Effect of OSEM Reconstruction Iteration Number and Monte Carlo Collimator Modeling on ¹⁶⁶Ho Activity Quantification in SPECT/CT

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

 

The paper investigates the image quality of 166Ho reconstructions generated using the vendor-neutral software package Hybrid Recon (Hermes Medical Solutions). Two reconstruction parameters are varied: the number of OSEM updates (iterations/subsets), and the inclusion or not, of Monte Carlo collimator modelling. First, the optimal number of iterations is determined based on convergence of the recovery coefficients. Subsequently, an additional choice of subsets and the application of MC collimator modelling is investigated. This research is of interest due to the growing adoption of 166Ho as a liver radioembolization device, and the relatively little available research on this specific reconstruction package in relation to 166Ho imaging.

However, I do think that the study is hampered in its design by the limited number of acquisitions (two for each scanner) and could benefit from additional reconstructions, possibly bootstrapped from the original data (from the Hermes website: “Hermia SPECT Reconstruction allows the simulation of SPECT studies with a user-defined percentage of the original counts. The simulation algorithm redistributes the data of the acquired SPECT scan using binomial deviates. The algorithm allows user to choose a fixed or unique seed for the random number generator. A fixed seed always produces the same (random) sample of acquisition data, whereas a unique seed provides unique noise realization.”). I think the following points should be addressed before considering publication.

Introduction

1. “Following this therapeutic activity administration, a post-therapy SPECT/CT scan is conducted to evaluate the doses delivered to both the liver and the tumor.”

Please note the time interval between therapy-injection and post-therapy SPECT.

2. “The OSEM (Ordered Subset Expectation Maximization) method is an iterative approach that involves discretizing continuous images and iteratively solving linear equations. The data is divided into a previously defined number of subsets and analyzed repetitively during N user defined iterations [8]”

Note that this section from ref 8 describes expectation maximization, not OSEM specifically. Further more, in the case of the Hermes software described in this paper, the forward projection step is the result of a monte carlo simulation, and not so much of solving equations. Consider leaving this paragraph out.

3. “Thus, properly determining the CF, which is required to convert SPECT image counts into radioactivity concentration units (MBq/ml), plays a critical role in activity quantification [10,11]”

It should be noted too, that the radioembolization procedure has the unique benefit (compared with systemic therapies), that practically all activity remains in the field of view of the SPECT scan. This allows for immediate conversion to activity without the need for a CF.

4. “The primary goal was to enhance the OSEM reconstruction method for 166Ho SPECT and assess the impact of advanced MC scatter correction through Hybrid Recon™ on both image quality and quantification.”

This seems to contradict the methods section, which stated that all reconstructions were done with the Hermes SPECT recon (which always relies on MC scatter correction), but the impact of MC collimator modelling was investigated.

Methods

5. General remark: The goal of the OSEM algorithm, compared to MLEM, is to decrease reconstruction time for a given number of updates (where updates = iterations * subsets). Note: MLEM is OSEM with just 1 subset. If reconstruction time would be of no importance (for instance if it is very fast, or run at night), then one could just use one subset.

When the number of iterations is doubled and the number of subsets is halved simultaneously, the number of updates is the same so the image quality will be very similar (but the computation time will be longer).

However, when changing from 15 subsets to 8 subsets while keeping the number of iterations constant, the number of updates nearly halves, having roughly the same effect as keeping the subsets constant at 15 and decreasing the iterations from 5 to 3.

Therefore, it is no surprise that if the highest CRC is found at 5it15subs (75 updates), it decreases again when reducing to 5it/8sub (40 updates).

Please consider removing the analysis on subsets completely, or alternatively, repeat the analysis with 8 subsets and vary the number of iterations from 1 to 14.

SPECT/CT imaging protocol

6. “The objective of the study is to obtain results that can be applicable to any system, regardless of the model or manufacturer. Thus, it does not correspond to a multi-center study but rather to an independent verification in two different systems/phantoms/acquisition parameters combinations.”

This is a big statement considering only two systems (with identical collimator/detector) were investigated. How do the results of this study extend to other systems?

Note: I would consider this paragraph to be in the introduction or discussion rather than materials and methods.

Phantom acquisition

7. A relatively large sphere-to-background ratio was applied (30:1 and 60:1), whereas typical values are 4:1 or 8:1. Often, the choice of the sphere-to-background ratio arises from typical tumor-to-normal activities found in patients. Can you elaborate on what the rationale was behind the choice for 30:1 or 60:1?

Image Quality

8. Equation 2 for the coefficient of variation seems inconsistent with the sentence above it, which states that “The CV value was obtained by averaging the calculated values from each slice”.

9. Consider including a contrast-to-noise calculation (i.e.: (Ch-Cb)/CV). In the Discussion section it is mentioned that: “Consequently, it was determined that 5 iterations achieved an optimal balance between noise reduction and accuracy. As the number of iterations was increased to 6 and 7, it was observed that the contrast coefficients remained relatively stable for both phantoms. Therefore, additional iterations would not improve the reconstructions adequacy or reliability but would merely lead to elevated noise levels”. I think that a CNR calculation could be a worthwhile addition.

Activity quantification

10. The CF was only calculated for one reconstruction (i.e., 5it/15sub full MC). It would be interesting to see a graph of the CV as a function of iteration number and the influence of collimator modelling.

