S2305
Physics - Machine learning and AI algorithms
ESTRO 2026
References: [1] A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” First Conference on Language Modeling, May 2024. Keywords: Synthetic CT, Adaptive Radiotherapy, Deep Learning Proffered Paper 3259 Catching MRI outliers: unsupervised detection and localization of MRI artefacts and clinical anomalies using deep learning Mustafa Kadhim 1,2 , Viktor Rogowski 1,2 , Emilia Persson 2,3 , André Haraldsson 1,2 , Sofie Ceberg 1 , Malin Kügele 2,4 , Sven Bäck 1,2 , Mikael Nilsson 5 , Christian Jamtheim Gustafsson 2,3 1 Medical Radiation Physics, Lund University, Lund, Sweden. 2 Radiation Physics, Department of Hematology, Oncology, and Radiation Physics, Skåne University Hospital, Lund, Sweden. 3 Medical Radiation Physics, Department of Translational Medicine, Lund University, Malmö, Sweden. 4 Department of Radiooncology, Rostock University Medical Center, Rostock, Germany. 5 Centre for Mathematical Sciences, Lund University, Lund, Sweden Purpose/Objective: Image anomalies may reduce the performance of deep learningmodels and obstruct automated radiotherapy workflows. We aimed to develop a comprehensive anomaly detection framework, trained only on normal clinical prostate data, to automatically detect and spatially locate magnetic resonance image (MRI) artefacts and clinical anomalies using unsupervised deep learning. Material/Methods: A cohort of 432 T2-weighted prostate MRI volumes without anomalies from the LUND-PROBE dataset[1] (voxel size 1x1x2.5 mm ³ ) was selected. Global image artefacts were simulated (motion, noise, ghosting, Fourier space spikes, blur). A local image artefact was generated by averaging the signal in the prostate target (CTV average signal). All artefacts and artefact- free counter parts were produced in one validation (n=15pat) and one test cohort (n=15pat). Generalizability was assessed on a clinical anomaly set (n=21pat) including hip implants, brachytherapy cervix applicators, and prostate cases with hydrogel spacer. The framework comprised of a 64 × 64 2D patch-based Vector Quantized Variational Autoencoder (VQ-VAE-2)[2] for image reconstruction and latent-tokenization where patches were extracted from 256x448x88 volumes, and a Masked Generative Image Transformer (MaskGIT)[3] modeling token- likelihoods conditioned on latent codes and patch-
were integrated exclusively at the bottleneck layer to maximize global receptive field coverage while maintaining computational efficiency. The composite loss function combined perceptual loss, body-contour- masked mean absolute error (MAE), and artifact- region-specific masked MAE. Performance was evaluated using MAE, peak signal-to-noise ratio (PSNR), and multi-scale structural similarity index (MS- SSIM), assessed globally within the whole-body contour and locally in artifact-specific regions (dental and shoulder). Results: Qualitative assessment demonstrated effective artifact reduction in both dental and shoulder regions, with Mamba model producing anatomically accurate tissue contrast and reduced streak artifacts compared to baseline and ViT architectures (Fig.1). Quantitative evaluation confirmed substantial improvements across all metrics for the proposed Mamba model (Table.1). Wilcoxon signed-rank testing revealed statistically significant differences (p<0.05) for all Mamba model metrics compared to both baseline and ViT architectures.
Conclusion: The proposed Mamba-enhanced architecture achieves high-fidelity sCT generation with statistically significantly improved HU accuracy, outperforming both baseline and ViT models. The demonstrated superior multi-artifact reduction capability supports enhanced dosimetric accuracy for ART in H&N cancer patients.
Made with FlippingBook - Share PDF online