Spectral Tomography for 3D Element Detection and Mineral Analysis

Round 1

Reviewer 1 Report

The paper presents a possible application of Sp-CT to mineral samples. However, only two experiments were show, one being a mixture of grains of pure minerals and the other a drill core sample. In order to better exemplify the power of the novel technique it will be better to compare more full characterized ore samples with the Sp-CT results. 

Some highlights:

1. The paper focus on applications of the Spectral Computed Tomography to mineral characterization;
2. The authors clearly presented the advantages and limitations of this technique, specially when compared with the CT (or Micro CT);
3. The principals of the Sp-CT applied to  mineral characterization were widely covered, including the experimental set up;
4. Little attention was given to the sample preparation, which could be better explored;
5. Only two samples were analyzed in the paper, which is a weakness. However, one sample is an artificial mixture of minerals, giving it a relatively high complexity;
6. The conclusions could be improved.

Author Response

Thank you for your time reviewing our paper.

We added additional information about sample preparation lines 170 and 208.

Indeed only two experiments are presented. They mark the beginning of a new technique that hopefully will stimulate further studies. Note our statement: “ The experiments were designed to provide the framework from where specific applications can be developed, rather than to provide an exhaustive study of the individual applications.”

We have improved the conclusions (from line 454)

Reviewer 2 Report

Dear Dr Godinho- 

Please find my comments on the very exciting manuscript ‘Spectral tomography for 3D element detection and mineral analysis’. Overall, the manuscript is well written and presents significant advance in leveraging all the material information present in a transmitted x-ray, not just the absorption intensity.  I particularly enjoyed the introduction and methods section presenting the full complexity of x-ray-matter interactions limit the information in the greyscale absorption tomogram, and the limitations of other spectrographic approaches for lab-based spectral tomography.  I can see potential for this instrument to improve understanding of rock composition in three -dimensions.  That said there are several points below that need addressing to improve to complete presentation of this new technique.  These are listed in order of appearance not importance.

• Lines 99 -112
  1. Line 102 states that the detector has 128 bins in the energy range from 20 keV to 160 keV, but then states that the data in Fig 1 is for 15 bins. Why is this reduced, and is this how the data in the later experiments was run?
  2. Figure 1- All the figures starting from this one need subfigure labels. It would make the reader work less to figure out what part of a figure they need to be reading. In this figure and it would also help in the text understanding eh difference between the four different spectral curves presented.
• Line 109 the text hints that for the sample examined here soft x-rays (i.e. <20 keV) are not transmitted through the sample. Would thinner samples enable spectroscopic investigation using the detector presented here?
• Line 124 – spell out acronyms the first time they are used. Not all readers will be X-ray tomography experts (and should not have to be if a broadly useful instrument).  As such they will likely not know that FBP means Filtered Backprojection.
• Line 137 – I do not understand how the 7.3 multiplier comes about. My simple calculations using the physical pixel size for the PolyDet and CoreTOM flat panel give difference of 5.3. What am I missing? This should be spelled out.
  1. Also, this raises the question of how tall the pixel array on the PolyDet is. Since really the data and most of the discussion while discussing the SP-CT data presents it as a flat 2D surface, whereas it has a thickness. What is this?
• Lines 143 -199(roughly)
  1. Line 143 – 154 The discussion of different modalities for using the instrument refer to the two experiments that will be presented, before explaining that there will be two experiments presented.
  2. Both experiments described would benefit from a workflow diagram demonstrating how the data gets collected and processed. This is especially true for ‘experiment 2’ but will discuss this more below.
  3. Line 174-176 Why are the SP-radiographs collected in experiment 1 using a lower beam energy than the CT data is collected at? Does this introduce distortions to the data? Also I do not understand what is meant by 10 average spectra and 30 average flat fields, since I thought for Experiment 1, only a single spectral radiographs were being collected.
  4. Similar, for experiment 2, does the 30 flat fields, mean that for the sp-CT experiments there were 30 projections taken, or roughly one line spectra every 12 degrees.
  5. Lines 188 – 199 As this is not an MLA paper, it is even more important that the acronyms and experimental conditions be explained a bit more. For instance, I do not know what GXMAP and SPL-LT mean.  Further, while I have some experience with QEMSCAN I have never used FEI’s MLA, so I do not know what it means to have pixel size of 2 µm with a step size of 6 pixels?  Is this to insure that the spectra recorded are not interacting and minimise volume effects? It is not clear that both MLA datasets are used, since further down the results just say MLA.  If only one data set is being used, only discuss that technique, and do not mention the other results.
• Line 232 again the discussion of specific spectral lines is confusing by lack of subfigure labels in Figure 3.
  1. On a related note, the spectra are too small to read easily if printed out.
  2. Would be helpful to refer to the spectra and particles by their numbers, especially for those who cannot differentiate colours.
• Line 250-251 here it states that elements higher than molybdenum are observable, where as the conclusion (line 420) it says heavier than silver. Which is it?
• Lines 259 – 260 In the caption for Figure 3 it says that the radiograph presented is the spectra presented. Could you please present the radiograph associated with the spectra presented?
• Line 263, what are these ‘Previous mineralogical investigations’? If these are published data, reference it; if it is optical characterisation state it. Otherwise, me and the readers have no idea where these statements are coming from.
• Line 269 – There is no mineral list in the appendix, and there was not supplemental information provided with the manuscript I am reviewing. Please provide this document.
• Line 275-277 – It is not clear how the MLA data is being used in the classification and analysis.
  1. Is it being used to classify in the postprocessing after the CT and sp-CT data have been processed?
  2. Or is it used to guide how the CT and sp-CT data classified?

