ESTRO 2023 - Abstract Book

S556

Sunday 14 May 2023

ESTRO 2023

cannot be easily implemented without purchasing a specific scanner. Deep learning approaches have potential for accurate use, but limited validation of commercial products has been undertaken. A commercial deep-learning, cycle generative-adversarial-network (cycleGAN) sCT generation model, which can be applied through the TPS for enhanced utility, has been developed. Here we validated the dosimetric accuracy of sCTs generated using this algorithm for pelvic, brain and head and neck (H&N) cancer sites using variable MRI data from multiple scanners. Materials and Methods The commercial cycle-GAN sCT algorithm was used to train individual sCT generation models, using paired MRI-CT patient data, for the following input MRI sequences; T2-SPACE for pelvis, T1 with gadolinium (T1Gd) and FLAIR for brain and T1 for H&N. The pelvis, brain and H&N models were trained on 46 rectum data sets, 36 data sets and 57 data sets respectively. The pelvis, brain and H&N model validation cohorts were comprised of 49 (16 anus, 28 rectum and 5 prostate), 30 and 25 patients respectively. All MRI scans were deformably registered to the planning CT prior to use and sCTs were generated using the appropriate model & MRI sequence (pelvis, brain or H&N). H&N patient externals and PTVs were cropped to the above the shoulders on CT and MRI as the field of view did not extend beyond this point. VMAT treatment plans, following local clinical planning protocols, were calculated on the planning CTs and recalculated on sCTs. HU and dosimetric differences were assessed between CT and sCTs, including DVH differences to target volumes and organs at risk as a percentage of the prescription dose (for example, PTV D95% dose difference) and gamma index (2%/2mm). Results Mean absolute error (MAE) HU differences were; 48.8 HU (pelvis), 118 (FLAIR brain), 126 (T1Gd brain) and 124 HU (H&N). Mean primary PTV D95% dose differences for all sites were <0.2 % (range: -0.9 to 1.0%). Mean 2%/2mm and 1%/1mm gamma indexes for all sites were > 99.6 % (min: 95.3 %) and >97.3 % (min: 80.1 %) respectively. For all OARs for all sites, mean dose differences were <0.4 %.

Figure 1. Axial slices of a pelvis CT (left) and T2-SPACE sCT (top), brain cancer CT (left) , T1Gd sCT (middle) and FLAIR sCT (right) (middle) and H&N cancer CT (left) and T1 sCT (right) (bottom). Conclusion Generated sCTs had excellent dosimetric accuracy for pelvic T2-SPACE sequences, brain FLAIR and T1Gd sequences and H&N T1 sequences. The commercial deep-learning cycle-GAN model is a feasible method for sCT generation with high clinical utility due to its ability to use variable input data from multiple scanners and sequences. PD-0664 How do motion-compensated MRI techniques affect visualization of a moving object for RT planning? A. van Lier 1 , P. Borman 1 , M. Fast 1 , B. Raaijmakers 1 1 UMC Utrecht, Radiotherapy, Utrecht, The Netherlands Purpose or Objective Compressed SENSE (CS) and 3D VANE are MR techniques which recently became available for the 1.5 T Unity MR-linac (Elekta AB, Sweden). While 3D VANE scans are aimed at reducing motion artifacts by sharing k-space data between acquisitions, CS is used to decrease scan times by sparsely sampling k-space. In this phantom study we investigate the effect of using these techniques in relation to imaging moving structures under free breathing conditions for radiotherapy treatment planning. Materials and Methods A 4D phantom (ModusQA, Canada) with a moving insert containing a 30 mm spherical target was used. Imaging was performed with and without motion. The motion pattern was cos4-shaped with a peak-to-peak amplitude of 20 mm and