Direct reconstruction of PET images via tuned CNNs demonstrates shortcomings in common image quality metrics

Peter Lindstrom; Stephen Pistorius

doi:10.1117/12.3005543

1 April 2024 Direct reconstruction of PET images via tuned CNNs demonstrates shortcomings in common image quality metrics

Peter Lindstrom, Stephen Pistorius

Author Affiliations +

Proceedings Volume 12925, Medical Imaging 2024: Physics of Medical Imaging; 129254J (2024) https://doi.org/10.1117/12.3005543
Event: SPIE Medical Imaging, 2024, San Diego, California, United States

Conference Poster

Abstract

While analytic and iterative reconstruction techniques have been studied extensively for several decades, direct reconstruction of medical images using convolutional neural networks has only recently received attention. In direct reconstruction schemes, network architecture plays a profound role, influencing how efficiently the network learns the necessary domain transform and how efficiently it learns and stores prior information. Previous studies have, however, not rigorously tested diverse architectures against each other. In this work, Monte Carlo simulations of realistic positron emission tomography data acquisition were performed, and the data were binned into 2-D sinograms. A flexible architecture was used to generate reconstruction networks whose characteristics depended on 15 hyperparameters. A Bayesian search algorithm was employed to efficiently search the hyperparameter space for the best-performing networks (tuned networks), according to two quality metrics: mean squared error (MSE) and structural similarity index metric (SSIM). A total of 341 networks were trialed. The best-performing networks consistently outperformed reconstruction by maximum likelihood expectation maximization, both in terms of mean SSIM (0.887 vs. 0.855) and MSE (0.711 vs. 0.854), but also on an image-by- image basis, as evidenced by 2D metric histograms. Furthermore, compared to untuned networks, tuned networks used less training data (70k vs. 105k training examples) and required far fewer epochs to converge (6 vs. 150). Compared to metrically inferior ML-EM images, network-reconstructed images suffered from over-smoothing, loss of finer details, and over-regularity of high-contrast regions. Since networks performed well by the metrics, this indicates that MSE and SSIM did not adequately quantify important image features.

(2024) Published by SPIE. Downloading of the abstract is permitted for personal use only.

Citation Download Citation

Peter Lindstrom and Stephen Pistorius "Direct reconstruction of PET images via tuned CNNs demonstrates shortcomings in common image quality metrics", Proc. SPIE 12925, Medical Imaging 2024: Physics of Medical Imaging, 129254J (1 April 2024); https://doi.org/10.1117/12.3005543

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available