Deep-Learning Dose-Driven Approach to Optimize Patient-Specific Quality Assurance Program in Radiotherapy

Abstract

Featured Application: Deep learning applied to dose distributions with engineered features enables accurate identification of radiotherapy plans requiring patient-specific quality assurance (PSQA). This approach allows more efficient allocation of medical physics staff and linear accelerator time, supporting safer, faster, and more streamlined PSQA workflows. Background: Radiotherapy (RT) is a cornerstone of cancer treatment, with patient-specific quality assurance (PSQA) being essential for patient safety. However, PSQA can be time-consuming, particularly in high-throughput centers. Machine learning and deep learning (DL) offer tools to efficiently analyze large datasets and support data-driven workflows. Methods: DL was applied to predict PSQA outcomes for clinical RT plans delivered with 6 MV and 6 FFF MV. A retrospective dataset of 839 VMAT plans was used to extract 4470 dosiomic features from five automatic isodose contours. The dataset was split into training and test subsets. Two hybrid models were developed to identify plans with γ-passing rates below the median QA value (97%). A prospective dataset of 201 PSQA measurements was used for validation. Results: Both models demonstrated strong predictive performance. In the test set, Hybrid-1 achieved 0.74 specificity and 0.64 sensitivity, while Hybrid-2 reached 0.72 specificity and 0.52 sensitivity. Similar results were observed in the prospective validation cohort. Conclusion: Applying deep learning to dose distributions with carefully engineered features allows accurate identification of RT plans that require PSQA. This approach facilitates more efficient allocation of both personnel and linear accelerator time, thereby supporting safer and more streamlined PSQA workflows.