ESTRO 2025 - Abstract Book

S2967

Physics - Image acquisition and processing

ESTRO 2025

[4] Wasserthal, J., et al (2023). TotalSegmentator: robust segmentation of 104 anatomic structures in CT images. Radiology: Artificial Intelligence , 5 (5). 1185 Digital Poster Robust and automated synthetic CT generation via deformable image registration for patients with head and neck cancer Adam Mylonas 1 , Brendan Whelan 1 , Robert Finnegan 1 , Doan Trang Nguyen 1 , Nicholas Hardcastle 1 , Shivani Kumar 1,2 , Paul Keall 1 , Tania Twentyman 3 , Michael G Jameson 3,2 , Lars Mejnertsen 1 1 SeeTreat Medical, -, Sydney, Australia. 2 University of New South Wales, -, Sydney, Australia. 3 GenesisCare, -, Sydney, Australia Purpose/Objective: Adaptive radiation therapy requires high-quality imaging to assess the need for adaptation, but daily CBCT has limitations, including poor image quality, artefacts, inaccurate Hounsfield units, low soft tissue contrast, and a limited field of view. These limitations can compromise the accuracy of dose calculation and structure delineation. To address these issues, multimodal deformable image registration (DIR) can be used to generate a synthetic CT (sCT) through registering the planning CT (pCT) onto the daily CBCT. The goal of this work was to develop a robust and automated sCT generation method via DIR for patients with head and neck cancer (HNC). Material/Methods: The registration algorithm was composed of rigid and deformable registration with intermediate image processing. CBCT truncation artefacts were handled using an optimisation procedure to mask truncated areas. Rigid registration was performed to get an initial alignment, followed by slab-wise histogram matching of the CBCT to the pCT. The DIR applied a multi-resolution demons algorithm to compute the deformation vector field (DVF). 1 The sCT was created by deforming the pCT via the DVF. A retrospective cohort of six patients with HNC were selected. For each patient the sCT method was evaluated on three daily CBCT images acquired at the start, middle, and end of the treatment course. The registration performance was quantified using the target registration error (TRE) of 12 to 20 manually annotated anatomical landmarks. The landmarks were labelled by a medical physicist and reviewed by a radiation therapist. The computation time of the sCT was measured. Results: The TRE mean (± standard deviation) was 2.6 ± 1.7 mm with a median of 2.1 mm (Fig. 1). The 5 th and 95 th percentiles were 0.7 mm and 6.6 mm. The maximum voxel dimension was 2 mm for CT and 3 mm for CBCT. Figure 2 demonstrates a substantial reduction in the registration error of the sCT compared to the pCT. On average, the sCT generation was completed in 85 seconds on a CPU.

Fig. 1 Box plots of the TRE for three CBCT images of each patient, with whiskers representing the minimum and maximum values.