ESTRO38 Congress Report

Physics

5. Correcting CBCT images for dose calculation using a U shaped deep convolutional neural network (E38-0312) Guillaume Landry 1,2 , David Hansen 3 , Florian Kamp 2 , Minglun Li 2 , Ben Hoyle 4 , Jochen Weller 4,5 , Katia Parodi 1 , Claus Belka 2,6 , Christopher Kurz 2,1 1 Department of Medical Physics, Faculty of Physics, Ludwig-Maximilians-Universität München, Garching, Germany, 2 Department of Radiation Oncology, University Hospital, LMU Munich, Munich, Germany, 3 Gradient Software, Aarhus, Denmark, 4 University Observatory, Faculty of Physics, Ludwig-Maximilians-Universität München, Munich, Germany, 5 Max Planck Institute for Extraterrestrial Physics, Garching, Germany

Context of the study Treatment plan optimization at each fraction, in an online- adaptive workflow, may provide optimal dose delivery in radiotherapy (RT). Online adaptive RT requires a combination of imaging, contouring, dose calculation and optimization performed in a narrow time window. When employing cone beamcomputed tomography (CBCT) based image guidance, a crucial step is the retrieval of accurate radiological properties allowing dose calculation. Given CBCT’s well documented image quality shortcomings, image correction is crucial when using CBCT for dose calculation. For online-adaptive RT, the corrections should be fast in addition to allowing accurate dose calculation. The success of deep convolutional neural networks (CNN) in a wide range of image processing tasks, and their near-real-time inference speed, suggest that this class of algorithms may be very well suited to tackle online- adaptive RT challenges. Overview of abstract We trained a U-shaped deep CNN (U-net), to generate dose-calculation-ready CBCT images of prostate cancer patients. The U-net consists of a contracting path similar to a conventional CNN, and an expansive path along which features from the contracting path are combined with up-sampled ones (see Figure 1). The goal was to train the U-net to provide corrected volumetric CBCT images in a few seconds. To ensure that the U-net output was adequate for online-adaptive radiotherapy, both photon and proton therapy doses calculations were performed, using one slower, established correction method as reference. What were the three main findings of your research? The same U-net could be trained to either correct raw CBCT projections suffering from scatter contamination, or already reconstructed CBCT images suffering from scatter related artefacts, highlighting the versatility of the U-net architecture. In both cases, the time to correct either a two-dimensional projection or an image slice was about 10 milliseconds, allowing fast application to an entire CBCT volume within about 3 seconds, a timeframe compatible with online-adaptive workflows. Volumetric arc therapy dose distributions calculated on U-net-corrected images were within 1% from those obtained with the reference method. For proton therapy, we could replicate range within 3 mm for 90% of depth dose curves and achieve >90% pass rates for 2%/2mm gamma evaluation. Figure 2 presents typical dose calculations and their differences for the U-net trained on reconstructed images.

What impact could your research have? The rapid inference of the trained U-net means that CBCT projections could be corrected as they are being acquired, leading to a dose-calculation-ready CBCT image immediately after reconstruction is completed. Thus, deep learning solutions, such as the one we present, may serve as crucial building blocks for online-adaptive RT when coupled to modern, fast graphical processor unit (GPU) based dose calculations. The versatility of the U-net also suggests that a similar method would be successful for other sites, as well as for pseudo CT generation in magnetic resonance (MR) imaging guided RT, as presented by other authors. Advanced CNNs based on U-nets, such as generative adversarial networks (GAN), may also simplify training owing to their unpaired nature, meaning they can be trained using CBCT and CT images of different patients. Is this research indicative of a bigger trend in oncology? The current RT enthusiasm towards the adoption of deep learning methods indicates that the field has identified their potential for workflow automation and acceleration. Besides the generation of dose calculation images, deep learning may help reduce workloads related to organ at risk and target delineation in online adaptive RT. Whether applied to conventional CBCT-linacs or to MR-linacs, deep learning will most likely play a role in the future of adaptive RT, as it currently does in many other fields.

Congress report | PHYSICS

19