Quality assurance incorporating artificial intelligence-generated reference contours in a phase II radiotherapy trial

Background and purpose: A dummy run quality assurance was conducted for a prospective phase II trial of tailored radiotherapy (RT) according to the response after neoadjuvant chemotherapy (NAC) followed by surgery in locally advanced breast cancer (RTaNAC). Since inter-institutional variations in dose distributions can impact clinical outcomes, this dummy run aimed to develop an RT plan protocol to minimize these variations. Material and methods: This study involved computed tomography images from three clinical scenarios: RT with no lymph node (LN) boost (Scenario 1), LN boost up to 60 Gy3.5, equivalent dose in 2 Gy fractions (EQD2) with α/β = 3.5 Gy (Scenario 2), and LN boost up to 66 Gy3.5 (Scenario 3). Seven institutions developed RT plans under a two-step process: first according to each institution’s policies (Step 1) and then using additional reference information, including artificial intelligence (AI) auto-contoured structures and LN boost target volume information (Step 2). Dose-volume histograms for breast and regional nodal areas were analyzed between the two steps. Results: Inter-institutional variation observed in Step 1 improved in Step 2 for breast and regional nodal areas. Specifically for Scenario 1, dose coverage for the nodal clinical target volumes of the axillary level I LN improved from 57.5% to 97.5% (p-value = 0.075), and that of the supraclavicular LN improved from 75.6% to 88.8% (p-value = 0.046). Conclusion: Variations in dose/volume metrics among institutions were mitigated by AI auto-contoured structures and LN boost target volume information. Through this dummy run, the participating institutions reached a consensus on an RT plan protocol to support multi-institutional expansion of RTaNAC.