Related Experiment Video
Updated: Dec 25, 2025
12:27
Large-scale Reconstructions and Independent, Unbiased Clustering Based on Morphological Metrics to Classify Neurons in Selective Populations
Published on: February 15, 2017
7.3K
DirectPET: full-size neural network PET reconstruction from sinogram data
William Whiteley1,2, Wing K Luk2, Jens Gregor1
1The University of Tennessee, Department of Electrical Engineering and Computer Science, Knoxville, Tennessee, United States.
Journal of Medical Imaging (Bellingham, Wash.)
|March 25, 2020
Summary
DirectPET, a novel neural network, reconstructs multislice PET images faster than traditional methods. This AI approach shows promise for clinical applications, maintaining image quality even with low-dose data.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Positron Emission Tomography (PET)
Background:
- Direct neural network image reconstruction from measurement data is an emerging research area.
- Previous methods were limited to small, single-slice images.
Purpose of the Study:
- Introduce DirectPET, an efficient neural network for reconstructing multislice PET image volumes from sinograms.
- Address memory challenges in large-scale direct neural network reconstruction.
Main Methods:
- Developed a novel Radon inversion layer to manage memory constraints.
- Compared DirectPET against the ordered subsets expectation maximization (OSEM) algorithm using patient data.
- Evaluated image quality using signal-to-noise ratio, bias, mean absolute error, and structural similarity.
Main Results:
- DirectPET achieved quantitative and qualitative image similarity to OSEM in significantly less time.
- Demonstrated the ability to maintain image quality in low-dose scenarios by training DirectPET to map low-count data to normal-count images.
- Provided lesion analysis using line profiles and full-width half-maximum measurements.
Conclusions:
- DirectPET's efficiency and ability to produce high-quality multislice PET images suggest clinical potential.
- Further research is needed to establish design parameters and performance boundaries for clinical adoption.