This is a clear point where a workflow chat showing the flow of information would help clarify how the different parts of the experiment relate to one another.

• Lines 278-283 / Figure 4 – I would recommend, that the spectral curves be labelled as sub figure ‘e’ as again it makes referring to features easier, and easier for the reader to find the information being discussed.
• Lines 310-313 This comment is interesting in that it strongly hints at how supervised machine learning techniques could be used to filter through and classify phases if reference spectra existed.
• Lines 339 This paragraph points out the long term goal for this combined sample characterisation, however, it does little to address that for most geological systems, the soft x-ray region contains much of the compositional information need for classification (ie why quartz is not fingerprinted in the analysis presented). This still shows the need for MLA and other data fusion approaches (for example De Boever et al, Micron, 2015) in order to have a widely applicable instrument. It would be constructive to discuss how to address this blind spot in the systems characterisation.
• Lines 344 though 374, spend lots of space discussing throughput constraints. This is rarely the reason for not doing an experiment.  If the data is critical for a project, the funds to run a long experiment will be found.  Further, this discussion ignores that in several years’ time the detector chip will likely be more sensitive, and potentially also be a square detector instead of a slit detector.  These very foreseeable and likely advances would make much of this discussion moot.  Instead, I would focus on building tight story around the opportunities for software and in particular machine learning tools to guide the different stages of selection through the process.  For example, it would make more sense to envision the use of active learning and gaussian processes to select where the spectra need to be collected (ref Kusne et al Nat Comm 2020).
• Line 396 – 402 This discussion of machine learning applications is unsupported. For applications of training a neural network for segmentation of x-ray absorption data, experience would suggest that one needs more than this number, usually on the order of 50 to 60 slices to capture the full variation in a data set). I would want to see this statement supported with either a reference to results elsewhere or with a demonstration in the supplemental materials.

I look forward to see the revisions to this manuscript describing a much needed advancement in lab-based x-ray technology.

Author Response

Please find our detailed reply attached

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

Dear Dr Godinho- 

Overall, this is a much-improved manuscript and is a great demonstration of the technique.  That said there are a few minor points that should be addressed. There are still a few gaps in documenting the methods that make interpretation of results and conclusions difficult.

