S2786
RTT - RTT contouring, target definition, and treatment planning
ESTRO 2026
reliability and patient safety. Material/Methods:
Since June 2023, AI-assisted OAR segmentation has been applied to all breast cancer patients receiving radiotherapy at a single institution, encompassing 1,085 cases. OAR contours were generated using a combination of CE-marked commercial and in-house AI models. In the clinical workflow, AI-generated contours were reviewed and adjusted by experienced radiographers before final approval. Both the initial AI- generated and the corrected contours were archived and evaluated using geometric metrics—Surface Dice at 2 mm (SD2mm) and mean Hausdorff distance (mHD). Clinical acceptance limits are listed in Table 1. The analysis focused on the esophagus, heart, humeral heads, lungs, and trachea, with non-relevant organs excluded based on treatment type. The first three months post-implementation served as a reference period, and both metric distributions and temporal trends were monitored. Results: For the esophagus, heart, lungs, and trachea, the proportion of outliers remained below 2% for both SD2mm and mHD, well within the defined clinical acceptance limits. In contrast, the humeral heads exhibited a high outlier rate (SD2mm: 92.7%; mHD: 69.2%), primarily due to differences between institutional contouring guidelines and those used in the commercial AI model. Figure 1 illustrates heart segmentation performance over time. Both SD2mm and mHD metrics remained consistent across all quarters (2023Q3–2025Q3), with no evidence of temporal drift or degradation. For all organs, the drift in corrections over the two-year period was minimal to non-existent, demonstrating stable AI performance and the absence of emerging AI bias in routine clinical use. This stability underscores the robustness of the implemented AI models and their suitability for sustained clinical deployment.
Conclusion: Routine geometric monitoring demonstrated that AI- assisted OAR segmentation for breast cancer patients is accurate, stable, and clinically acceptable for most organs. The observed discrepancies for the humeral heads emphasize the need to align AI model training data with institutional standards. Continuous performance tracking enables early detection of deviations, supports adaptive model improvement, and ensures sustained clinical reliability of AI tools in radiotherapy workflows. Keywords: AI contouring, delination, quality assurance RTT-Led Treatment Planning meets AI Dose Prediction: Benchmarking and Dosimetric Eligibility Screening in Prostate Radiotherapy Miriam Kerr, Pierre Thirion, Louise O'Neill, Sarah McDermott, Claire Fitzpatrick, Ciaran Malone Radiation Oncology, St. Luke's Radiation Oncology Network, St. Luke's Hospital, Dublin, Ireland Purpose/Objective: Adopting new radiotherapy techniques is rarely straightforward, even in experienced departments. In clinically eligible patients, hypofractionated radical radiotherapy delivering a simultaneous integrated boost (SIB-HypoRT) to the prostate (68Gy/25 fractions) with concurrent pelvic lymph node irradiation (45Gy/25 fractions) reduces treatment appointments without compromising efficacy. During pilot Poster Discussion 2893 implementation of SIB-HypoRT, we evaluated whether AI Dose Prediction tools could reliably benchmark our RTT-planned cases against international practice, and screen patients for dosimetric eligibility for the
Made with FlippingBook - Share PDF online