11. Eq 4: add index i to ARC

12. It seems to me that the two methods for ARC determination following Eq 4 and 5 are trivially related to each other. The CF calculated from Eq3 is either used directly in Eq4 to determine the ARC from the count density, or indirectly in Eq5 through the Hermes reconstruction software (which also applies the CF as a simple scaling factor).

Consider removing one of these ARC methods.

13. Please explicitly state the method for determining the last ARC method (based on the EANM guidelines).

Results

14. General comment: the results section contains many paragraphs which seem to belong in the methods section or in the discussion section. Please thoroughly revisit this section.

15. In Figure 3 a typical scatter plot with connected dots is used (which indicates a continuous x-axis), but in Figure 5 a bar chart is used, which is typical for a categorical x-axis. However, the x-axis in Fig.5 (and some other figures) is sphere size, which would logically be plotted on a standard scatter plot with connected dots, such as Fig.3. Please change the figure type from bar chart to normal graphs when plotting as a function of sphere size (preferably sphere diameter, but otherwise sphere volume such as Fig. 9)

16. In Fig. 7 and 8, consider using the same range for the y-scale in all figures for comparability (i.e, 0 – 0.8)

17. From the manuscript, it is not clear to me how the statistical analysis is performed. With my limited expertise in statistics, this sample size does not lend itself to do any meaningful statistical analysis other than denoting which of the methods performs best (note, the results are highly correlated as they all share the same underlying imaging data). A possible way for improvement could be to repeat the reconstructions multiple times using different random seeds to initialize the MC algorithm.

18. A new fit function was introduced in the results section (rather than methods). What is the rationale for this particular function other than that it gave a better fit? Note that with 4 variables to fit 6 datapoints, many functions would fit very well, including a simple polynomial.

19. As the authors mention in the discussion, the original fit function performed best on the ARCs as produced with the method in the EANM guidelines, for which it was introduced.

20. Especially in the context of radioebolization dosimetry, concerning patient safety, the introduction of a predictive function to compensate for PVE should be more rigorously substantiated. Please consider removing this function or explain the underlying model.

21. It would be interesting to see a side-by-side image of reconstructions with and without the MC collimator modelling.

Discussion

22. “A limitation of this study was the consideration of a low number of subsets, preventing a conclusion on whether 5 iterations remain optimal with a varied subset count.”

Why was there any limitation in the number of iterations/subsets?

23. In light of point 20 above, please remove or reconsider the discussion about the performance of the fit routines. Extrapolating to larger volumes makes sense only when the underlying model can be relied upon.

Conclusion

24. In “OSEM reconstruction, utilizing 5 iterations and 15 subsets and enhanced through full MC collimator modeling, significantly improved 166Ho SPECT/CT image quality compared to other examined reconstruction methods”, please be more specific about the ‘other examined reconstruction methods’.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors tested different reconstructions with the aim of optimising image quality and quantification in 166Ho SPECT/CT imaging. However, the study design has some deficiencies and needs improvement:

·         I would request the authors to include the reconstructed PET images in the manuscript, presented in a panel showing the differences across reconstructions. It is always helpful to include the images from which data are extracted in order to provide context and illustrate the process. I believe it should be a standard requirement for any paper about imaging.

·         Furthermore, some of the trends identified by the authors are already well-established, for instance, that as the number of iterations increases, the level of noise, expressed as CV, also rises. I would therefore propose the inclusion of a qualitative evaluation by expert opinion or a detectability criterion, such as the Rose model. In the present manuscript we are not able to know if some spheres are undetectable in the PET images, as the PET images are not included.  While it is possible to draw ROIs over the CT image, it may not be meaningful to quantify non-visible spheres. If this is the case, please, reconsider and redo statistics using only detectable spheres.

·         Please justify the selected activity concentration within the phantom. I believe that the sphere-to-background ratios of 30:1 and 60:1 are not realistic.

·         Why did the authors set the number of iterations to 5 to vary the number of subsets? This is not justified at all. Also, only two different numbers of subsets were used. Why is this? In general, in OSEM algorithms, the final image quality depends on the effective number of iterations, which is the product of iterations and subsets. Therefore, it is not logical to play with both. It would be more beneficial to test different filters to optimise the results. In this design study, which is not comprehensive with all possibilities, it is questionable whether the "5 iterations/15 subsets" is the best option.

·         One of the main contributions of this paper is the evaluation of collimator modelling within the OSEM algorithm. Therefore, I would suggest to choose the best reconstruction parameters (iterations, subsets and filtering) and then, in a subsequent step, to study the impact of this modelling independently.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

The Authors have addressed all of my concerns with the original manuscript. The revised

manuscript is ready for publication.

I would only suggest a minor change. Please, crop the images in figures 4 and 7 to make them look bigger in the final presentation and present a single color scale. Here I include an example:

 

Author Response

Comments 1: 

The Authors have addressed all of my concerns with the original manuscript. The revised

manuscript is ready for publication.

I would only suggest a minor change. Please, crop the images in figures 4 and 7 to make them look bigger in the final presentation and present a single color scale. Here I include an example:

 

Response 1:

Thank you for your feedback and for taking the time to review our manuscript.

We followed your suggestion. We have cropped the images in figures 4 and 7 to make them more perceptible and have standardized them with a single color scale for consistency and clarity.