S2062
Physics - Image acquisition and processing
ESTRO 2026
could highlight anomalous regions for further human- expert review. Inference on a full 3D MR volume took less than 2 minutes for both models, suggesting that they could potentially run in parallel while additional MR sequences are being acquired with the patient still in the scanner. Results from the additional experiment with distorted MR images showed strong activation for DRAEM in the areas of low-quality (Figure 2) and highlighted the need for further finetuning of RD4AD.
Recognition (CVPR). 2017:5967-5976. doi:10.1109/CVPR.2017.632 Keywords: MR-only radiotherapy, synthetic CT, pelvic implant
Digital Poster Highlight 1606 Synthetic CT—to be or not to be generated: towards automating MR-only radiotherapy quality assurance with deep learning out-of-distribution detection Mariia Lapaeva 1,2 , Nico Camillo Zala 1,2 , Manuel Günther 2 , Nicolaus Andratschke 1 , Matthias Guckenberger 1 , Stephanie Tanadini-Lang 1 , Riccardo Dal Bello 1 1 Department of Radiation Oncology, University Hospital Zurich and University of Zurich, Zurich, Switzerland. 2 Department of Informatics, Artificial Intelligence and Machine Learning Group, University of Zurich, Zurich, Switzerland Purpose/Objective: In MR-only radiotherapy, where synthetic CTs (sCTs) are generated from MR images using deep learning (DL) methods, models may hallucinate1 when encountering out-of-distribution (OOD) MR inputs containing artefacts, implants or rare anatomies. This study investigates DL-based OOD detection to automatically identify atypical MR images before the step of sCT generation, potentially during MR simulation, allowing prompt decision making (sCT or CT) to increase the efficiency of MR-only quality
assurance workflow. Material/Methods:
In this study, two unsupervised DL OOD detection models2, namely DRAEM and RD4AD, were evaluated. These models learn the distribution of anomaly-free data and, during inference, detect spatial deviations from this distribution to generate anomaly maps and predict OOD cases. The models were trained on 4119 anomaly-free T1 Dixon in-phase axial MR slices from 36 pelvic tumor patients and tested on 796 slices from 12 patients with hip implants (OOD) and 807 slices from 8 patients without implants (anomaly-free cases). Metal-containing OOD slices were initially detected by thresholding CT images, marking regions with voxel values above 2000 HU. An additional experiment was conducted on MR slices from 2 patients with MR imaging artifacts. Model classification performance was evaluated using slice-level Area Under the Curve (AUC) and F1-score. Results: DRAEM and RD4AD showed promising performance in detecting out-of-distribution MR slices, with AUC/F1- score pairs of 0.87/0.77 and 0.75/0.72, respectively. Visual analysis (Figure 1) indicated that the models
Conclusion: Our findings indicate that DL-based OOD detection models hold the potential to enhance the safety and reliability of MR-only workflows by automatically assessing the suitability of MR images for sCT generation, thereby supporting informed decision- making early in the clinical workflow prior to CT acquisition or sCT generation. This motivates further research into extending the dataset and detecting more complex OOD scenarios with DL methods as well as developing 3D OOD prediction methods on a
Made with FlippingBook - Share PDF online