Automated Radiotherapy Treatment Planning for Breast Cancer: A Robust Tool for Global Deployment

Hana Baroudi, MS (Advisor: Laurence Court, PhD)

Breast cancer incidence continues to rise worldwide, particularly in low- and middle-income countries, where limitations in resources already constrain access to timely care. Radiotherapy is a cornerstone of breast cancer management, proven to significantly lower both recurrence and mortality. However, a growing shortage of radiation staff worldwide threatens the prompt delivery of these treatments. This thesis proposes an automated solution to address the limited accessibility of radiotherapy planning in breast cancer management by developing an end-to-end automated treatment planning model and evaluating its performance and limitations across diverse patient populations.

An automated contouring model was trained using data from 104 whole-breast patients. The model’s outputs were quantitatively assessed using the Dice similarity coefficient (DSC) and mean surface distance (MSD), and qualitatively reviewed by physicians to determine clinical acceptability. Five automated conventional planning approaches were developed, complemented by an established RapidPlan model for volumetric arc therapy. These included conventional tangents for whole-breast treatment, variations for supraclavicular node (SCLV) irradiation with or without axillary nodes, and two approaches for comprehensive regional lymph node irradiation—either photon wide tangents with an SCLV field or photon tangents with a matched electron field targeting the internal mammary nodes (IMN).

All algorithms begin by generating contours automatically for the breast clinical target volume, regional lymph nodes, and relevant organs at risk. Subsequently, gantry angles and field shapes are created and optimized to ensure adequate target coverage while constraining doses to nearby critical structures. Optimization relies on field weighting for the lymph node fields and a field-in-field technique for the tangents. These algorithms were tested for clinical validity on 15 internal whole-breast patients (150 plans) and 40 external patients from four institutions across Switzerland, Argentina, Iran, and the United States (360 plans). Plan evaluations focused on target coverage and adherence to normal tissue dose limits, and were reviewed by a radiation oncologist (5-point scale for the internal dataset) and a medical physicist (accept or edit for the external dataset). Further large-scale testing was performed on 272 internal and 285 external patients from six different countries (Argentina, Iran, Jordan, United States, South Africa and Switzerland).), enabling a comprehensive assessment of plan quality via dosimetric analysis and physicist review. The rate of automated plans requiring edits was compared across treatment sites, and fault tree analysis was employed to elucidate underlying reasons for any modifications.

Automated contouring achieved DSC values above 0.70 for target volumes and an MSD under 3 mm. Two physicians deemed 63% of automatically generated contours acceptable without further edits. For the 15 internal patients, physician review indicated that 74% automated plans were clinically acceptable as is. Notably, when using automated contours in conjunction with the RapidPlan model, 73% of plans required no further modifications. Similarly, physicist review of 40 multi-institutional cases confirmed that 79% of automated plans were ready to use, with the rest needing minor edits. A broader evaluation encompassing 540 patients (4,860 plans) revealed that 78% of automated plans required no revisions, whereas 22% needed adjustments. The rate of edits was statistically comparable across institutions, except for data from Jordan and Switzerland, which exhibited better performance, and Yale, which included intentionally challenging cases. Automated planning treating IMNs with tangent fields was more robust than using electron fields. Fault tree analysis identified decisions about clinical compromises (target coverage vs. normal tissue doses) as the primary cause of plan modifications, followed by patient anatomical and positioning variations.

In conclusion, this thesis demonstrates the feasibility of a fully automated radiotherapy planning model for whole-breast cases. The approach successfully accommodates a wide range of clinical protocols and achieves robust performance across diverse patient populations.

Advisory Committee:

• Laurence Court, PhD, Chair
• Adam Melancon, PhD
• Tucker Netherton, PhD
• Joshua Niedzielski, PhD
• Simona Shaitelman, PhD
• Sanjay Shete, PhD

Join via Zoom (Please contact Ms. Hana Baroudi for her Zoom meeting info.)