1. Line 123 – you state that the number of pixels on the detector is 384, but in your reply letter you give it as 387. The detector has only a fixed number of pixels, which is it 387 or 384?
   1. Additionally, you have not addressed my question regarding the height of the PolyDet sensor. While you give the total width at 307 mm divided into 387 pixels, this does not tell me how tall the single slice is.  From the above two figures I can see that each pixel is 793 µm wide.  Are these square pixels? I ask as this helps me / other readers familiar with tomographic reconstruction, understand how thick a slice each spectral radiograph represents.  Further with out knowing how tall a pixel is, it is very hard to understand the actual size of the spCT voxel, especially as you note the actual focal plane of the spCT and the XCT datasets are not the same, meaning even the geometrical magnifications are not the same.
2. Line 389 - Related to the above I am glad to see that resolution difference between modalities has been measured at 7.3. However, this statement leaves a few hanging concepts, especially as the concept of resolution is non-trivial to measure, and it opens the door on greater scrutiny of what is being done (ie there are great claims at ‘resolution’ that do not hold up when pushed). I think it is the right thing to quantify the difference between the sensors, but this needs to be done carefully.
   1. How was this measured? Is this described I the earlier paper introducing the technology? (Reference 15 in the current manuscript) As noted above, since the magnifications between the two XCT techniques are not exactly the same so how these volumes are compared should be clarified.
   2. Interestingly in your reply you calculate ratio of the number of pixels in a single dimension at 7.38. This is the ratio between a single side of a voxel (ie the pixel width), whereas the difference in volumetric resolution will be a function of the ratio of the volumes. Th this in mind I would suggest that some additional detail in how the voxel resolution was measured.  If simply the smallest feature after a segmentation / phase labelling process than spell this out for the reader.
3. Line 207-226 Again this is a much-improved simple presentation of automated mineralogy, and it really helps to way mark how the technique was used. That said, I am a little confused by the parameters given for SPL- LT as it is supposed to be a higher resolution (spatial or spectral dimension?) technique.  In reading the description, the step size is given as 6 um and 7ms.  This is the same step-size as GXMAP mode.  If this is a higher spatial resolution, I would expect the step size to be smaller, if higher spectral resolution I would expect the dwell time to be longer.  But it is given as the same for both techniques. Please clarify, as I can not see a difference in the operating conditions.
4. Fig 3 / lines 286-287 While I am encouraged that the actual radiographs producing the spectra in Fig 3c, are now included, I am still not clear why they are not presented in this figure. Fig S1 so clearly shows the relationship between the radiograph produced for CT, the spCT radiographs and then when assembled with Fig 3c the final analytic spectra. Whereas the presented sp-radiographs (Fig 3b) are clearly mixtures of particles, and not the single particle data described in the methods for this experiment.  I do not see how Fig 3B presents any data that is discussed. I would strongly recommend that S1 be replaced for Fig 3b with discussion in the manuscript pointing out that two different angles of incidence were needed for this sample to acquire isolated spectra.

Finally, addressing these minor points would ensure a quality manuscript presenting this novel technique and instrumentation. I am encouraged by your interest in active leaning, but my efforts in this direction are a little too premature to be able to collaborate beyond my current network. I look forward to seeing this publication and follow on results demonstrating new insights into the distribution of heavy metal grains.  

Author Response

Overall, this is a much-improved manuscript and is a great demonstration of the technique.  That said there are a few minor points that should be addressed. There are still a few gaps in documenting the methods that make interpretation of results and conclusions difficult.

1. Line 123 – you state that the number of pixels on the detector is 384, but in your reply letter you give it as 387. The detector has only a fixed number of pixels, which is it 387 or 384?
1. Additionally, you have not addressed my question regarding the height of the PolyDet sensor. While you give the total width at 307 mm divided into 387 pixels, this does not tell me how tall the single slice is.  From the above two figures I can see that each pixel is 793 µm wide.  Are these square pixels? I ask as this helps me / other readers familiar with tomographic reconstruction, understand how thick a slice each spectral radiograph represents.  Further with out knowing how tall a pixel is, it is very hard to understand the actual size of the spCT voxel, especially as you note the actual focal plane of the spCT and the XCT datasets are not the same, meaning even the geometrical magnifications are not the same.

Reply: We can see the source of confusion, although the answer relates to more technical details that are described in more details in the referred paper [15]. There are indeed only 384 physical pixels in the detector. Nevertheless, because the detector is composed of 3 sequential modules, there is a physical gap between the modules, thus no information can be measured within that gap. To minimize the artefacts caused by this problem, the gap in the resulting radiograph image is filled with a pixel for which the transmission is calculated by the average transmission physically measured in the neighbour pixels. Therefore, the radiograph image has 3 more pixels than the number of detection units in the detector. Note that this limitation from the detector construction does not affect any of our interpretation of the results.

