Low dose SPECT image denoising using a generative adversarial network

Greta S. P. Mok; Jingzhang Sun; Qi Zhang

arxiv: 1907.11944 · v1 · pith:U47PDOVCnew · submitted 2019-07-27 · 📡 eess.IV

Low dose SPECT image denoising using a generative adversarial network

Qi Zhang , Jingzhang Sun , Greta S. P. Mok This is my paper

Pith reviewed 2026-05-24 14:36 UTC · model grok-4.3

classification 📡 eess.IV

keywords SPECT imagingimage denoisinggenerative adversarial networklow-dose imagingXCAT phantomnoise reductionmedical image processing

0 comments

The pith

A generative adversarial network trained on simulated SPECT pairs substantially reduces noise in high-noise reconstructions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper applies generative adversarial networks to denoise static SPECT images from low count rate acquisitions. It generates training pairs by simulating projections from XCAT phantoms at two noise levels, one matching standard clinical counts and the other one-eighth that rate, then reconstructs the images. The GAN learns to map the high-noise versions to the low-noise versions using data from nine patients and is tested on a tenth. The central claim is that the processed high-noise images show substantially lower noise while retaining diagnostic quality comparable to the originals, which would allow trading noise reduction for lower injected dose or shorter scan times.

Core claim

The GAN trained on paired high-noise and low-noise SPECT reconstructed images from XCAT phantoms substantially reduces noise in the high-noise images, making their quality similar to the low-noise originals and thereby supporting the possibility of reduced injection dose or acquisition time in clinical SPECT while preserving diagnostic utility.

What carries the argument

Generative adversarial network trained to translate high-noise SPECT reconstructions into low-noise ones using paired data simulated via analytical projection of XCAT phantoms at two count rates.

If this is right

High-noise SPECT images processed by the GAN show substantially reduced noise levels.
The denoised images maintain similar quality to the original low-noise images for clinical diagnosis.
The noise reduction can be traded for a lower radiotracer injection dose.
Acquisition time can be shortened while still yielding diagnostic-quality images.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Real clinical data may contain noise patterns or artifacts absent from the phantom simulations, so direct testing on patient scans is required to confirm performance.
The same paired-simulation training strategy could be applied to dynamic SPECT or to PET if suitable low-dose and standard-dose pairs can be generated.
Combining the GAN output with iterative reconstruction methods might yield further gains in resolution or quantitative accuracy.

Load-bearing premise

The noise statistics and image properties produced by the XCAT phantoms and analytical projector match those of real clinical SPECT acquisitions closely enough for the trained GAN to generalize to patient scans.

What would settle it

Apply the trained GAN to real low-dose patient SPECT acquisitions and check whether the output images match the diagnostic quality and quantitative accuracy of standard-dose images from the same patients.

read the original abstract

The image noise level and resolution of SPECT images are relatively poor attributed to the limited number of detected counts and various physical degradation factors during acquisitions. This study aims to apply and evaluate the use of generative adversarial network (GAN) method in static SPECT image denoising. A population of 4D extended cardiac-torso (XCAT) phantoms were used to simulate 10 male and female patients with different organ sizes and activity uptakes. An analytical projector was applied to simulate 120 projections from right anterior oblique to left posterior oblique positions with two noise levels. The first noise level was based on a standard clinical count rate of 987 MBq injection and 16 min acquisition (low noise) while the other was 1/8 of the previous count rate (high noise). The high noise and low noise SPECT reconstructed images of 9 patients, i.e., 1026 images (9*114 axial slices) respectively, were paired for training. High noise SPECT images of 1 patient were tested using the trained GAN. The noise level is substantially reduced in high noise SPECT reconstructed images after using the GAN. Our method has the potential to decrease the noise level of SPECT images, which could be traded for a reduced injection dose or acquisition time while still maintaining the similar image quality as compared to the original low noise images for clinical diagnosis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This applies a standard GAN to noise reduction in simulated low-count SPECT from XCAT phantoms but stays entirely in simulation with no metrics or real scans.

read the letter

This paper takes an existing GAN architecture and applies it to denoising reconstructed SPECT images from low-count simulations using XCAT phantoms. The core result is that after training on paired high-noise and low-noise images from nine simulated patients, the network reduces noise in a tenth simulated case. The authors suggest this could support lower radiation doses or shorter scans in practice. The work is new in the sense that it targets static SPECT specifically with this count-rate reduction factor and uses a range of phantom variations for training data. Generating 1026 axial slices from the analytical projector at two noise levels provides a controlled paired dataset, which is a practical step for supervised learning in this domain. It handles the simulation setup reasonably by including different patient anatomies and activity distributions. The focus on trading noise reduction for dose or time savings aligns with real clinical needs in nuclear medicine. The evaluation has clear limitations. No quantitative measures appear in the abstract, such as contrast-to-noise ratios, structural similarity indices, or any statistical comparison against the low-noise ground truth or alternative methods. The test is limited to one simulated patient, and there is no validation on actual clinical SPECT acquisitions. Real scans introduce variables like respiratory motion, attenuation map inaccuracies, and scatter that the analytical model omits. This makes the generalization to clinical use an open question rather than a demonstrated outcome. Readers working on machine learning methods for emission tomography might find the data generation approach helpful. Those seeking proven clinical tools will see this as an early feasibility check rather than a ready method. The paper deserves peer review. Referees can evaluate the network details and training procedure once the full text is considered, and they can advise on the necessary next steps for real-data testing. It is not ready for strong claims yet but has enough structure to benefit from external input.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a GAN-based method to denoise high-noise (low-count) static SPECT reconstructions. It generates paired training data from 9 XCAT phantoms using an analytical projector at two count rates (standard clinical vs. 1/8), trains on 1026 axial slices, and reports qualitative noise reduction on one held-out simulated patient, claiming potential to trade noise reduction for reduced injected dose or acquisition time while preserving diagnostic quality.

