parameters to design an optimization algorithm that could generate a patient-specific plan with a homogenous dose to the target and minimal dose to healthy tissues. We defined clinical acceptability as achieving 95% V95 to the clinical CTV volume for each plan generated. Each patient had 8 treatment plans (4 contours and 2 wedge angles) which gave a total of 104 treatment plans to analyze. Of these, 85% were clinically acceptable. We also we able to minimize dose to healthy tissues such as the cochlea.

In conclusion, we successfully designed and tested a fully automated CSI autocontouring and treatment planning pipeline. Moreover, we successfully tested the tool on data from another institution, proving that our algorithms successfully accommodated different patient populations. Additionally, we successfully autocontoured the post-operative GTV volumes for patients using CT, MRI, or both images, and automated the boost treatment planning process to treat each respective target volume. Automating both aspects of radiation therapy for medulloblastoma has the potential to decrease demanding workflows and increase global access to high-quality pediatric radiation therapy.

Keywords

pediatrics, radiation therapy, global health, automated treatment planning, automated contouring, CT, MRI

Download

Share

COinS