S1575
Physics - Autosegmentation
ESTRO 2026
Digital Poster Highlight 2855
Potential of semi-automated uncertainty-guided segmentation to reduce editing effort in online adaptive treatment of the prostate Dagmar M. C. Eek, Joëlle E. van Aalst, Anneke M. Middelveld, Arjen van der Schaaf, Shafak Al-Uwini, Charlotte L. Brouwer Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands Purpose/Objective: Magnetic resonance-guided radiotherapy (MRgRT) enables adaptation to inter-fraction deformation in prostate cancer patients [1]. However, it requires rapid contouring, which can be accelerated using auto- segmentation. Despite extensive training, auto- segmentation performance varies across patients. Uncertainty quantification (UQ) can highlight regions likely needing correction. However, voxel-wise maps are complex and challenging to act on [2]. This study investigates a clinically actionable workflow that thresholds uncertainty maps into regions requiring correction and those that can be left unchanged, and assesses the resulting impact on segmentation accuracy and dose evaluation. Material/Methods: Pre-treatment T2-weighted MRI scans and manually delineated prostate CTVs and organs of interest (OOIs) were retrospectively collected from 10 patients with localised prostate carcinoma. Auto-segmentations were generated using an in-house DeepLabV3+ model with an Xception backbone, and UQ was performed using Monte Carlo dropout. To evaluate threshold- based corrections, voxels exceeding uncertainty thresholds, varying from 0 to 1 in 0.1 increments, were replaced with ground truth labels, simulating uncertainty-guided editing. For an example of uncertainty-guided editing, see Figure 1. Geometric accuracy was quantified using Dice similarity coefficient (DSC) and added path length (APL). Dosimetric impact was assessed using manually created mimicked MR-linac treatment plans with a 9- beam step-and-shoot IMRT workflow and extreme hypofractionation (5 x 7.25 Gy) in RayStation Research 2023B using the PTV V99% and CTV V37.5Gy. A new treatment plan was optimised for every uncertainty threshold per patient, leading to a total of 100 plans.
Results: Segmentation performance improved as the uncertainty threshold decreased (i.e. more was corrected), with DSC increasing and remaining APL decreasing (Figure 2). Correcting only the most uncertain voxels (T > 0.8) yielded the largest geometric gains (DSC: 0.80 --> 0.87) and reduced the remaining APL to 50%, suggesting a potential 50% reduction in manual contouring effort. Dosimetric evaluation showed a corresponding improvement, with voxels above T > 0.8 correcting approximately 50% of the target dose error (PTV V99% and CTV V37.5Gy) while dose differences stabilized at lower uncertainty thresholds. A similar trend was observed for OOI Dmean (not shown), though maximum dose metrics were less consistently affected.
Made with FlippingBook - Share PDF online