MLPET: a localized neural network approach for probabilistic post-reconstruction PET image analysis using informed...

Area of Science:

• Medical Imaging
• Machine Learning
• Nuclear Medicine

Background:

• Positron Emission Tomography (PET) imaging is challenged by noise and blur, hindering small lesion detection.
• Current methods for improving PET image quality are often computationally intensive.

Purpose of the Study:

• To develop and evaluate MLPET, a fast, localized machine learning approach for PET image analysis.
• To reduce noise and enhance spatial resolution in PET images for improved lesion detectability.

Main Methods:

• MLPET utilizes a probabilistic deconvolution framework, replacing Markov chain Monte Carlo sampling with a neural network.
• The method incorporates scanner-specific point spread functions, correlated noise modeling, and flexible priors.
• Performance was assessed using NEMA phantom data across three different PET systems.

Main Results:

• MLPET achieved higher contrast-recovery coefficients and reduced background noise compared to conventional PET.
• Effective point-spread function (PSF) full width at half maximum (FWHM) decreased from ~2 mm to <1 mm, a 2.5x reduction in blur.
• Comparable image quality was achieved with MLPET in 40-80 seconds versus 900 seconds for conventional PET.

Conclusions:

• MLPET offers a computationally efficient method for quantitative probabilistic PET image analysis.
• The approach successfully combines noise suppression and resolution enhancement using machine learning and informed priors.
• MLPET shows potential for improving small-lesion detectability and quantitative accuracy in clinical PET imaging.