Good DQE in Counting Mode, Poor DQE in Super-Resolution Mode, K2 camera
Forums
I recorded images on a FEI TF20 at 200 kV on a K2 Summit (model XP). In counting mode, the DQE plots are equal to or better than the comparable curves posted at http://grigoriefflab.janelia.org/dqedata (see attached). But, when I tried super-resolution mode, the DQE curves are much worse (see attached). Here is how I recorded the pointer images, and the values I entered into FindDQE:
-Direct alignments updated.
-Although it doesn't, of course, affect the size of the pointer image, magnification was set to a typical magnification we use for recording images.
-Dark and gain references updated for the K2.
-With pointer retracted, dose set to desired value in DigitalMicrograph "Camera Monitor" window. Values were approximately 10, 5, and 3 electrons/pixel/sec in counting mode or approx. 2.5, 1.25, and 0.75 in super-resolution mode.
-Pointer inserted and image recorded and saved with DigitalMicrograph. I just recorded with DM set to regular record mode (no dose fractionation used).
-I recorded counting and SR mode images for the same amount of time, e.g., exposure times for 10 e/p/s counting and "2.5 e/p/s" SR images both were 6.0 sec.
-DM output images were in floating-point format.
-In FindDQE, for counting mode 10 e/p/s, the total dose per pixel was 59.838 electrons. For "2.5 e/p/s", the total dose per physical pixel was 59.928 electrons. (The counting and SR images shown in the attached graphs were recorded on different days.) The counts multiplier was 1.0.
So, if the counting-mode DQE looks as good as or better than the published Grigorieff lab results, why is the SR-mode DQE worse? Did I use the wrong overall dose? The advice in the forum entry "Dose and dose rate for DQE calculation at 200kV" suggests that the dose per pixel would be the same for counting and SR modes. (Please also see my comment below* about the red curve in the SR plot.) As far as I can see (see attached figs), the input and FindDQE-output images for counting mode look very similar to those posted at http://grigoriefflab.janelia.org/dqedata. The SR output image seems to have a gradient or highlight at the edge of the pointer (see attached). In a private conversation with Niko, he said gradients except at the image edge were bad. If this is what he warned me about, how do I get rid of it?
*A note on the red curve in the super-resolution DQE plot. It is the result of a 72.0-sec exposure with the dose at 0.75 e/p/s (measured in SR mode, the dose is 3.0 e/p/s for the physical pixel). Assuming originally that the dose was not based on physical pixels, I entered 54.0 electrons/pixel for the total dose. This is correct if I assume SR pixels and desire an exposure of 50-60 total electrons/pixel. But if I entered the dose based on the physical pixels, it would be 216 electrons/pixel. In a private conversation with Niko, he told me to keep the total dose between 50 and 60 electrons/pixel. Would someone also please comment on the appropriateness of using ranges outside 50-60 electrons/pixel (total dose). What happens to the DQE if I use a total dose less or more than 50-60 electrons/pixel? Won't the image intensities compensate for the higher overall dose?
Thanks, David
Hi David, Have you contacted
Hi David,
Have you contacted Gatan? I wonder if this may be a camera or software issue.
Axel
I looked at the old post
I looked at the old post again (Dose and dose rate for DQE calculation at 200kV) and think that my statement was incorrect. The total exposure per pixel refers to the super-resolution pixel. When you look at the log files posted with the example data (e.g. K2super-300kv-50e-07eps.log) you see that the counts per pixel are close to 50 outside of the pointer area. Basically, FindDQE does not know if the image was collected in super-resolution mode or not (it only detects if the detector was a counting detector). It will deal with all images the same way. You should check that the calculated "Gain conversion factor (counts/e)" is close to 1 for images collected in counting and super-resolution mode when you enter the total exposure per pixel (physical pixel in normal counting mode and super-resolution pixel in super-resolution mode). If it is not close to 1, the exposure might not have been what you think.
Results redone, same overall conclusion--SR worse than C
Axel and Niko,
Thank you for your replies. Based on Niko's reply, I looked at the gain conversion factors. Only the ones taken at 5 e/p/s (counting mode) or 1.25 e/p/s (SR mode) were close to 1 (0.99 or 1.01). The others were 4% to 9% less than or greater than 1. I redid the calculations and entered 1 / (gain conversion factor) as the count multiplier value. This gave new GCFs at 1. The results for counting mode changed very little. The results for super-resolution mode changed significantly because I used different pointer pictures that had a dose of 50-60 electrons/pixel (for SR pixel). (When recalculated, the pictures I used last time showed some DQE points greater than 1. I assume this was because I noted this time that “counting” mode was not detected by FindDQE and the number of outliers was greater than 1.5 million.) Again, however, SR mode does worse than your previous results. Again, in counting mode my K2 XP camera does as good as (10 e/p/s) or a bit better (3 e/p/s) than your older (I think) K2 that was used for your previous results.
Please see the following new files that I posted. The new files are attached to my first posting:
dqe-K2counting-200kV-UofU-2.pdf
dqe-K2superresolution-200kV-UofU-2.pdf
dqe-K2-SRvsCounting-200kV-UofU-2.pdf
counting3-res-small.tif
SR1-res-small.tif
SR3-res-small.tif
I did contact Gatan, as Axel suggested. Gatan said, “When you interpret these curves please keep in mind that counting mode is simply a bin2 version of super-resolution mode. The Counting and Super-resolution mode DQE curves must super-impose almost identically until about 3/4 of the way along the counting mode curve. This could be a good way to check if you are taking and calculating the measurement correctly.” Clearly, this is not the case for my FindDQE results (see dqe-K2-SRvsCounting-200kV-UofU-2.pdf).
