Uncertainty-aware gamma interaction localization and reconstruction in PET | JoVE Visualize

Area of Science:

Medical Imaging
Nuclear Physics
Machine Learning

Background:

Precise gamma-ray interaction localization is crucial for high-resolution positron emission tomography (PET) imaging.
Existing machine learning methods for gamma interaction positioning often neglect event-level uncertainty, limiting their information utility.

Purpose of the Study:

To quantify event-wise positional uncertainties in gamma interaction localization using machine learning.
To demonstrate the utility of uncertainty estimation for improving PET image quality through filtering and weighting strategies.

Main Methods:

Employed multilayer perceptrons (MLPs) and convolutional neural networks (CNNs) for gamma interaction positioning in a semi-monolithic scintillation detector.
Trained regression models using Gaussian negative log-likelihood to estimate coordinates and positional variance, reflecting aleatoric uncertainty.
Integrated predicted variances into time-of-flight ordered-subsets expectation maximization (TOF-OSEM) reconstruction for event-level filtering and line of response (LOR) weighting.

Main Results:

Uncertainty-aware models achieved high positioning accuracy, with CNNs excelling in planar and depth-of-interaction (DOI) estimation.
Variance- and energy-based filtering significantly improved positioning accuracy (MAE down to 0.47 mm) and LOR precision (median distance as low as 0.98 mm).
Image reconstruction demonstrated improved quality, enabling visualization of 0.8 mm rods with a peak-to-valley ratio (PVR) of 1.184, alongside increased signal-to-noise ratio (SNR).

Conclusions:

Event-level uncertainty estimation provides a framework for effective filtering and weighting, enhancing interaction positioning and PET image quality.
The uncertainty-aware approach offers a new reconstruction parameter for improving image quality and supporting quantitative PET imaging, despite a trade-off with noise.
This method advances the field by leveraging previously unused uncertainty information for more reliable PET analyses.