Dynamic PET image reconstruction utilizing intrinsic data-driven HYPR4D denoising kernel.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Image Reconstruction

Background:

• Dynamic PET imaging generates noisy reconstructed images, limiting the precision of image features.
• Conventional denoising methods often reduce noise at the expense of image accuracy.

Purpose of the Study:

• To develop and validate a novel four-dimensional (4D) denoised image reconstruction framework.
• To achieve 4D noise reduction while preserving spatiotemporal patterns and minimizing denoising-induced errors.

Main Methods:

• A 4D denoising operator/kernel (HYPR4D) based on Highly Constrained BackProjection (HYPR) was developed.
• The HYPR4D kernel utilizes spatiotemporal high-frequency features from a 4D composite to preserve patterns and constrain noise.

Main Results:

• HYPR4D outperformed standard OSEM with filters and HYPRC3D-HTR in 4D noise reduction and spatiotemporal pattern preservation.
• Error in outcome measures was less dependent on region size, contrast, and patterns with HYPR4D.
• For nondisplaceable Binding Potential (BPND), HYPR4D reduced root mean squared error to ~2% in small structures, outperforming HYPRC3D-HTR.

Conclusions:

• The proposed HYPR4D method offers more robust and accurate image features compared to conventional techniques.
• No prior information is required for the HYPR4D method, simplifying its application.