Automated MR-Only RT for H&N cancer: synthetic CTs, autocontouring and automated treatment planning
Iris Lauwers,
The Netherlands
PD-0324
Abstract
Automated MR-Only RT for H&N cancer: synthetic CTs, autocontouring and automated treatment planning
Authors: Steven Petit1, Sandeep Kaushik2, László Rusko3, Cristina Cozzini2, Borbala Deák-Karancsi3, Vanda Czipczer3, Bernadett Kolozsvári3, Katalin Hideghéty4, András Frontó3, Botond Maros3, Jean-Paul Kleijnen5, Jonathan Wyatt6, Hazel McCallum7, Gerda Verduijn1, Florian Wiesinger2, Juan Hernandez-Tamames8, Marta Capala1
1Erasmus MC Cancer Institute, University Medical Center Rotterdam, Radiotherapy, Rotterdam, The Netherlands; 2GE Healthcare, MR Applied Science Laboratory, Munich, Germany; 3GE Healthcare, Healthcare Digital, Budapest, Hungary; 4University of Szeged, Department of Oncotherapy, Szeged, Hungary; 5Haaglanden Medical Center, Department of Medical Physics, The Hague, The Netherlands; 6Newcastle University, Translational and Clinical Research Institute, NewCastle, United Kingdom; 7Newcastle University, Translational and Clinical Research Institute, Newcastle, United Kingdom; 8Erasmus MC, University Medical Center Rotterdam, Radiology and Nuclear Medicine, Rotterdam, The Netherlands
Show Affiliations
Hide Affiliations
Purpose or Objective
Head-and-neck (H&N) cancer patients often have fast growing tumors and therefore benefit from an early start of treatment. Three factors that cause substantial time delay are (i) acquisition of both a MR and CT scan (ii) contouring of organs at risk and (iii) treatment planning. The goal of this study was to evaluate automated MR-only RT for H&N, where (i) synthetic CTs (sCT) are generated from Zero TE MR scans (i.e., omitting the need for a separate planning CT), (ii) where the organ-at-risks (OARs) are automatically segmented from T2 MR and (iii) treatment planning is performed fully automatically.
Material and Methods
Thirty-two patients from the Deep MR-Only RT trial (NL8167) were used to train the methods for sCT generation and auto-contouring. Patients were scanned with 3D Zero TE (2min33sec) and 2D T2 (3min29sec) MR sequences on a 1.5T GE SIGNA MR450w scanner using a dedicated H&N RT coils setup. A supervised DL algorithm was developed to convert the Zero TE MR scans into sCT images [1]. A combination of 2D (UNET) and 3D (VNET) was trained for automated OAR contouring on T2 MR images [2].
The sCT and autocontouring methods were validated for the 8 patients in the validation cohort. The clinical treatment VMAT plans were generated fully automatically using Erasmus-ICycle and Monaco [3] based on the clinical (i.e. manual) contours and the planning CT. Next the plans were recalculated on both the planning CT and sCT scan using the Monte Carlo algorithm of Scimoca 1.5 (Scientific RT GmbH). The mean absolute error (MAE) in HU and DVHs differences of the clinical contours calculated on the sCT vs. the planning CT were calculated. The cumulative effect of automated OAR contouring in combination with sCT for dose calculations was assessed by comparing DVHs of the automated OAR contours on the sCT with DVHs of the clinical contours on the planning CT (=ground truth).
Results
For the 8 patients of the validation cohort, the MAE between the planning CT and sCT was 104 ± 9.1 HU within the body region. Excellent agreement was found in DVH parameters between the sCT and planning CT (evaluated on the clinical contours) with differences in PTV Dmean of 0.4 Gy (range 0.1 -0.6), V95 0.8% (0.1- 1.4%), and V107 0.2% (0 – 0.7%) and 0.1 Gy (-0.5, 0.7) for the Dmean of the OARs (Fig 1a). The difference in mean dose (mean ± SD) between the automated contours and sCT vs the clinical contours and planning CT was only 1.1 ± 2.9 Gy in Dmean for parallel organs and for spinal cord and brain stem the near maximum dose (D0.05cc) difference was 0.9 ± 4.5 Gy (Fig 1b). Fig 2 shows an example patient.
Conclusion
The dosimetric accuracy of the synthetic CTs was clinically acceptable. Automated MR-Only RT yielded dose differences below 1.1 Gy, which is similar to typical differences due to manual contour inter-observer variation. To conclude, with the aim to reduce time to treatment, we demonstrated the feasibility of automated MROnly RT with sCT generation, automated OAR contouring and automated treatment planning.
