S2334
Physics - Quality assurance and auditing
ESTRO 2026
Digital Poster 879 Statistical process control for performance monitoring and continuous quality assurance of deep learning segmentations in radiotherapy Niels van Acht 1,2 , Dave van Gruijthuijsen 1 , Johanna Bluemink 1 , Els Hagelaar 1 , Coen Hurkmans 1,2 1 Radiation oncology, Catharina hospital, Eindhoven, Netherlands. 2 Electrical engineering, Eindhoven university of technology, Eindhoven, Netherlands Purpose/Objective: After clinical implementation of deep learning segmentation (DLS) models it is highly recommended and will soon be mandated by the EU AI act to perform continuous quality assurance (CQA) and introduce alarms. Therefore, the goal was to implement a CQA framework for DLS in radiotherapy that utilises statistical process control to monitor performance. Material/Methods: The direct output of the DLS models and the clinically approved structures (CS) were automatically exported, after which the volumetric Dice similarity coefficient (VDSC), surface Dice similarity coefficient (SDSC) with 3 mm tolerance, 95th percentile Hausdorff distance (HD95) and added path length (APL) were calculated. For each region of interest (ROI)-metric combination, a target, lower and upper control limit were determined, based on statistical process control (SPC). Adapted versions of the first three Nelson rules were used. Rule 1 stated that a ROI was an outlier when 3 out of the 4 metrics were outside of the control limits. Rule 2 stated that a trend shift occurred if 9 consecutive ROIs were above or below the target per metric, with 0.5 % tolerance. Lastly, rule 3 stated that a trend drift occurred when 6 consecutive ROIs were either increasing or decreasing per metric. The CQA framework is represented in Figure 1. Results: In the first six months, 545 DLS and corresponding CS RT structure files were logged containing 3093 ROIs. From these ROIs, 3.0 % was automatically reported as outlier and manually investigated. Based on the control limits 76 ROIs were more and 5 less adjusted than expected. Additionally 12 were identified as labelling errors and five patients, whose anatomy was deemed interesting, were saved for potential model re-training. Twelve trend shifts and one trend drift were detected that caused minor temporary deviations. However, four trend shifts identified a major permanent performance drop for the humerus (Figure 2) after upgrading to a new version of the DLS model in week 15. This was an unexpected change as the new DLS model passed commissioning. Investigation pointed out that this drop was caused by a combination of the new DLS model with a new CT scanner. Patients imaged on the new CT scanner were
performing dose accumulation calculations. Despite these differences, most reported doses remained within clinically acceptable limits.
Figure 1: Cumulative dose (EQD2) for centres using RIR and DIR, including those using TRFs for both patient cases. Conclusion: This study demonstrates variability in re-irradiation dose accumulation practices across Irish radiotherapy centres, influenced by registration and dose summation techniques. The results underscore the need for national benchmarking and standardisation of reRT dose accumulation protocols. Establishing such guidelines would enhance consistency, support training and implementation of new software, and ultimately improve clinical decision-making and patient safety in re-irradiation planning. References: Andratschke N, et al. (2022). ESTRO–EORTC consensus on re-irradiation: definition, reporting, and clinical decision making. Lancet Oncol, 23(10). https://doi.org/10.1016/S1470-2045(22)00447- 8Hardcastle, N., Vasquez Osorio, E., Jackson, A., Mayo, C., Aarberg, A. E., et al. (2024). Multi-centre evaluation of variation in cumulative dose assessment in reirradiation scenarios. Radiotherapy and Oncology, 194, 110184. https://doi.org/10.1016/j.radonc.2024.110184Paradis, K. C., Mayo, C., Owen, D., Spratt, D. E., Hearn, J., et al. (2019). The special medical physics consult process for reirradiation patients. Advances in Radiation Oncology, 4(4), 715–722.
https://doi.org/10.1016/j.adro.2019.05.007 Keywords: Re-irradiation, Image registration
Made with FlippingBook - Share PDF online