This report investigates deep learning-based anomaly detection as a patient-specific quality assurance step for MR-only radiotherapy, focusing on detecting metal-induced artifacts in pelvic MR images before synthetic CT generation. The study evaluates multiple unsupervised and one-class approaches, including reconstruction-based, knowledge distillation-based, flow-based, and memory-bank methods, on a curated subset of the SynthRAD2023 pelvis dataset. The models are compared across pixel-, slice-, and patient-level metrics, and the impact of input representations (replicated MR channels, 2.5D adjacent slices, and bone colormaps) as well as ImageNet versus RadImageNet pretraining for FastFlow is analyzed. Among all evaluated configurations, the memory-bank method PatchCore achieves the most favorable trade-off between localization and patient-level detection, reaching a post-processed best patient-level Dice of 0.90 on the single-center setting, and remaining comparatively robust under multi-center domain shifts. These findings suggest that anomaly detection on MR alone can support automatic pre-screening for MR-only workflows and that PatchCore-style memory methods are a promising component for future clinically oriented MR-only QA pipelines, although further prospective and multi-center validation remains necessary.
MR-OOD
Popular repositories Loading
Repositories
Showing 6 of 6 repositories
- Anomaly-Detection Public
MR-OOD/Anomaly-Detection’s past year of commit activity - ood-train Public
MR-OOD/ood-train’s past year of commit activity - mp_visualizations Public
MR-OOD/mp_visualizations’s past year of commit activity - OOD-Data-Preprocessing Public
MR-OOD/OOD-Data-Preprocessing’s past year of commit activity - .github Public
MR-OOD/.github’s past year of commit activity - KD-OOD Public
MR-OOD/KD-OOD’s past year of commit activity
People
This organization has no public members. You must be a member to see who’s a part of this organization.
Top languages
Loading…
Most used topics
Loading…