Self-supervised parametric map estimation for multiplexed PET with a deep image prior.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Multiplexed positron emission tomography (mPET) enables simultaneous multi-tracer imaging.
• Supervised deep learning shows promise for mPET image separation but requires extensive paired data, posing practical challenges.
• Generalizability of supervised methods is limited by patient-specific tracer kinetics outside training distributions.

Purpose of the Study:

• To develop a self-supervised learning framework for mPET image separation using the deep image prior (DIP).
• To integrate a multi-tracer compartmental model within DIP for estimating tracer-specific parametric maps.
• To recover separated dynamic single-tracer activity images from estimated parametric maps.

Main Methods:

• A self-supervised learning framework based on DIP was proposed for mPET image separation.
• The multi-tracer compartmental model was integrated into DIP to estimate parametric maps from dynamic dual-tracer activity images.
• Dynamic dual-tracer images served as training labels, and static dual-tracer images as network input.

Main Results:

• The proposed method demonstrated superior performance compared to conventional voxel-wise multi-tracer compartmental modeling (vMTCM) and a two-step DIP-Dn+vMTCM method.
• Lower bias and standard deviation were observed in separated single-tracer images and estimated parametric maps at both voxel and ROI levels.
• The method was validated on a simulated brain phantom for dynamic dual-tracer ([18F]FDG+[11C]MET) separation and parametric map estimation.

Conclusions:

• The proposed self-supervised DIP framework effectively separates dual-tracer mPET images and estimates parametric maps.
• This approach overcomes the data acquisition limitations of supervised methods.
• The method offers improved accuracy and generalizability for mPET image analysis.