Significance. If the approach generalizes, it addresses a practical need in nuclear medicine for lower-dose SPECT. The supervised training on paired simulated data is a clear methodological strength that enables direct comparison to ground-truth low-noise images. However, the simulation-only scope limits the assessed clinical significance.

major comments (3)

[Evaluation / Results] Evaluation section / results: All training and testing use only XCAT phantom simulations with an analytical projector; no real patient SPECT acquisitions are included. This leaves the central clinical claim (generalization to actual data with motion, scatter, and collimator effects) untested and load-bearing for the stated potential to reduce dose/time in clinical diagnosis.
[Abstract / Results] Abstract and results: No quantitative metrics (PSNR, SSIM, contrast-to-noise, or reader studies), error bars, or baseline comparisons (e.g., to filtered back-projection or other denoisers) are supplied to support the assertion of 'substantially reduced' noise or 'similar image quality'; evaluation is described only qualitatively on a single simulated test patient.
[Methods] Methods: The manuscript provides no details on the GAN architecture, loss functions, training hyperparameters, or convergence criteria, which are required to assess reproducibility and to determine whether the reported outcome depends on extensive tuning of the free parameters listed in the axiom ledger.

minor comments (2)

[Abstract] The abstract states '10 male and female patients' but then uses 9 for training and 1 for testing; clarify the exact split and whether the 10th patient is distinct.
[Methods] Notation for the two noise levels (987 MBq / 16 min vs. 1/8 count rate) should be defined consistently in the methods to avoid ambiguity when readers attempt to replicate the count-rate scaling.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the constructive feedback. We address each major comment below, indicating planned revisions where appropriate. The work is a simulation-based proof-of-concept using controlled XCAT data to evaluate the GAN approach with known ground truth.

read point-by-point responses

Referee: [Evaluation / Results] All training and testing use only XCAT phantom simulations with an analytical projector; no real patient SPECT acquisitions are included. This leaves the central clinical claim untested.

Authors: We agree this is a substantive limitation. The XCAT phantoms with analytical projection provide a controlled setting to isolate noise reduction effects with paired ground truth, which is a methodological strength for initial validation. However, real acquisitions introduce additional factors (motion, scatter, collimator response) not fully modeled here. We will revise the abstract, introduction, and discussion to explicitly frame the study as a simulation-based demonstration and moderate claims about immediate clinical dose reduction, while noting the need for future real-data validation. revision: partial
Referee: [Abstract / Results] No quantitative metrics (PSNR, SSIM, contrast-to-noise, or reader studies), error bars, or baseline comparisons are supplied; evaluation is only qualitative on a single simulated test patient.

Authors: We acknowledge that quantitative evaluation would strengthen the results. In revision we will compute and report PSNR, SSIM, and contrast-to-noise ratio on the test patient, include standard deviations across slices, and add comparisons against at least one conventional baseline (e.g., Gaussian post-filtering). We note that expanding to multiple test patients or reader studies would require additional simulation or acquisition work beyond the current scope. revision: yes
Referee: [Methods] The manuscript provides no details on the GAN architecture, loss functions, training hyperparameters, or convergence criteria.

Authors: We apologize for this omission, which hinders reproducibility. The revised Methods section will specify the generator and discriminator architectures (including layer counts and filter sizes), the composite loss (adversarial plus pixel-wise term), optimizer settings, learning rate schedule, batch size, number of epochs, and early-stopping or convergence criteria used during training. revision: yes

standing simulated objections not resolved

Validation on real clinical SPECT acquisitions (including motion and scatter) cannot be added without new data collection outside the scope of this revision.

Circularity Check

0 steps flagged

No circularity; standard supervised GAN training on simulated pairs with no derivation reducing to inputs

full rationale

The paper applies a standard conditional GAN to denoise high-noise SPECT reconstructions, training on 1026 paired axial slices from 9 XCAT phantoms (analytical projector, two count rates) and testing on one held-out phantom. No equations, uniqueness theorems, ansatzes, or self-citations are invoked to derive the noise reduction; the output is the empirical result of adversarial training. The central claim is therefore an observed performance on simulated data rather than a quantity forced by construction from fitted parameters or prior self-referential results. This is the most common honest non-finding for applied ML papers.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim depends on the realism of the XCAT-based simulation pipeline and the generalization ability of the trained GAN; these are domain assumptions rather than derived quantities. No free parameters are explicitly named in the abstract, but network training implicitly involves many hyperparameters chosen to fit the denoising task.

free parameters (1)

GAN training hyperparameters
Network architecture, loss weighting, learning rate, and number of epochs are selected to produce the reported denoising effect on the simulated data.

axioms (1)

domain assumption XCAT phantoms with analytical projection accurately reproduce the noise and degradation statistics of clinical SPECT acquisitions
This assumption enables creation of paired high-noise and low-noise training data.

pith-pipeline@v0.9.0 · 5769 in / 1302 out tokens · 26468 ms · 2026-05-24T14:36:40.051423+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

conditional GAN... high noise SPECT images were input to the generator while the discriminator compares the generated samples with the real samples, i.e., the low noise SPECT images
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

XCAT phantom... analytical projector... OS-EM reconstruction

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.