S2105
Physics - Image acquisition and processing
ESTRO 2026
Radiation Oncology, University of Texas MD Anderson Cancer Center, Houston, Texas, USA. 3 Division of Cancer Sciences, School of Medical Sciences, Faculty of Biology, Medicine and Health, The University of Manchester, Christie NHS Foundation Trust Hospital, Manchester, United Kingdom. 4 Division of Cancer Sciences, School of Medical Sciences, School of Medical Sciences, Faculty of Biology, Medicine and Health, The University of Manchester, Christie NHS Foundation Trust Hospital, Manchester, United Kingdom Purpose/Objective: Magnetic resonance-guided adaptive radiation therapy (MRgART) enables daily imaging and treatment adaptation, especially for localised prostate cancer. However, significant anatomical differences between patients complicate voxel-based analysis and cohort- level integration. This study develops and validates robust CT and MRI-based deformable image registration (DIR) for spatially normalising anatomical and dosimetric data across patient populations, using a data-driven reference-patient selection algorithm. Material/Methods: Forty-six prostate cancer patients received 5-fraction MRgART (7.25 Gy per fraction) on a 1.5 T MR-Linac (2021–2024); four patients with prior transurethral resection of the prostate (TURP) were excluded, leaving 42 patients for analysis. We investigate the factors influencing registration performance in the male pelvis region and propose a three-factor anatomical characterisation framework that incorporates pelvic volume proportion (Factor 1), organ shape morphology and texture heterogeneity (Factor 2), and normalised spatial positioning (Factor 3), as shown in Figure 1. The patient at the centroid of the factor distribution was identified and selected as the best reference (BestRef), with the nearest centroid patient designated as near BestRef. We further investigated the registration performance on the farthest patient. The rest of the cohort was spatially normalised to each investigated reference using B- spline DIR, with similar parameters, including normalised mutual information and bending energy regularisation. The DIR performance was evaluated by normalised cross-correlation (NCC), Dice similarity coefficient (DSC), Mean Distant to Agreement (DTA), and Hausdorff distance (HD) across both CT and MRI modalities. DIR performance between each reference was compared by paired test, considering statistically significant at p<0.05.
Results: Identical deformable parameters achieved consistent accuracy across modalities, demonstrating cross- modality applicability. BestRef registration yielded mean NCC values of 0.92 (CT) and 0.93 (MRI) with higher DSC, MDA, and HD (Figure 2) with no significant with near BestRef. The lower value of all metrics with larger deviation was found when using extremely anatomical patients as a reference (p<0.05). The near- identical NCC across modalities validates parameter consistency and physical plausibility of registrations.
Made with FlippingBook - Share PDF online