S2049
Physics - Image acquisition and processing
ESTRO 2026
with Syngo.via RT Image Suite, n.d.[3] K.N.D.D. Brou Boni, J. Klein, A. Gulyban, N. Reynaert, D. Pasquier, Med Phys 48 (2021) 3003–3010. Keywords: synthetic CT, autosegmentation, data QC
structures. Low- and high-resolution segmentations achieved comparable results, with negligible degradation inaccuracy (volume: p = 0.98, p= 0.93 and ED: p= 0.86, 0.82 for sCT and ksCT respectively), while low-resolution showed a 5-fold increase in segmentation speed (67-75 vs. 398-468 s). ROI-based volume and ED evaluations showed strong agreement between sCT and KsCT datasets, supporting the robustness of the automated QC process (Figure1). The generated QC report included demographic data, 3D visualization snapshots, and per-ROI thresholds to highlight potential outliers, as in Figure2. Total QC execution time, including segmentation and report generation, was under 10 minutes using GPU- accelerated hardware (CUDA 12.4, GPU 4GB Nvidia Quattro P1000).
Digital Poster 1037 How to fix a synthetic CT: a Deep Learning model to predict the voxel-wise conversion error Lorena Romeo 1 , Ciro Benito Raggio 2 , Michela Destito 1 , Carlo Cosentino 1 , Stefan Both 3 , Maria Francesca Spadea 2 , Paolo Zaffino 1,3 1 Department of Experimental and Clinical Medicine, Magna Graecia University, Catanzaro, Italy. 2 Institute of Biomedical Engineering, Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany. 3 Department of Radiation Oncology, University Medical Center Groningen, Groningen, Netherlands Purpose/Objective Synthetic computed tomography (sCT) images generated from MRI scans play a fundamental role in the MRI-only radiotherapy workflows [1]. However, unlike automatic segmentations, whose output can be visually verified, the quality of sCT generation can not be directly assessed by humans. Consequently, residual inaccuracies in Hounsfield Unit (HU) estimation may lead to dosimetric errors, compromising clinical reliability. Recent works have proposed deep learning (DL) frameworks to generate sCTs and estimate the DL model uncertainty. However, such information could be misleading, as a model can be highly confident while being completely wrong. The key aspect lies in predicting the actual conversion error, rather than the model’s confidence [2]. We propose a method that directly predicts the voxel-wise signed error within the sCT, enabling fine-grain correction of local intensity inaccuracies and providing clinicians with a spatial error map. Material/Methods A 2D DL model was developed to predict the signed pixel-wise deviation between sCT and ground truth CT images (as in Fig. 1), in which positive and negative values represented local over- and under-estimations of HU compared to the ground truth CT. The 2D maps were stacked to obtain the final 3D voxel-wise error volume. This information is used both to inform clinicians and to correct the sCT. The proposed approach was evaluated using a leave-one-out cross- validation strategy on a head dataset of 15 brain tumor patients. Performance was assessed in terms of mean absolute error (MAE) improvements and bias reduction.
Conclusion: The presented workflow delivers rapid, reproducible QC of pelvis sCTs without requiring CT references. The combination of independent GAN generation and TotalSegmentator analysis provides objective validation for MR-only clinical workflows. Future work will refine ROI selection and implement automatic DICOM listening for real-time integration. References: [1] J. Wasserthal, H.-C. Breit, M.T. Meyer, M. Pradella, D. Hinck, A.W. Sauter, T. Heye, D.T. Boll, J. Cyriac, S. Yang, M. Bach, M. Segeroth, Radiol Artif Intell 5 (2023).[2] M. Hoesl, N.E. Corral, N. Mistry, MR-Based Synthetic CT Reimagined An AI-Based Algorithm for Continuous Hounsfield Units in the Pelvis and Brain-
Made with FlippingBook - Share PDF online