S2350
Physics - Quality assurance and auditing
ESTRO 2026
1 Radiation Oncology, Iridium, Wilrijk, Belgium. 2 Machine Learning, Raysearch Laboratories, Stockholm, Sweden. 3 Radiation Oncology, University of Antwerp, Edegem, Belgium Purpose/Objective: Accurate delineation of target volumes and organs at risk is essential for effective radiotherapy. Artificial intelligence (AI)-based segmentation can reduce inter- and intra-observer variability and time burden, but may still introduce clinically relevant inaccuracies. While previous studies have evaluated the overall effectiveness of quality assurance (QA) methods for AI- generated contours [1], limited attention has been given to the specific types of errors these methods can detect. This study investigates how effectively QA approaches identify general segmentation errors and explores their sensitivity to different error types in prostate cancer patients. Material/Methods: A retrospective dataset of 70 prostate cancer patients was analysed, including the anorectum, bladder, and both femoral heads. Initial AI-generated segmentations were reviewed by a deep learning expert and a radiation oncologist to assign a binary clinical relevance score. Segmentation errors were categorized as model-related (i.e. over-/under- segmentation) or anatomy-related (i.e. non-standard anatomy or imaging artefacts) (Figure 1). Two QA approaches were evaluated: (i) an independent model comparison [2] and (ii) entropy-based uncertainty estimation. Logistic regression classifiers per organ were trained using stratified three-fold cross- validation. Performance was averaged across folds, and sensitivity analyses were performed to determine which error types were most effectively detected.
Conclusion: These findings point to potential improvements needed in the dose reconstruction model, particularly for X6MV energies in accurately representing target dose distributions. Although not the primary focus of this study, the software’s workflow proved time- consuming and computationally demanding, which may limit its practicality for routine clinical implementation without further optimisation. References: Lehmann J et al., SEAFARER - A new concept for validating radiotherapy patient specific QA for clinical trials and clinical practice. Radiother Oncol. 2022 Jun; 171:121-128. doi: 10.1016/j.radonc.2022.04.019. Epub 2022 Apr 21. PMID: 35461949. Keywords: PSQA, Delta4, DVH Anatomy
Results: For the femoral heads, only anatomy-related errors were present due to artificial hip implants, and these were all successfully detected (accuracy = 0.99 ± 0.01; sensitivity = 1.00). For the bladder and anorectum, both model- and anatomy-related errors occurred, with moderate overall classification performance (accuracy = 0.60 ± 0.02 and 0.65 ± 0.13, respectively). Anatomy-related errors were detected more reliably (bladder = 76%, anorectum = 82%) than model-related errors (55% and 67%, respectively). Figure 2 illustrates
Poster Discussion 1585
Assessing quality assurance methods sensitivity to detect clinically relevant errors in prostate cancer auto-segmentations Michael Claessens 1 , Geert De Kerf 1 , Carole Mercier 1 , Gabriele Balletti 2 , Jonas Söderberg 2 , Dirk Verellen 1,3
Made with FlippingBook - Share PDF online