S2342
Physics - Quality assurance and auditing
ESTRO 2026
corrected (“reviewed”) contours. Performance was evaluated using Dice Similarity Coefficient (DSC) and Distance to Agreement (DTA) metrics. The DTA quantifies the average (or 95th percentile) spatial distance between voxels of the evaluated contour and the closest corresponding voxels in the VAE reconstructed contour. Higher DTA values indicate greater deviations from the training population. Results: Compared with the original contours, the reviewed contours showed statistically significant improved metrics (Table 1). The DTA maps produced by the tool clearly highlight areas of contour disagreement, with regions of high DTA closely matching those corrected during manual review (Figure 1). Figure 1 shows a case where three slices at the superior end of the spinal cord had been incorrectly identified as brainstem, and six slices were missing at the inferior end of the spinal cord. Potential errors can be automatically identified by applying a DTA threshold and filtering connected regions above a specified size. By optimising these parameters, a small number of regions (3.4 on average for the brainstem and 3.8 for the spinal cord) were detected, successfully capturing all reviewed regions in 10/10 and 7/8 cases, respectively. The remaining regions likely correspond to minor deviations which are clinically acceptable.
used. The reduction of MPC tolerance levels can improve the sensitivity of quality assurance by allowing earlier detection of errors. This helps identify issues such as gradual beam output drifts, geometric misalignments, or calibration errors before they affect treatment accuracy. References: 1. Barnes, M.P. and Greer, P.B. (2017), Evaluation of the truebeam machine performance check (MPC): mechanical and collimation checks. J Appl Clin Med Phys, 18: 56-66 Keywords: Radmachine, Quality Assurance, MPC, Automation An automatic tool to guide the review of OAR contours for clinical trials quality assurance Clea Dronne 1,2 , Catharine H Clark 3,2 , Xavier Loizeau 4 , Elizabeth Miles 5 , Peter Hoskin 5,6 , Jamie R McClelland 1 1 UCL Hawkes Institute, Department of Medical Physics and Biomedical Engineering, University College London, London, United Kingdom. 2 Radiotherapy and Radiation Dosimetry Group, National Physical Laboratory, Teddington, United Kingdom. 3 Department of Medical Physics and Biomedical Engineering, University College London, London, United Kingdom. 4 Data Science Department, National Physical Laboratory, Teddington, United Kingdom. 5 National Radiotherapy Trials Quality Assurance Group (RTTQA), Mount Vernon Hospital, Northwood, United Kingdom. 6 Division of Cancer Sciences, University of Manchester, Manchester, United Kingdom Purpose/Objective: Quality assurance (QA) of OAR contours in radiotherapy clinical trials is a critical but time- Proffered Paper 1291 consuming task for clinical trial QA groups. With the increasing adoption of AI-generated contours in routine practice, new types of errors may arise, distinct from those made by clinicians, further increasing the task difficulty1. We present an automatic tool to support OAR contour review, based on a variational autoencoder2 (VAE). The model jointly reconstructs CT images and OAR contours to learn expected anatomical variability. Material/Methods: The VAE was pre-trained on the open-source RADCURE head-and-neck dataset3 (2850 patients) and fine- tuned on the ARTDECO clinical trial dataset4 (262 patients). By modelling acceptable anatomical variability, the model can flag contours that deviate from learned patterns, thereby identifying potential errors. Our test set included cases from the ARTDECO trial comprising 10 brainstem and eight spinal cord delineations, each with original and manually
Made with FlippingBook - Share PDF online