S2310
Physics - Machine learning and AI algorithms
ESTRO 2026
technique, and patient setup metrics across different tumor locations. For example, we observed a Csim value of 0.88 and 0.30 between Maastro and BU in head-and-neck (H&N) and gastrointestinal (GI) prescription and patient setup parameters respectively (Figure 2), which indicates a similar practice in H&N and more diverge practice in GI between the two clinics. Conclusion: Vector space modeling provides a novel, quantitative framework to measure variations in radiation oncology practices. This approach enables objective and quantitative comparison of clinical data while preserving data privacy, offering potential applications
Imaging Archive. Keywords: Image Registration, Deep Learning, DEEDS
Proffered Paper 3730 Vector Space Model and Statistics for Quantifying Multi-Institutional Differences in Radiation Oncology Clinical Practices Petros Kalendralis 1 , Shaotai Hu 2 , Kaivalya Bhatt 2 , Andre Dekker 3 , Alan Kalet 4 , Samuel Luk 2,5 1 Department of Radiation Oncology (Maastro), GROW Research Institute for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, The Netherlands, Netherlands. 2 Department of Radiation Oncology, Boston University, Boston, MA, United States, USA. 3 Department of Radiation Oncology (Maastro), GROW Research Institute for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, Netherlands. 4 Department of Radiation Oncology, University of Washington Medical Center, Seattle, WA, USA. 5 Department of Radiation Oncology, Boston Medical Center, Boston, MA, United States, USA Purpose/Objective: Inconsistencies in radiation oncology clinical practices across institutions may impact treatment outcomes and AI model performance. This multi-institutional study introduces a novel vector space model (VSM) approach to quantify and compare clinical practice variations between radiation oncology centers using natural language processing techniques. Material/Methods: Anonymized radiation oncology datasets including tumor locations, prescription, treatment planning and setup parameters from four institutions (Boston University (BU), University of Washington (UW), University of Vermont (UVM), and Maastro Clinic) were analyzed. Each clinical data was preprocessed locally into tokenized corpora and converted into multidimensional vectors using Word2Vec model with continuous bag-of-words methodology. The generated word vectors are then shared and weighted cosine similarity (Csim) between the vectors were calculated to quantify differences in prescription, plan complexity, and treatment approaches of various anatomic tumor locations across institutions. Results: Differences between clinical practices on prescription patterns and treatment planning parameters are observed in the multi-institutional data (Figure 1). The VSM approach successfully quantified institutional differences using Csim. The Csim values are paired with the differential term frequency plots to demonstrate variations between two institutions in prescription dose patterns, treatment planning
for quality assurance, treatment protocol standardization, and improving AI model generalizability across institutions.
Figure 1: Distribution of prescribed target (PTV) dose across four radiation oncology institutions (Maastro Clinic, Boston University, University of Vermont, and University of Washington) stratified by anatomic tumor location.
Figure 2:Weighted cosine similarity (Csim) comparison of clinical practices between Maastro Clinic and Boston University for gastrointestinal (left panel, Csim: 0.2952) and head-and-neck (right panel, Csim: 0.8797) Keywords: AI, Vectors. Modelling
Made with FlippingBook - Share PDF online