ESTRO 2021 Abstract Book

S525

ESTRO 2021

Conclusion For multimodality deep learning-based HNSCC GTV segmentation, DR reduces MSD but has no significant effect on Dice or HD95 compared to RR. Both DCN and CHTL with prior RR achieved end-to-end solutions with comparable performance to DR UNet. CHTL requires less computing resources and shorter training time. PH-0655 Filling in the allegedly unknown: Detection and inpainting of interpolation artifacts in 4DCT images F. Madesta 1 , T. Sentker 1 , T. Gauer 2 , R. Werner 1 1 University Medical Center Hamburg-Eppendorf, Department of Computational Neuroscience, Hamburg, Germany; 2 University Medical Center Hamburg-Eppendorf, Department of Radiotherapy and Radio-Oncology, Hamburg, Germany Purpose or Objective Interpolation artifacts are one of the most common artifacts in 4D CT imaging of lung and liver lesions, mainly caused by non-optimal imaging protocols and irregular breathing patterns. Both may lead to a violation of the data sufficiency condition essential for image reconstruction, which results in naïve image interpolation for the non-reconstructed slices. As, for instance, deformable image registration algorithms rely on image intensities, interpolation artifacts severely affect the plausibility and decrease the overall utility of calculated displacement vector fields for treatment planning. In this study, we propose a deep learning-based approach for detection and inpainting of interpolation artifacts and evaluate it in terms of image quality improvement. Materials and Methods The study includes lung (N=75) and liver (N=49) tumor patients with a 10 phase 4D CT each. We used 93 randomly picked patients for training both the interpolation artefact detection and inpainting model. We inserted simulated interpolation artifacts into artifact-free areas of the phase images (1-4 artifacts per image, 1-10 slices in size). A modified U-Net was trained to detect these artifact-affected slices, while a second U- Net was optimized for inpainting of said slices. The latter network is trained in a conditional fashion by feeding both the patient’s average CT as well as the artifact-affected image into the network. The inpainting process is performed by partial convolutions inside the U-Net iteratively filling the artifact from the outside in. This ensures that inpainting only occurs in the neighborhood of non-artifact/already inpainted voxels. To evaluate image quality improvement, we utilized the remaining 31 4D CT sets and calculated the root-mean- square error (RMSE) and normalized cross-correlation (NCC) for the artifact-affected and the inpainted slices. Since the artifacts were simulated, the ground truth is known and acts as the reference image for calculation of the measures. Results The artifact detection robustly detects the simulated interpolation artifacts even with the smallest artifact