Normal twin PET: personalized generative modeling for confounder correction and anomaly detection in whole-body...

Area of Science:

• Nuclear Medicine
• Artificial Intelligence
• Medical Imaging

Background:

• Distinguishing normal from pathological [18F]FDG uptake in PET/CT is challenging due to variable patterns and limited labeled data.
• Automated analysis requires robust methods to account for physiological variations and patient-specific factors.

Purpose of the Study:

• To develop a deep learning approach for generating patient-specific normal twin PET images (ntPET) for quantitative analysis.
• To enable unsupervised detection of pathological anomalies in whole-body PET/CT imaging.
• To improve the accuracy of quantitative measurements by correcting for confounding physiological variables.

Main Methods:

• Developed an image-to-image generative model synthesizing ntPET from CT scans, demographics, and acquisition parameters.
• Trained the model on 2,538 pseudo-normal PET/CT studies, including disease-masked scans.
• Introduced a 'twin correction' method to reduce SUVmean variance and account for patient-specific factors.

Main Results:

• Model achieved 89.3% explained variance and 18.0% mean absolute relative error on 177 test studies.
• Twin correction reduced SUVmean variance by up to 90% in organs, mitigating effects of sex, age, fat mass, and uptake time.
• Achieved 49.3% Dice score on the AutoPET dataset for unsupervised tumor segmentation without requiring annotations.

Conclusions:

• The proposed ntPET methodology provides personalized normal references for disease-agnostic PET image analysis.
• This approach enhances quantitative accuracy and facilitates unsupervised anomaly detection in PET/CT.
• Enables more precise and individualized interpretation of whole-body PET/CT scans.