ESTRO 2020 Abstract book

S989 ESTRO 2020

PO-1701 Optimized source-detector trajectories for low dose CBCT S. Hatamikia 1,2 , A. Biguri 3 , G. Kronreif 1 , H. Furtado 4 , M. Buschmann 4 , J. Kettenbach 5 , E. Steiner 6 , M. Gober 6 , D. Georg 4 , W. Birkfellner 2,4 1 Austrian Center for Medical Innovation and Technology- Wiener Neustadt- Austria, Austrian Center for Medical Innovation and Technology- Wiener Neustadt- Austria, Vienna, Austria ; 2 Medical University of Vienna- Austria, Center for Medical Physics and Biomedical Engineering, Vienna, Austria ; 3 University of Southampton- United Kingdom, Institute of Sound and Vibration Research, Southampton, United Kingdom ; 4 Medical University of Vienna- Austria, Department of Radiotherapy, Vienna, Austria ; 5 Department of Diagnostic and Interventional Radiology and Nuclear Medicine- Wiener Neustadt- Austria, Department of Diagnostic and Interventional Radiology and Nuclear Medicine- Wiener Neustadt- Austria, Vienna, Austria ; 6 Institute for Radiooncology and Radiotherapy- Landesklinikum Wiener Neustadt, Institute for Radiooncology and Radiotherapy- Landesklinikum Wiener Neustadt, Vienna, Austria Purpose or Objective Novel radiotherapy imaging devices might allow non- coplanar on-board imaging. We develop a target-based cone beam computed tomography (CBCT) imaging framework in order to optimize the source–detector trajectory by incorporating prior information available from prior CT images. Our main aim is to enable the CBCT system to provide topical information about the target using a short-scan trajectory with a minimal number of projections and therefore applying minimum dose to the patient. Material and Methods A patient specific model from a prior diagnostic CT is used as a digital phantom for CBCT trajectory simulations. We propose a trajectory optimization scheme which aims to minimize the radiation dose by means of reducing projections while keeping a reasonable imaging performance at the region of interest (ROI). Selection of the best projection views is accomplished through maximizing an objective function fed by the imaging quality of different arbitrary x-ray position in 3D space on the digital phantom data. We use Structural Similarity Index (SSIM) as the objective function for trajectory optimization and Adaptive steepest descent projection onto convex sets (ASD-POCS) algorithm for the reconstruction. The final optimized trajectory selected includes a minimal number of projections which can be applied to a C-arm device capable of general source–detector positioning. We use a Philips Allura FD20 Xper C-arm for our experiments. The performance of the proposed framework is investigated with an Alderson-Rando head phantom with considering neck as the ROI. The dose delivered with the optimized trajectory compared to standard CBCT was measured using computed tomography dose index (CTDI). Results Our experiments based on the head phantom showed that our optimized trajectory could achieve a comparable image quality with respect to the reference C-arm CBCT while using one quarter of projections. The reconstruction results related to the optimized trajectory based on both real data and simulations are presented in Figure 1. (b, c). In addition, C-arm CBCT based on the standard circular trajectory is presented in Figure 1.a for the comparison. We use Feature Similarity Index (FSIM) in order to quantify the reconstruction results based on the real data. Relative deviation of FSIM value was achieved 7.54% between the reconstructed image (Figure 1. b) and the reference CBCT (Figure 1.a). 3D visualization of the optimized compared with standard C-arm trajectory is shown in Figure 2. The CTDI measured for the standard CBCT trajectory and the optimized trajectory was 15.5 mGy and 3.38 mGy,

The CT images of these cases were converted from DICOM RT formats to arrays of 32 × 128 × 128 voxels and input into both 2-D and 3-D U-Net. The number of training, validation and test sets were 160, 40 and 32, respectively and a 5- fold cross-validation was performed to make models more generalizable. Dice similarity coefficients (DSCs) of the test set were evaluated employing Smart segmentation as well as the 2-D and 3-D U-Net. Results The optimal learning rate was found to be 0.001 for 2-D U- Net and 0.0001 for 3-D U-Net. Table 1 shows the mean DSC with the standard deviation, 95% confidence interval (CI), maximum, median and minimum of DSC for each technique. The mean DSCs of the test set were 0.964 (95% Confidence Interval, 0.960–0.968), 0.990 (95% CI, 0.989– 0.992) and 0.990 (95% CI, 0.989–0.991) with Smart segmentation, 2-D and 3-D U-Net, respectively. Compared to Smart segmentation, both U-Nets presented significantly higher DSCs by the Wilcoxon signed-rank test (p <0.01). There was no difference in mean DSC between the 2-D and 3-D U-Net systems. Fig. 1 shows the main bronchi present in a slice selected arbitrarily from the test set of 32 cases. Compared with the ground truth image, Smart segmentation depicted the main bronchi and other structures incorrectly, while 2-D and 3-D U-Net showed all but the main bronchi correctly.

Conclusion The newly-devised 2-D and 3-D U-Net approaches were found to be more effective than a commercial auto- segmentation tool. Even the relatively shallow 2-D U-Net which does not require high-performance computational resources was effective enough for the lung segmentation. Semantic segmentation using deep learning was useful in radiation treatment planning for lung cancers.