Image reconstruction and elongation artifact reduction for a dual‐panel dedicated prostate PET scanner

TL;DR

This paper introduces new image reconstruction techniques for a prostate-specific PET scanner, showing that deep learning improves image quality and lesion detection.

Contribution

The study introduces a deep learning-enhanced reconstruction method for a prostate-dedicated PET scanner, demonstrating improved image quality and lesion detection.

Findings

Swin-UNETR-based deep learning model achieved highest CNR and CNC for small lesions in simulations.

Experimental results showed Swin method outperformed others in CNR for both large and small prostate lesions.

Model-based and learned methods showed complementary strengths depending on lesion contrast and size.

Abstract

The development of PET scanners dedicated to high temporal and spatial resolution organ‐specific imaging is an active research area, motivated by the need for cost reduction, improved lesion detectability and quantification in specific clinical scenarios, as well as by ongoing hardware and software innovations.

This study investigates and compares various image reconstruction strategies for a dual‐panel prostate‐dedicated PET scanner (ProVision), which features four‐layered dual‐readout time‐of‐flight depth‐of‐interaction detectors and a 22‐position acquisition protocol to improve angular coverage.

A list‐mode MLEM algorithm with multi‐ray modeling was developed and optimized using a scaled NEMA image quality phantom to determine optimal number of rays and iterations. These parameters were then used to reconstruct data from both simulation and experimental acquisitions, including an anthropomorphic pelvis phantom, named Adam‐PETer. Four reconstruction approaches were evaluated: classical MLEM; MLEM with embedded shift‐variant point spread function (PSF) modeling; a hybrid list‐mode reconstruction; and a Swin‐UNETR‐based deep learning model applied as a post‐reconstruction enhancement to MLEM images. Performance was assessed using contrast recovery coefficient (CRC), contrast‐to‐noise ratio (CNR), and contrast‐to‐noise consistency (CNC), on both a scaled NEMA phantom and an experimental anthropomorphic phantom.

In the scaled NEMA phantom simulation, the Swin‐based method yielded the highest CNR and CNC, especially for the smallest spheres, thereby outperforming both standard MLEM and the hybrid algorithm. In the Adam‐PETer experimental prostate phantom, the CNR was 10.43 for MLEM, 14.48 for Hybrid, and 13.85 for Swin for the larger lesion (10 mm). The CNR values were 2.28, 3.03, and 4.35, respectively, for the smaller lesion (8 mm). CNC values also varied across methods, with Swin achieving the best result for the smaller lesion. These findings indicate that model‐based and learned methods offer complementary strengths depending on lesion contrast and size.

The PET scanner‐adapted reconstruction combined with deep learning refinement improves image quality in dedicated, limited‐angle PET systems.

Figures