Related Experiment Video
Updated: Sep 12, 2025
12:24
Computed Tomography-guided Time-domain Diffuse Fluorescence Tomography in Small Animals for Localization of Cancer Biomarkers
Published on: July 17, 2012
12.5K
IgCONDA-PET: Weakly-supervised PET anomaly detection using implicitly-guided attention-conditional counterfactual
1Department of Physics & Astronomy, University of British Columbia, Vancouver, BC, Canada; Department of Integrative Oncology, BC Cancer Research Institute, Vancouver, BC, Canada.
Summary
This study introduces IgCONDA-PET, a new weakly-supervised method for detecting anomalies in PET scans. It reduces the need for expert annotations by generating healthy versions of scans to identify differences.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Pixel-level annotations for PET lesion detection are time-consuming and costly.
- Current unsupervised/weakly-supervised methods often use autoencoders or GANs, with GANs facing training instability.
- There's a need for efficient anomaly detection in PET imaging with reduced annotation dependency.
Purpose of the Study:
- To present IgCONDA-PET, a novel weakly-supervised diffusion model for anomaly detection in PET images.
- To reduce reliance on pixel-level annotations for training PET lesion detection and segmentation networks.
- To enable robust anomaly detection in multi-center, multi-cancer PET datasets.
Main Methods:
- Developed a weakly-supervised Implicitly guided COuNterfactual diffusion model for Detecting Anomalies in PET images (IgCONDA-PET).
- Utilized attention modules for class label conditioning (healthy vs. unhealthy) and implicit diffusion guidance.
- Employed counterfactual generation for "unhealthy-to-healthy" domain translation to detect anomalies via image differences.
Main Results:
- Validated on 6 retrospective cohorts (2652 cases) using multi-cancer, multi-tracer PET scans.
- Demonstrated the effectiveness of attention modules for detecting small anomalies.
- Outperformed traditional methods (e.g., SUVmax thresholding) and other deep learning approaches in weakly-supervised anomaly detection.
Conclusions:
- IgCONDA-PET offers a transformative approach to PET anomaly detection by minimizing annotation requirements.
- The method shows strong performance across diverse PET datasets, highlighting its clinical potential.
- Publicly available code facilitates further research and application in medical image analysis.
