ParaPET: non-invasive deep learning method for direct parametric brain PET reconstruction using histoimages

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Machine Learning in Medical Imaging

Background:

• Traditional indirect methods for PET parametric imaging are sensitive to noise, leading to poor parameter estimation.
• Direct methods improve image quality but require significant computational resources and paired training data.
• Current arterial input function estimation relies on invasive blood sampling or additional MRI scans.

Purpose of the Study:

• To develop a novel machine learning approach for reconstructing high-quality brain parametric images from time-of-flight PET histoimages.
• To eliminate the need for invasive arterial sampling, MRI scans, or paired training data.

Main Methods:

• A direct, deep learning-based reconstruction method using histoimages from time-of-flight PET data.
• The method incorporates kinetic and noise models directly into the image reconstruction process.
• No paired training data or arterial input function measurements are required.

Main Results:

• The method achieved strong correlations for kinetic parameters (K1, k2, k3) in simulated phantoms (Pearson correlation coefficients > 0.91).
• Significantly improved contrast-to-noise ratio compared to conventional nonlinear least squares methods (p < 0.05).
• The proposed method demonstrated a 37% increase in speed over conventional techniques.

Conclusions:

• A direct, non-invasive deep learning-based reconstruction method successfully produced high-quality parametric brain maps.
• Histoimages offer a promising avenue for enhancing parametric image estimation in PET.
• The method is applicable to subject-specific dynamic PET data without external inputs.