CBCT-to-CT synthesis with a single neural network for head-and-neck, lung and breast radiotherapy
PD-0310
Abstract
CBCT-to-CT synthesis with a single neural network for head-and-neck, lung and breast radiotherapy
Authors: Maspero|, Matteo(1)*[m.maspero@umcutrecht.nl];Houweling|, Antonetta C(1);Savenije|, Mark H F(1);van Heijst|, Tristan C F(1);Verhoeff|, Joost J C(1);Kotte|, Alexis N T(1);van den Berg|, Cornelis A T(1);
(1)UMC Utrecht, Radiation Oncology, Utrecht, The Netherlands;
Show Affiliations
Hide Affiliations
Purpose or Objective
CBCT plays a crucial role in IGRT for patient position verification. The HU consistency of CBCT is not sufficient to perform accurate dose calculations due to image artefacts such as scatter and streaking. However, if CBCT can be used to perform dose calculations, online adaptive radiotherapy can be facilitated on conventional linacs.
Techniques as look-up tables, DIR and model- or Monte Carlo-based methods for scatter correction have been proposed to improve the quality of CBCTs for dose calculations. These techniques can be effective but are generally too slow (e.g. time scale of minutes) to enable online dose evaluation or online plan adaptation. Recently, deep learning has been proposed for fast CBCT artefact correction resulting in CBCT converted to CT in a matter of seconds.
In this study, we investigated whether CBCTs converted with convolutional networks may be used for dose calculations. We employed networks trained in an unpaired manner to convert CBCT-to-CT of head-and-neck (HN), lung, or breast cancer patients investigating whether a single network can generalise for the three anatomical sites compared to site-specific networks.
Material and Methods
Ninety-nine patients diagnosed with HN, lung or breast cancer undergoing radiotherapy with CBCT-based position verification were included in this study. CBCTs were registered to planning CTs according to the clinical procedure. Per anatomical site, a cycle-consistent generative adversarial network (cycle-GAN) was trained on 15 patients. Next, a network was trained with all the anatomical sites together. For each site, 10 synthetic-CTs (sCTs) were generated using the single and site-specific networks. The sCTs of each network were compared in terms of mean absolute error (MAE) against CT. To overcome the limited FOV of the CBCT and enable planning, the missing information outside the CBCT-FOV was filled from the CT. Clinical plans were recalculated on sCT and analysed through voxel-based dose differences and γ-analysis for the single network.
Results
An sCT was generated in 10 seconds (Fig1). Image similarity was comparable between networks trained on different anatomical sites and a single network trained for all sites: average MAE differed up to 3HU only (Tab1). Mean dose differences of less than 0.5% were obtained in high-dose regions. Mean gamma (2%,2mm) pass-rates >89.4% were achieved for all sites (Tab1).
Conclusion
Cycle-GANs reduced CBCT artefacts and increased HU accuracy, enabling CBCT-based dose calculations. A single network was evenly accurate to site-specific networks. The speed of the network can facilitate online adaptive radiotherapy using a single network for head-and-neck, lung and breast cancer patients.