Room: AAPM ePoster Library
Online adaptive radiotherapy with the first AI-driven linear accelerator is underpinned by the concept of a synthetic CT (sCT) derived from planning CT-CBCT deformable image registration (DIR). This study develops a framework for independent assessment of the underlying DIR.
DIR analysis was performed retrospectively for 3 bilateral head and neck (HN) patients, and for the first adaptive HN patient with online adaptation. For each fraction (N=20) the CBCT, structure set, and sCT were analysed. Qualitatively a blended comparison of the CBCT and sCT was completed and scored with a ranking system adapted from the TG132 recommendations. The consistency of the sCT HU mapping was assessed with HU histogram analysis. As the planning CT-CBCT deformation vector field is not accessible, an indirect assessment of sCT-CBCT target registration error (TRE) was carried out using the Scale Invariant Feature Transform (SIFT) algorithm to identify matching sCT-CBCT characteristic points. The TRE was characterised as the 3D displacement between corresponding point pairs.
The visual alignment of anatomic structures between sCT and CBCT appeared to be within 2 voxels (4 mm) and was ranked ‘Usuable with risk of deformity’. Peak histogram differences of up to 50 HU were observed. For the target and key OAR structures the mean pCT-sCT HU difference was up to 20 HU. SIFT identified 300 – 700 points per sCT/CBCT comparison. The mean, median, and standard deviation TRE between match points identified with SIFT was 2-4 mm, 2-3mm, and 3–6 mm respectively, with 85% of points within 5mm.
The method described enables the evaluation of sCT generated for online AI-driven adaptive radiotherapy.
Not Applicable / None Entered.