S2306
Physics - Machine learning and AI algorithms
ESTRO 2026
coordinates. Inference used a cascading two-criteria strategy. First, perceptual reconstruction loss (PL; ResNet-50 pre-trained on ImageNet) identified volumes with global anomalies. PL patch-anomalies were defined by percentile thresholds ( ≥ 95th or ≤ 10th) and patch z-score>1.5; a volume was anomaly-positive if ≥ 10% of its patches were flagged. Second, volumes passing PL underwent patch-wise token negative log- likelihood (NLL) scoring against the learned normal- token distribution. Background patches were excluded. NLL anomalies used thresholds ( ≥ 98th or ≤ 2nd) and patch z-score>2.5; a volume was anomaly- positive if ≥ 5% of patches were flagged. Thresholds were optimized using the validation cohort. Results: The framework detected and localized global, local, and multiple clinical anomalies, achieving mean patient-level accuracy of 95.6% across simulated artefacts and 100% sensitivity for clinical anomalies. Qualitative examples (Fig.1) show spatial anomaly detection maps for simulated motion and noise (global), prostate CTV average signal (local), and clinical anomalies (hip implants, hydrogel spacer, brachytherapy cervix applicator). Quantitative detection metrics showed excellent performance (Table.1).
Conclusion: This study is the first to combine state-of-the-art deep learning methods to enable unsupervised MRI anomaly detection for global, local and multiple clinical anomalies simultaneously. All clinically relevant artefacts and anomalies, such as hip implants, spacers, and brachytherapy applicators, had a 100% detection sensitivity or accuracy with explainable and informative spatial locations. Thereby, image anomalies can be automatically detected, facilitating more reliable and robust automated clinical workflows. References: [1] Rogowski, V., Olsson, L. E., Scherman, J., Persson, E., Kadhim, M., af Wetterstedt, S., ... & Jamtheim Gustafsson, C. (2025). LUND-PROBE–LUND Prostate Radiotherapy Open Benchmarking and Evaluation dataset. Scientific Data, 12(1), 611. [2] Razavi, A., Van den Oord, A., & Vinyals, O. (2019). Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32. [3] Chang, H., Zhang, H., Jiang, L., Liu, C., & Freeman, B. (2022). MaskGIT: Masked Image Generative Transformers. Keywords: Anomaly detection, Deep learning, MRI
Made with FlippingBook - Share PDF online