To avoid confusion we no longer mention the real number of pixels because in practice the number of pixels in the radiograph is what is shown in the paper. We also now clarify that the detector is only one pixel height, which is common to line detectors. We added the real pixel size of the sensor 0.8 mm. (lines 123-126)

1. Line 389 - Related to the above I am glad to see that resolution difference between modalities has been measured at 7.3. However, this statement leaves a few hanging concepts, especially as the concept of resolution is non-trivial to measure, and it opens the door on greater scrutiny of what is being done (ie there are great claims at ‘resolution’ that do not hold up when pushed). I think it is the right thing to quantify the difference between the sensors, but this needs to be done carefully.
1. How was this measured? Is this described I the earlier paper introducing the technology? (Reference 15 in the current manuscript) As noted above, since the magnifications between the two XCT techniques are not exactly the same so how these volumes are compared should be clarified.
2. Interestingly in your reply you calculate ratio of the number of pixels in a single dimension at 7.38. This is the ratio between a single side of a voxel (ie the pixel width), whereas the difference in volumetric resolution will be a function of the ratio of the volumes. Th this in mind I would suggest that some additional detail in how the voxel resolution was measured.  If simply the smallest feature after a segmentation / phase labelling process than spell this out for the reader.

Reply: We agree that the concept of resolution is not trivial and not even uniform for different research groups. Being aware of this, we believe that resolution is not appropriate to compare the 2 imaging modalities in this otherwise qualitative work (spCT is used to identify phases not to quantify them). Instead we use a pixel size factor (7.3) simply measured on cylindrical samples with known diameter.  It is fair to say the spCT has worse resolution than normal CT, although we do not know exactly how much worse, and that would also be very dependent on the sample. Note that we do not claim to have done real resolution measurements as suggested by the reviewer.

Regarding point two. Since the detector is only one pixel height the reconstruction is done using fan beam reconstruction algorithm, thus the result of a spectral scan is a 2D representation in space (plane perpendicular to the rotation axis) plus the 3^rd dimension is energy. Therefore, normal and spectral CT can only be compared on 2D planes and not by volume.

We added a brief description of how the 7.3 factor was measured. (lines 143-144) and corrected that we are measuring pixels instead of voxels in spCT

1. Line 207-226 Again this is a much-improved simple presentation of automated mineralogy, and it really helps to way mark how the technique was used. That said, I am a little confused by the parameters given for SPL- LT as it is supposed to be a higher resolution (spatial or spectral dimension?) technique.  In reading the description, the step size is given as 6 um and 7ms.  This is the same step-size as GXMAP mode.  If this is a higher spatial resolution, I would expect the step size to be smaller, if higher spectral resolution I would expect the dwell time to be longer.  But it is given as the same for both techniques. Please clarify, as I can not see a difference in the operating conditions.

Reply: The higher spatial resolution is given by the backscattered image, which is used to better define the boundaries of the grains. Higher spatial EDS resolution is possible but it was not necessary in this case due to the clear difference between the phases of interest.

We now clarify that the higher resolution was spatial. (i.e. not chemical) (lines 113, 220)

1. Fig 3 / lines 286-287 While I am encouraged that the actual radiographs producing the spectra in Fig 3c, are now included, I am still not clear why they are not presented in this figure. Fig S1 so clearly shows the relationship between the radiograph produced for CT, the spCT radiographs and then when assembled with Fig 3c the final analytic spectra. Whereas the presented sp-radiographs (Fig 3b) are clearly mixtures of particles, and not the single particle data described in the methods for this experiment.  I do not see how Fig 3B presents any data that is discussed. I would strongly recommend that S1 be replaced for Fig 3b with discussion in the manuscript pointing out that two different angles of incidence were needed for this sample to acquire isolated spectra.

Reply: Good idea. We have replaced figure 3b with figure S1 and added the appropriate discussion to the text. This also demonstrates the need to take the radiographs at angles where the particles are not aligned. (Fig.3 and caption was changed and lines 249-255 modified).

Finally, addressing these minor points would ensure a quality manuscript presenting this novel technique and instrumentation. I am encouraged by your interest in active leaning, but my efforts in this direction are a little too premature to be able to collaborate beyond my current network. I look forward to seeing this publication and follow on results demonstrating new insights into the distribution of heavy metal grains.