MR Anomaly Detection
Overview
Developed deep learning-based anomaly detection methods 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.
🎯 Goal: Enable automatic pre-screening for MR-only radiotherapy workflows by detecting out-of-distribution patterns (metallic implants, artifacts) in MR images that could compromise synthetic CT quality.
Key Contributions
- Dataset Annotation: Annotated an OOD dataset from SynthRAD2023 for metal artifact detection in pelvic MR images, preprocessed NIfTI files into anomaly detection datasets
- Model Benchmarking: Evaluated 10 unsupervised anomaly detection models across 3 input formats using Anomalib framework
- Multi-level Evaluation: Comprehensive evaluation at pixel-, slice-, and patient-level metrics
- Post-processing Pipeline: Built morphological filtering and volumetric consistency pipeline achieving 0.90 patient-level Dice with PatchCore
- Multi-center Validation: Validated generalizability across different scanners with 0.77 Dice
Technical Approach
Anomaly Detection Methods Evaluated
- Memory-Bank Methods: PatchCore, PaDiM, CFA
- Knowledge Distillation: RD4AD, STFPM
- Flow-Based: FastFlow, CFlow
- Reconstruction-Based: DRAEM
- One-Class: DeepSVDD, CutPaste
Input Representations
- Replicated MR: Single MR slice replicated to 3 channels
- 2.5D Adjacent: 3 consecutive slices as RGB channels for spatial context
- Bone Colormap: MR + bone segmentation overlay for anatomical guidance
Post-Processing Pipeline
- Morphological operations (erosion, dilation) for noise removal
- Connected component analysis for artifact region refinement
- Volumetric consistency checks across adjacent slices
- Threshold optimization for clinical sensitivity requirements
Results
🏆 Single-Center Performance
0.90 Patient-Level Dice achieved with PatchCore after post-processing optimization
🌐 Multi-Center Generalization
0.77 Dice Score maintained across different scanner types and imaging protocols
Key Finding: Memory-bank method PatchCore achieves the most favorable trade-off between localization accuracy and patient-level detection, remaining comparatively robust under multi-center domain shifts.
Clinical Impact
- Supports automatic pre-screening for MR-only radiotherapy workflows
- Enables early detection of problematic MR acquisitions before synthetic CT generation
- Reduces manual QA burden while maintaining patient safety standards
- PatchCore-style memory methods identified as promising for future clinical MR-only QA pipelines