Geometry-invariant abnormality detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Cancer is a heterogeneous disease, making detection challenging.
• Unsupervised anomaly detection models show promise for general-purpose cancer detection.
• Existing models struggle with variations in data geometry (e.g., resolution, field of view).

Purpose of the Study:

• To develop a novel spatial conditioning mechanism for unsupervised cancer detection models.
• To enhance the adaptability of anomaly detection models to varying data geometries.
• To improve the accuracy and robustness of cancer detection in positron emission tomography (PET).

Main Methods:

• Applied a spatial conditioning mechanism to a Vector-Quantized Variational Autoencoder + Transformer (VQ-VAE+Transformer) model.
• Utilized unsupervised learning for anomaly detection in medical images.
• Evaluated model performance on whole-body PET data with varying geometries.

Main Results:

• The proposed spatial conditioning mechanism significantly improved model performance.
• The enhanced model demonstrated increased robustness to changes in image resolution and field of view.
• The model successfully performed inference across different data geometries.

Conclusions:

• Spatial conditioning is a statistically significant improvement for VQ-VAE+Transformer based abnormality detection.
• This approach enhances the reliability of AI models for cancer detection in diverse PET imaging scenarios.
• The method allows for adaptable and accurate cancer detection despite variations in imaging data geometry.