Unsupervised anomaly detection for longitudinal comparison in whole-body PET/CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Longitudinal comparison of 18F-fluorodeoxyglucose (FDG)-PET/CT scans is crucial for disease monitoring.
• Traditional subtraction-based methods often yield high false-positive rates.
• Supervised approaches require extensive lesion-annotated datasets, limiting their applicability.

Purpose of the Study:

• To evaluate an unsupervised anomaly detection method for longitudinal FDG-PET/CT analysis.
• To compare its performance against subtraction-based methods.
• To assess its ability to detect new lesions across various anatomical regions without prior annotation.

Main Methods:

• A 3D U-Net model was trained on normal PET/CT images.
• The model predicts current PET image parameters based on prior PET, prior CT, and current CT.
• Anomaly detection was performed using voxel-wise Z-scores derived from predicted parameters.

Main Results:

• The proposed unsupervised method achieved a higher area under the receiver operating characteristic curve (AUC) of 0.881 compared to the subtraction method (0.757).
• Superior performance was observed across neck (AUC=0.925), chest (AUC=0.918), and abdominal (AUC=0.824) regions.
• Sensitivity was significantly higher for the proposed method (0.728) versus the subtraction method (0.375) at 5.0 false positives per case.

Conclusions:

• Unsupervised anomaly detection effectively reduces false positives in longitudinal FDG-PET/CT comparisons.
• The method successfully highlights newly appearing lesions in diverse anatomical areas.
• It offers an alternative to supervised methods by not requiring lesion-annotated datasets.