S143
Brachytherapy - Physics
ESTRO 2026
architecture. The models were trained using five-fold cross-validation and the nnU-Net ResEncL residual encoder preset. Independent models were trained for each of the OARs routinely contoured for cervix brachytherapy plans locally, following EMBRACE II: bladder, bowel, rectum, and sigmoid [3].
Conclusion: In both cases, the combination of the two systems improved error detection robustness. The two treatment verification systems uniquely complement each other, allowing for improved error detection sensitivity and specificity. The absolute positional measurements of the FPD tracking and imaging system allow for better localisation of the ISD, and the ISD in the treatment volume reduces the reliance of the FPD source tracking on pre-treatment coordinate space registration. Keywords: Verification, Source Tracking, In vivo dosimetry References: [1] Smith et al. An integrated system for clinical treatment verification of HDR prostate brachytherapy combining source tracking with pretreatment imaging. Brachytherapy (2018) 17(1):111-121[2] Johansen et al. Time-resolved in vivo dosimetry for source tracking in brachytherapy. Brachytherapy (2018) 17(1):122-132 Proffered Paper 4446 Developing an in-house autocontouring solution for MR-only gynae Brachytherapy James Cummings 1 , Rachel Joshi 1 , Samuel Ingram 1,2 1 Christie Medical Physics & Engineering, The Christie NHS Foundation Trust, Manchester, United Kingdom. 2 Division of Cancer Sciences, University of Manchester, Manchester, United Kingdom Purpose/Objective: At our centre gynae brachytherapy patients are imaged para-axially, aligned with the in-situ applicator following GEC-ESTRO recommendations [1], resulting in a range of slice tilt angles. Commercial AI contouring solutions that are not trained on tilted MR scans will often not run, or produce clinically unsuitable contours. A feasibility study was undertaken to develop an AI autocontouring model capable of working with tilted MR slices. Material/Methods: The clinical MR (T2 TSE) treatment planning scans and structure sets for 182 patients (534 fractions treated 2023-2025) with a range of tilt angles (Figure 1) were used to train an autocontouring model using the nnU- Net v2 framework [2] with 3D full resolution
Once trained, the model was qualitatively reviewed on 5 test patients (15 fractions) within the treatment planning system alongside clinical contours for comparison. Results: Models were successfully trained on a wide range of tilted MR scans. Training cross-validation mean DICE scores of 0.91, 0,79, 0.75 and 0.53 were achieved for bladder, rectum, sigmoid and bowel respectively (with bowel results largely due to superior/inferior contouring variation).The autocontouring models were successfully able to identify the OARs and contour them in all 15 test cases. A qualitative review by an independent Medical Physics Expert found that the contours capture the OARs (Figure 2), with some differences seen in comparison to clinical contours in areas of greater ambiguity/ finer features.
Made with FlippingBook - Share PDF online