[Paper Review] Quantitative Comparison of Planar Coded Aperture Imaging Reconstruction Methods

Tobias Meißner, V. Rozhkov+3 more|arXiv (Cornell University)|Apr 29, 2022

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper presents a quantitative comparison of analytical and deep learning-based reconstruction methods for planar coded aperture imaging (CAI) using experimental gamma-camera data. It demonstrates that a custom-tailored Convolutional Encoder-Decoder (CED) model outperforms traditional methods like MURA Decoding and MLEM in both reconstruction quality (1.37–2.60× better contrast-to-noise ratio) and speed (300 ms), despite using low-fidelity synthetic training data, highlighting the potential of data-driven approaches in nuclear imaging.

ABSTRACT

Imaging distributions of radioactive sources plays a substantial role in nuclear medicine as well as in monitoring nuclear waste and its deposit. Coded Aperture Imaging has been proposed as an alternative to parallel or pinhole collimators, but requires image reconstruction as an extra step. Multiple reconstruction methods with varying run time and computational complexity have been proposed. Yet, no quantitative comparison between the different reconstruction methods has been carried out so far. This paper focuses on a comparison based on three sets of hot-rod phantom images captured with an experimental Gamma-camera consisting of a Tungsten-based MURA mask with a 2mm thick 256x256 pixelated CdTe semiconductor detector coupled to a Timepix readout circuit. Analytical reconstruction methods, MURA Decoding, Wiener Filter and a convolutional Maximum Likelihood Expectation Maximization (MLEM) algorithm were compared to data-driven Convolutional Encoder-Decoder (CED) approaches. The comparison is based on the contrast-to-noise ratio as it has been previously used to assess reconstruction quality. For the given set-up, MURA Decoding, the most commonly used CAI reconstruction method, provides robust reconstructions despite the assumption of a linear model. For single image reconstruction, however, MLEM performed best among analytical reconstruction methods, but took the longest with an average of 13s run time. The fastest reconstruction method is the Wiener Filter with 67ms and mediocre quality. The CED with a specifically tailored training set was able to succeed the most commonly used MURA decoding on average by a factor between 1.37 and 2.60 and a run time of around 300ms.

Motivation & Objective

• To quantitatively evaluate the performance of analytical and data-driven reconstruction methods for planar coded aperture imaging (CAI) in a real-world experimental setup.
• To assess the trade-off between reconstruction quality, computational time, and robustness across multiple methods using a standardized metric: contrast-to-noise ratio (CNR).
• To investigate whether deep learning-based reconstruction, particularly with application-specific training data, can surpass established analytical methods like MURA Decoding and MLEM.
• To provide a publicly available dataset of high-resolution coded aperture images from three hot-rod phantoms for future benchmarking.
• To re-implement and validate a convolution-based MLEM algorithm on experimental data for the first time in this context.

Proposed method

• The study uses experimental data from a gamma-camera with a 256×256 pixelated CdTe detector and a 2 mm thick tungsten-based MURA mask, acquiring images of three hot-rod phantoms at a fixed object-to-camera distance.
• Analytical reconstruction methods include MURA Decoding (linear model), Wiener Filter (frequency-domain deconvolution), and a convolutional MLEm algorithm based on a previously proposed method.
• Data-driven methods employ a Convolutional Encoder-Decoder (CED) neural network architecture trained on synthetic data, with two variants: one using natural photographs and another using source-specific synthetic phantoms.
• Reconstruction quality is evaluated using the contrast-to-noise ratio (CNR), a standard metric in nuclear imaging, computed across all phantoms and reconstruction types.
• The CED is trained on simulated data that neglects near-field effects, transmission, and scattering, creating a domain gap with real experimental data.
• All methods are evaluated on both single-image and dual-image reconstruction scenarios to assess performance under varying photon statistics and exposure times.

Experimental results

Research questions

• RQ1 How do established analytical CAI reconstruction methods (MURA Decoding, Wiener Filter, MLEM) compare in terms of reconstruction quality and computational time on real experimental data?
• RQ2 Can a data-driven deep learning approach (CED) outperform analytical methods in both reconstruction quality and speed when trained on synthetic data?
• RQ3 Does incorporating application-specific a-priori knowledge into the training data of a CED improve reconstruction performance compared to generic training data?
• RQ4 To what extent does the domain gap between synthetic training data and real experimental data affect the performance of data-driven CAI reconstruction?
• RQ5 Can the convolution-based MLEm algorithm be successfully re-implemented and validated on experimental CAI data, and how does it compare to other analytical methods?

Key findings

• MURA Decoding, the most widely used analytical method, provides robust reconstructions despite relying on a linear system model, making it a reliable baseline.
• Among analytical methods, MLEM achieves the highest reconstruction quality but requires an average of 13 seconds per reconstruction, making it impractical for real-time applications.
• The Wiener Filter is the fastest method, with a run time of 67 ms, but delivers only mediocre reconstruction quality, indicating a poor quality-speed trade-off.
• The CED with application-specific training data outperforms MURA Decoding by a factor of 1.37 to 2.60 in contrast-to-noise ratio across all three phantoms, demonstrating superior image quality.
• The CED with tailored training data achieves this improvement in only about 300 ms, significantly outperforming MLEM in speed while surpassing it in quality.
• Even with low-fidelity synthetic training data, the CED method shows strong generalization to real experimental data, suggesting that domain gap mitigation through better simulation could further enhance performance.