S1343
Interdisciplinary - Education in radiation oncology
ESTRO 206
Keywords: Autosegmentation, OARs, Systematic errors
results the local group established categories to quantify the model outputs. Step 2 – To test if regular users of the model would agree, they were asked to review individual OARs and assign a category. Results: Expert evaluation categorised each of the organs into four distinct categories: (1) AI-correct cases requiring practice change, (2) systematically biased AI, (3) randomly incorrect AI and (4) minor changes only (Figure 1). 10 regular users of the model (including both radiation therapists and radiation oncologists) rated 10 structures each, sampled from 127 structures across 13 OARs, in random order, viewing post-AI raw contours only, and selected one of the categories. Despite regular use of the model, consensus was not achieved on the correct categorisation of all structures (Figure 2). Overall ability to correctly identify structure category was only 57%. Sensitivity for identifying systematic error was only 50%, with specificity of 79%. While 65% of random gross outliers were correctly identified. OARS which had previously been flagged in our network with targeted education interventions due to systematic errors scored better (e.g. spinal cord 83% sensitivity) compared to OARs with no targeted education (e.g. parotid 17% sensitivity).
Mini-Oral 4610 A Structured approach to training medical physicists during rapid industry transformation and limited educational resources Ruslan Zelinskyi 1 , Serhii Brovchuk 2,3 , Natalka Suchowerska 4 , Viktor Yakovenko 5 , Nataliya Kovalchuk 6 1 Medical Physics, National Cancer Institute, Kyiv, Ukraine. 2 Radiation therapy, National Institute "O.O. Shalimov National Scientific Center of Surgery and Transplantology", Kyiv, Ukraine. 3 Radiation therapy, LISOD - Israeli Oncology Hospital, Pliuty, Ukraine. 4 Radiation oncology, University of Sydney, Sydney, Australia. 5 Radiation oncology, UT Southwestern Medical Center, Dallas, USA. 6 Radiation oncology, Stanford University, Stanford, USA Purpose/Objective: At the beginning of 2022, Ukraine had 33 Linacs and 44 Co-60 units in clinical use. Following the onset of the full-scale war, it became evident that Co-60 sources, originally manufactured in the USSR, will be impossible to replace. To address this gap, the Ministry of Health launched the procurement of 21 Linacs, supplementing 5 under installation.This abrupt transition from outdated to modern technologies created an urgent need for qualified personnel, particularly medical physicists. However, medical physics in Ukraine faced challenges, including limited access to high-quality university programs, insufficient postgraduate education, and weak regulatory oversight. These issues have become especially critical amid the rapid transformation and the influx of
modern equipment. Material/Methods: Given the impossibility of establishing a
comprehensive medical physics education system in a short timeframe, an adaptive training algorithm was developed to ensure safe and effective use of new equipment under conditions of limited educational resources.At the initiative of Help Ukraine Group, a three-phase training program for Ukrainian medical physicists was implemented in 2024–2025:Phase 1: An online course mostly covering theoretical fundamentals (50 lectures). Phase 2: Four 3-day in- person sessions in groups of 15-20 combining lectures and hands-on exercises covering commissioning, dosimetry, quality control and planning.Phase 3: Fifteen 3-day hands-on training sessions in groups of four covering commissioning, dosimetry, quality assurance, and treatment planning for different localizations Phase 1 conducted by AAPM lecturers in English with Ukrainian subtitles, Phase 2 and 3
Conclusion: Successful AI implementation requires understanding specific model limitations. Our four-category framework provides a practical approach for risk- stratified monitoring protocols. These results show that although users remained somewhat vigilant to random errors, there was complacency and bias towards systematic errors. This could be explained by passage of time or by newer users not having a reference of pre automation era, and highlights the importance of targeted education interventions, structured monitoring, and periodic re-evaluation.
Made with FlippingBook - Share PDF online