S2100
Physics - Image acquisition and processing
ESTRO 2026
Digital Poster 4917 Motion Correction in 4D CBCT Projections for Enhanced Reconstruction Johannes B Gebauer 1 , Laura E Büttgen 1,2 , Frederic Madesta 1 , Tobias Gauer 2 , René Werner 1 1 Institute for Applied Medical Informatics, Image Processing and Medical Imaging, University Medical Center Hamburg-Eppendorf, Hamburg, Germany. 2 Department of Radiotherapy and Radiooncology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany Purpose/Objective: Reconstructing 4D CBCT phase-images is typically a severely ill-posed inverse problem due to the low number and non-uniform distribution of phase- projections around the patient. 4D CBCTs therefore often exhibit streaking and blurring artifacts. To enhance the reconstruction with established methods and in turn increase the image quality, we aim to correct the motion in projections yielding us more projections per phase. For this task, we train and compare two 4D CBCT specific image-to-image deep neural networks. Material/Methods: Our dataset consists of 57 (34:8:15) lung 4D CTs which we forward-project to simulate CBCT projections needed for supervised network training and evaluation. As our model, we choose a 2D U-Net architecture. The network should generate a novel projection p̂ at gantry angle θ and phase k. Its input consists of the projection p* acquired at θ of another phase, plus the two nearest existing phase- k projections in terms of the gantry angle. One network variant should directly return p̂ (Direct U-Net) and the other a 2D optical flow field with which p* can be warped into p̂ (Flow U-Net). We train the networks with an image loss consisting of L1, SSIM, and a perceptual loss, combined with an additional regularization term for the flow field. For evaluation, we calculate the MSE of the generated normalized projections on forward-projected body and lung masks. Results:
The table in Sup. 1 shows the average MSE over the test set cases. The values for p* confirm the high similarity betweenp̂ and p*, especially outside of the lung region where little motion is expected (body w/o lung). The Flow U-Net better captures these expected static regions but falls behind the Direct U-Net in the lung mask, the latter overall achieving the best results. Examples of generated images and their differences to the ground truth are depicted in Sup. 2. Conclusion: Our models can generate novel 4D CBCT projections which closely resemble simulated ground truth data. We identify regions of the motion we aim to correct and observe, that the model via the 2D optical flow falls behind the direct method in regions of greater expected motion. Generating novel projections in this manner for each phase can in the next step lead to a significant increase in reconstruction quality. In the future, we aim to investigate the application on real clinical datasets. Keywords: 4D CBCT Reconstruction, Projection Prediction Digital Poster 4924 ACCURACY AND PRECISION OF THE 5DCT APPROACH Daniel A Low, Claudia Miller, Ryan Andosca, Michael Lauria, Jie Deng, Drew Moghanaki Radiation Oncology, UCLA, Los Angeles, USA Purpose/Objective: In our institution we have supplemented 4DCT with a reliable and quantitative model-based CT workflow that uses a mathematical model based on real-time surrogates of breathing amplitude and rate (5D)1. 5DCT separates breathing motion into the real-time component vis a vis the amplitude and rate and a mathematical model that provides the relationship between the real-time surrogates and the tissue motion. This work reports a retrospective evaluation of the 5DCT approach. Material/Methods: Our protocol acquires 25 fast helical CT scans that have reduced mAs, with one assigned as the reference scan. The session breathing amplitude is defined as the 5th to 95th percentile amplitudes. The surrogate is calibrated using one of the two diaphragm dome positions as detected in each of the 25 scans, the residual of which is considered a measure of surrogate
Made with FlippingBook - Share PDF online