Author: Christian Fiandra, Marco Fusella, Gianfranco Loi, Silvia Pesente, Lorenzo Placidi, Claudio Vecchi, Orlando Zaccaria, Stefania Zara 👨🔬
Affiliation: Abano Terme Hospital, University of Turin, Maggiore della Carità, Tecnologie Avanzate Srl, Fondazione Policlinico Universitario Agostino Gemelli IRCCS 🌍
Purpose: Deformable-image-registration (DIR) is essential in modern radiotherapy for adaptive RT, re-irradiation, and other clinical applications. Multimodal DIR is especially important in MRI-only workflows and MRI Linacs. This study introduces a new general-purpose multimodal (GPM) DIR algorithm, integrated into the iTAData research platform (https://eng.tecnologieavanzate.com/prodotto/ita-data/), based on the Self Similarity Context (SSC) metric. Its performance was evaluated against a commercial and clinically validated algorithm (ANACONDA, RayStation), analyzing Target Registration Error (TRE) and inverse consistency error (ICE) on MRI and CT images of the pelvis.
Methods: MR and CT images of ten prostate cancer patients, acquired in treatment position, were retrospectively selected. Ten anatomical landmarks were identified for each patient on both datasets. Rigid and deformable registrations were performed using GPM and RayStation’s hybrid algorithm with bladder, femoral heads, and rectum as control ROIs. Deformable registrations in both directions (CT→MR and MR→CT) were performed. Identical initial rigid registration was applied to both systems to prevent bias in deformation vector field (DVF) generation. ICE and TRE were computed from these DVFs and landmarks. ICE and TRE were computed applying these DVFs to landmarks and evaluated using descriptive statistics. Differences and Level of Agreement (LOA) between the two algorithms were assed respectively by Wilcoxon test and Bland-Altman analysis.
Results: The GPM algorithm demonstrated similar spatial accuracy to ANACONDA, with comparable median TRE values (GPM: 2.48 mm, ANACONDA: 2.29 mm; p=0.916) and high agreement (LOA [2.3;2.4] mm). However, GPM achieved significantly lower ICE (1.05 mm, 1.44 mm; p<0.001), demonstrating better consistency (LOA [-2.8;1.5] mm, bias=-0.7mm).
Conclusion: In this study, a newly developed DIR algorithm was compared with a widely validated one. The results obtained demonstrated the former algorithm's robustness in multimodal scenarios revealing promising performances in the analysed site; however, further testing is required in different anatomical sites to ensure the generalisability of these results.