Related Experiment Video
Updated: May 10, 2026
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
7.4K
Generative Modeling for Interpretable Anomaly Detection in Medical Imaging: Applications in Failure Detection and
McKell E Woodland1,2, Mais Altaie1, Caleb S O'Connor1
1Departments of GI Radiation Oncology, Imaging Physics, Interventional Radiology, and Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX 77030, USA.
Bioengineering (Basel, Switzerland)
|October 29, 2025
Summary
Generative models using StyleGAN2 effectively detect anomalies in medical images, improving AI failure detection interpretability and aiding large-scale data curation for datasets like ChestX-ray14.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Machine Learning
Background:
- AI systems in medical imaging require robust failure detection and efficient data curation.
- Generative models offer potential for anomaly detection and interpretability.
Purpose of the Study:
- To leverage generative modeling for enhanced AI failure detection interpretability.
- To utilize generative models for improved data curation in large medical image repositories.
Main Methods:
- Retrospective study using CT scans and ChestX-ray14 radiographs.
- StyleGAN2 networks modeled training data distributions.
- Anomaly detection via image reconstruction scoring (MSE, WD) and AUROC analysis.
Main Results:
- Generative models successfully detected anomalous attributes (needles, ascites) in unseen data.
- Mean AUROC for failure detection was 0.86 (±0.13), and for data curation was 0.82 (±0.11).
- Accurate localization of anomalies (81% ±13%) and differential performance of MSE/WD metrics observed.
Conclusions:
- Generative models demonstrate promise for interpretable AI failure detection.
- This approach facilitates unsupervised anomaly detection and aids large-scale medical data curation.
