Quality Assurance of Field-in-Field and Hybrid Breast Radiotherapy: Moving Towards Machine Learning-Based Prediction
Authors: Md Akhtaruzzaman1, Md Ahad Mia1, Mohammad Ullah Shemanto1, Martin Andrew Ebert2,3
Institution: 1Evercare Hospital Chattogram, Chattogram, Bangladesh 2The University of Western Australia, Perth, Australia 3Sir Charles Gairdner Hospital, Perth, Australia
Introduction
Field-in-field (FiF) and hybrid radiotherapy techniques are commonly used for breast cancer treatment to improve dose homogeneity and organ-at-risk sparing. Hybrid breast radiotherapy using conventional tangential fields (70–75%) and volumetric modulated arc therapy (VMAT; 25–30%) faces challenges regarding increased delivery complexity and quality assurance (QA). The VMAT component is routinely verified by a measurement-based patient-specific QA (PSQA) while FiF plans are typically delivered without a dedicated measurement-based verification. The aim of this study is to develop a technique-specific QA framework and to explore the use of machine learning (ML) for predicting uncertainties related to delivery in FiF and hybrid breast radiotherapy.
Methods
A combination of a retrospective study and an experimental dosimetric verification is suggested. FiF and hybrid breast radiotherapy plans will be analysed using DICOM RTPLAN, RTDOSE and RTSTRUCT datasets. Parameters of the treatment plan, such as monitor units, multileaf collimator aperture characteristics, segment complexity, beam geometry and dose distribution parameters, will be extracted. Independent dose verification and phantom measurements will be used to define reference QA results for select FiF plans. For VMAT components, measurement-based PSQA results will be collected. Supervised ML algorithms will be evaluated for their ability to predict dose agreement and gamma passing rates, such as Random Forest, Support Vector Regression and gradient boosting. Validated component-specific predictions will then be combined into a composite hybrid QA framework to classify the plans as low-, intermediate-, and high-risk. Model performance will be evaluated using independent validation, regression and classification metrics, and calibration analysis.
Results
The proposed framework is expected to identify planning and delivery complexity parameters associated with QA variability in FiF and VMAT components. ML models will be tested on their ability to predict measured QA outcomes and flag plans needing further verification. Integrated analysis may allow better characterisation of delivery-related risks in hybrid breast radiotherapy compared to VMAT-component assessment only.
Conclusion
Experimental dosimetric verification combined with ML-based prediction could provide a practical basis for thorough, risk-informed QA for FiF and hybrid breast radiotherapy. This approach may aid in identifying potentially challenging treatment strategies earlier, utilise QA resources more effectively and improve the safety of the treatment delivery process, without replacing current verification procedures.