# Iterative projected gradient descent for dynamic PET kernel reconstruction

Source: [Iterative projected gradient descent for dynamic PET kernel reconstruction](<https://www.sota2.com/research/paper/6955ff6653fdd5c36173de30>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-01-02T05:00:04.167Z

Authors: Alan Miranda, Steven Staelens

Published: 2025-12-30T16:21:23.000Z

## Abstract

Dynamic positron emission tomography (PET) reconstruction often presents high noise due to the use of short duration frames to describe the kinetics of the radiotracer. Here we introduce a new method to calculate a kernel matrix to be used in the kernel reconstruction for noise reduction in dynamic PET. We first show that the kernel matrix originally calculated using a U-net neural network (DeepKernel) can be calculated more efficiently using projected gradient descent (PGDK), with several orders of magnitude faster calculation time for 3D images. Then, using the PGDK formulation, we developed an iterative method (itePGDK) to calculate the kernel matrix without the need of high quality composite priors, instead using the noisy dynamic PET image for calculation of the kernel matrix. In itePGDK, both the kernel matrix and the high quality reference image are iteratively calculated using PGDK. We performed 2D simulations and real 3D mouse whole body scans to compare itePGDK with DeepKernel and PGDK. Brain parametric maps of cerebral blood flow and non-displaceable binding potential were also calculated in 3D images. Performance in terms of bias-variance tradeoff, mean squared error, and parametric maps standard error, was similar between PGDK and DeepKernel, while itePGDK outperformed these methods in these metrics. Particularly in short duration frames, itePGDK presents less bias and less artifacts in fast kinetics organs uptake compared with DeepKernel. itePGDK eliminates the need to define composite frames in the kernel method, producing images and parametric maps with improved quality compared with deep learning methods.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2512.24322>)

[PDF](<https://arxiv.org/pdf/2512.24322.pdf>)