I have the following questions:
1. Do you calculate DQE in a manner different from Gatan? They also suggested to me to ask how you deweight based on coincidence loss. Is coincidence loss accounted for in FindDQE?
2. Assuming I measured the dose correctly, did I use FindDQE correctly? Did I apply the count multiplier correctly?
3. In the plot dqe-K2superresolution-200kV-UofU-2.pdf, the SR4a and SR4b curves show vastly different results for two different doses. SR4a was a low dose image (20 e/p) and has a high DQE. SR4a also had a high number of outliers (514337). SR4b was a high dose image (76 e/p), and its DQE declines rapidly beyond 0.3 Nyquist. What is the relationship of dose to the results of FindDQE? Clearly, a low dose gives too many outliers. Will a dose greater than 60 e/p also create problems in FindDQE? In other words, why do you suggest a total dose of 50-60 e/p?
4. You suggested that I check the output image from FindDQE and beware of gradients in that image. I’ve posted counting and SR images. I clearly see a bright “halo” around the edge of the SR images that is missing in the counting images. The halo gives the pointer the appearance of being “underfocused.” Does the halo in the SR results indicate a problem?
Thank you,
David
Log files uploaded
In reply to Results redone, same overall conclusion--SR worse than C by dmbelnap
I uploaded two log files in case they were helpful to answer my questions.
David
The highlighted edges in the
The highlighted edges in the SR output images you posted suggest that the fitting algorithm did not fit the edge successfully. Therefore, the MTF will be wrong and with it the DQE. Not sure why this is happening. The only possible idea would be to increase the exposure so that the pointer image contains less noise. I do not think that there is a problem with too high exposures, just with too low exposures. 216 electrons/pixel should not be a problem.
Higher dose makes DQE and highlights worse
In reply to The highlighted edges in the by niko
I took pointer pictures at a higher overall dose, but the problem persists. Here is what I observed (please look at files counting6*-res-small.tif, SR6b-res-small.tif, SR7*-res-small.tif, dqe-K2counting-200kV-UofU-3.pdf, and dqe-K2-superresolution-200kV-UofU-3.pdf posted with my first message):
-At higher dose, the DQE curve is worse, not better, for both counting and SR modes. Compare a and b curves for SR5, SR6, SR7, counting6, and counting7.
-At higher dose, the highlight around the edges increases. Again this occurs for both counting and SR modes. Compare the output FindDQE images counting6*-res-small.tif and SR7*-res-small.tif.
-After I took the counting5 and SR5 images, I saw that increasing the dose did not improve the DQE curve. I turned the K2 off and let it stay off overnight. I restarted the camera, redid the gain and dark references. About 5 hours after restarting the camera, I took two more sets of pictures (the 6 and 7 data sets). The DQE plots did improve in SR mode for comparable conditions (see SR5b and SR6b & SR3 and SR7a). However, in counting mode, the DQE was worse for comparable conditions (see counting5 and counting6a). I also noted that in the FindDQE output images, I seem to see a gradient in the SR pictures (view SR6b-res-small.tif zoomed out for the most striking example, where the pointer tip is lighter than the rest of the pointer).
So, how does the dose affect the FindDQE outcome? What is the lowest dose I can use to get an effective FindDQE analysis?
Thanks.
David
The exposure should only
In reply to Higher dose makes DQE and highlights worse by dmbelnap
The exposure should only effect the noise in the image and a higher exposure should produce lower noise and more accurate results. It looks to me as if the MTF fit does not work properly because the outline of the pointer is clearly visible. You may have to measure the DQE with more traditional methods, such as a knife edge.
How can the problem be the pointer?
In reply to The exposure should only by niko
How can the problem be the pointer if the silhouette method is not dependent on the shape of the silhouetted object (http://grigoriefflab.janelia.org/node/4236)? Is your assertion that "higher exposure should produce lower noise" theoretical or something you have measured? I only see doses of 50 electrons/pixel in your K2 data set (http://grigoriefflab.janelia.org/dqedata) and mentioned in the paper (http://grigoriefflab.janelia.org/node/4236).
Yes, FindDQE should not
In reply to How can the problem be the pointer? by dmbelnap
Yes, FindDQE should not depend on the shape of the pointer. However, it does not seem to work for your data and therefore it might be better to use an entirely different method, such as the more traditional method using a straight edge.
Data on the effect of dose on DQE?
In reply to Yes, FindDQE should not by niko
Thank you, Niko for your prompt replies.
In the paper, you have clear data on the effect of dose rate on DQE. Do you have any data on the effect of dose on DQE? All I see in the paper and on the website is K2 data recorded at a total dose of 50 electrons/pixel. My data is consistent with your results for dose rate. My data disagrees with your assertion about total dose. I wonder if your assertion is based on theory or observations.
Thanks.
David
Increasing the total exposure
In reply to Data on the effect of dose on DQE? by dmbelnap
Increasing the total exposure should only make the DQE measurements better since there will be less noise in the images you are trying to analyze. I am not sure what the problem is with your super-resolution images but as long as the diagnostic output images of FindDQE show a clear pointer outline it means that the MTF was not correctly fitted by the algorithm. Maybe go through all your parameters again. Make sure that: