Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 26·5 min read
Catching MRI outliers: unsupervised detection and localization of MRI artefacts and clinical anomalies using deep learning
TL;DR
MRI scans can contain artefacts and anomalies that affect diagnosis, but detecting these issues has been challenging. A two-stage unsupervised anomaly detection framework was developed, utilizing deep learning to analyze MRI images.
✦ Why It Matters
Engineers can implement this framework to improve MRI quality control and enhance diagnostic accuracy in clinical settings.
Key Takeaways
- Implement unsupervised anomaly detection in MRI workflows.
- Utilize deep learning for token-based image analysis.
- Evaluate model performance using AUC metrics.
- Incorporate heatmap visualizations for anomaly localization.
Full Summary
MRI scans are crucial for accurate medical diagnoses, yet they can contain artefacts and clinical anomalies that compromise their reliability. To address this, a two-stage unsupervised anomaly detection framework was developed, leveraging deep learning techniques.
In the first stage, MRI slices were compressed into discrete tokens, and in the second stage, the distribution of normal tokens was modeled. Anomaly detection was evaluated using synthetic and real clinical anomalies in pelvic MRI and clinically annotated abnormalities in brain MRI.
The framework achieved an area under the receiver operating characteristic curve (AUC) of 0.97 for pelvic MRI and 0.81 for brain MRI, indicating robust detection capabilities. Heatmap analysis confirmed strong spatial agreement between detected anomalies and their true locations, enhancing interpretability.
These findings suggest that this framework could serve as an automated quality-control layer in radiotherapy workflows.