ESTRO 2020 Abstract book

S1016 ESTRO 2020

Purpose or Objective The aim of this study was to evaluate the feasibility of a deep convolutional neural network (CNN) based method for generating synthetic CT (sCT) data in the head and neck (H&N) region. In an MRI only workflow, sCT data generated from MRI data are used as basis for absorbed dose calculation in radiotherapy, thus excluding the standard dose planning CT from the workflow. The quality of sCT data has to this point not been thoroughly validated for the H&N region. Material and Methods The method for sCT generation was evaluated using 44 H&N cancer patients with both CT and MRI data. A T1w Dixon Vibe (3D spoiled GRE) sequence was used for sCT data generation via MRI Planner v2 (Spectronic Medical AB, Sweden). The software generates sCT data via a CNN based method, tra ined on 66 MRI and CT data pairs (44 H&N separate from the evaluated data, 20 brain data pairs). The original clinical CT based treatment plan was transferred to and recalculated based on sCT data. Relative local absorbed dose difference (sCT – CT)/CT)*100 was calculated between the two treatment plans for a subset of DVH parameters. Paired statistical tests of equivalence were conducted for the relative absorbed doses. The absorbed dose distribution was further investigated using 3D gamma evaluation with a gamma pass rates of 2%/1mm. To validate the geometry of the sCT data sets, mean absolute error (MAE), Dice similarity coefficient (DSC) and water equivalent depth (WED) difference were evaluated against deformed CT data that was non-rigidly registered to the MRI data. Results Generation of sCT data was completed within a few minutes. Mean MAE (± 1 SD) for all data set were 66.8 ± 13.7, 195 ± 27.9, and 198 ± 67.8 HU for overall body, bones and air structures respectively. Mean DSC (± 1SD) for all data sets were 0.98 ± 0.048 for overall body, 0.80 ± 0.071 for bones and 0.81 ± 0.12 for air. Mean WED difference (± 1SD) was 0.324 ± 3.40, -1.08 ± 2.03 and - 0.730 ± 3.78 mm for the Th1-C7, mid mandible and mid nose region in that order. The maximum mean deviation in absorbed dose for all parameters was 0.30%, corresponding to 0.12 Gy. The absorbed dose difference for all 44 paired data sets can be seen in Figure 1. The absorbed doses were considered equivalent (p-value < 0.001) according to the paired statistical test of equivalence (equivalence bonds: -1%, 1%). The mean gamma passing rate for the absorbed dose distributions was 91.7 ± 5.7%. CT data with severe dental streak artifacts and MRI and generated sCT data with minor dental artefacts can be seen in Figure 2. Particularly, treatment sites in the dental region could benefit from using sCT data compared to CT data.

Conclusion It can

be the method evaluated in this study generates sCT data for H&N cancer that can be used in an MRI only workflow for radiotherapy. The sCT data allows for accurate dosimetric calculations and the absorbed doses were considered equivalent and close to zero compared to CT data. PO-1739 A deep learning neural network to remove metal artefacts via residual learning for cone-beam CT M. Skaarup 1 , J. Edmund 2 , M. Kachelriess 3 , I. Vogelius 1 1 The Finsen Center - Rigshospitalet, Clinic of Oncology, Copenhagen, Denmark ; 2 Herlev and Gentofte Hospital, Radiotherapy Research Unit- Department of Oncology, Herlev, Denmark ; 3 German Cancer Research Center, Division of X-Ray Imaging and CT, Heidelberg, Germany Purpose or Objective Cone-beam CT (CBCT) is used for daily imaging for treatment setup and monitoring, but image quality is often compromised by metal artefacts. We investigate the ability of a deep learning algorithm to remove these artefacts while preserving the correct anatomical information. Material and Methods Metal artefacts were artificially simulated to train a supervised convolutional neural network (CNN). These data were created using clinical CBCT scans. 51 CBCT scans from different patients without metal artefacts and 34 CBCT scans with metal artefacts were used. These included scans from the head and neck, thorax and pelvis scan regions and artefacts from dental implants, metal markers, pedicle screws and hip protheses. Each sample for training was created by segmenting a random metal implant and subsequently rotate it and place it in a random concluded that