S1597
Physics - Autosegmentation
ESTRO 2026
Results: In total 782 structures (534 OARs and 248 Targets) were contoured by the prototype. The prototype could detect Targets with a sensitivity of 82.5% and a false alarm rate of 17.24%.The median score was 3 [IQR 2-3] for both OARs and GTVs [Fig.1]. The best clinically rated structures were Lens and Cornea with a median score of 4 and the worst were Cochlea and Lacrimal Gland (median of 2). The mean MDA for OARs and GTVs was 1.33 mm ± 3.03 and 1.03 mm ± 6.74, respectively [Fig.2]. Retina had the lowest mean MDA (0.44 mm) and Lacrimal gland (2.33 mm) had the highest among all OARs. The mean DSC for all OARs was 0.58 ± 0.21 and 0.67 ± 0.18 for Target delineation. Brainstem had the highest DSC of 0.87 and the lowest belonged to Cochlea (0.32).
Poster Discussion 4270
Bayesian deformable registration for automatic probabilistic localisation of the intraprostatic lesion for precision boosting on the MR-linac Sarah A Mason 1,2 , Imogen Thrussell 2 , Sophie E Alexander 1,2 , Dow-Mu Koh 2 , Alison Tree 2 , Helen McNair 1,2 , Matthew D Blackledge 2 1 Academic Radiography, The Royal Marsden NHS Foundation Trust, Sutton, United Kingdom. 2 Radiotherapy and Imaging, The Institute of Cancer Research and the Royal Marsden NHS Foundation Trust, Sutton, United Kingdom Purpose/Objective: Delivering a focal boost to the intraprostatic lesion (IPL) may improve disease-free survival in radiotherapy (RT) for prostate cancer1. As part of online adaptive RT (oART), localisation of the gross tumour volume ([GTV], which encapsulates the IPL) is important to ensure tumour control. However, accurate localisation is challenging as the IPL is typically not visible on magnetic resonance (MR) images.We developed an automated pipeline that uses Gaussian Processes (GPs) for non-parametric registration to probabilistically localise the GTV given planning (MRref) and daily (MRdaily) MR contours. Material/Methods: Data from 29 patients treated in two MR-linac fractions2 were used to train (n=15), calibrate (n=5), and test (n=9) the GTV localisation model. The prostate and GTV were manually contoured on the MRref using diagnostic MRs as a guide. Prostate contours on MRdaily images were generated by editing propagated MRref contours per standard oART protocol. The MRref GTV (GTVref) was propagated to the MRdaily and manually shifted for visual alignment (guided by surrounding anatomy as lesions were not visible) to generate current-standard GTVdaily.A 3D deformable mask-to-mask demons registration (DIR)3 extracted 50 corresponding point pairs from the MRref and MRdaily prostate contours. These points were inputs to a GP model for posterior estimation of population GP parameters from the training set using Hamiltonian Monte Carlo sampling in Stan4 (3 chains, 1000 samples, no thinning), with convergence confirmed by a Gelman–Rubin statistic Rhat ≤ 1.01.Using estimated population parameters and the 50 DIR-derived points, GP fitting generated a probabilistic deformation vector field (DVF) applied to the GTVref to create 3000 GTVGP contour samples for the calibration and test sets. Each sample was converted to a mask, and heatmaps were obtained by averaging the superimposed masks.Calibration of GP posterior estimates against true probabilities was performed using the calibration set, and the resulting calibration function was applied to the test set heatmaps to obtain probabilistic
Conclusion: The MRI-based auto contouring prototype demonstrated promising performance, achieving high accuracy for several OARs that cannot be easily delineated on CT images. More interestingly, the prototype created clinically acceptable delineations for more than 85% of Targets. While variability persists in structures, mean MDA for all the structures including GTVs, was less than 2 mm. Overall, the tool shows potential for clinical integration following refinement
of underperforming anatomical regions. Keywords: MR-guided radiotherapy, Target delineation
Made with FlippingBook - Share PDF online