Testing the Prime95B – a back-illuminated sCMOS camera with 95% QE

A few weeks ago I mentioned that Photometrics had released a new camera, the Prime95B, featuring a back-illuminated sCMOS sensor with 95% peak QE. I got a chance to play with it last week, and I’m pleased to say that it performs as well as you would expect. We compared it to an iXon 888 EMCCD mounted on our CSU-W1 spinning disk confocal. We had purchased this EMCCD for imaging those samples that were too dim to get good images with the Zyla 4.2 sCMOS camera we also have mounted on the confocal. (You can see a sketch of how everything is configured in this previous post). For testing the Prime95B, we replaced the Zyla 4.2 with the Prime95B, allowing us to directly compare the Prime95B and the iXon 888.

Before I get to the data, however, what performance do we expect? To get a sense of what to expect I wrote a Matlab script that calculates the theoretical performance for a number of different cameras, using their quantum efficiency, read noise, and excess noise factor (for EMCCDs). You can get the script here.   You can read more about how to calculate the signal-to-nosie ratio for a camera in this Hamamatsu white paper. Here, I’m ignoring the different pixel sizes of the various cameras by assuming that they all receive the same photon flux per pixel, as if the magnification had been adjusted to produce the same effective pixel size at the sample.

Theoretical performance of different cameras.

Theoretical performance of different cameras. Ideal is a theoretical ideal camera with QE=1 and no read noise. EMCCD assumes a high EM gain, ~200x; 82% QE sCMOS is a Flash4.0v2 or Zyla 4.2 Plus; 72% QE sCMOS is a Flash 4.0 or Zyla 4.2; ICX285 is a Coolsnap HQ2 or similar interline CCD camera.

As can be seen here, the Prime95B is expected to outperform all other cameras at high light levels. In particular, it even should outperform an EMCCD camera for light levels greater than about 2 photons/pixel. This calculation does not take into account fixed pattern noise or dark current. Fixed pattern noise is not so easy to quantify, but my sense is that it is a significant contributor to the noise of some sCMOS cameras. Finally, it’s interesting to see how much better all of these cameras are than the ICX285 cameras, which were state of the art only a few years ago.

Now on to the data. We ran a number of tests on different samples, comparing the Prime 95B to the iXon 888 EMCCD. In all cases, we saw what you would expect from the graph above – it’s very difficult to find a regime where the EMCCD outperforms the Prime95B. Here are some example images from a test slide with our laser launch at 0.1% power and a 50 msec exposure time. For comparison, the white level was chosen to saturate the top 0.01% of pixels in each channel and the black level was set to the mode of the background peak. All the images below are scaled like this.

EMCCD at 0.1% laser power; click to see full size image.

EMCCD at 0.1% laser power; click to see full size image. Download the 16-bit tiff here.

Prime95B

Prime95B at 0.1% laser power; click to see full size image. Download the 16-bit tiff here.

I think the Prime95B image looks better than the EMCCD, but this is a pretty bright image, even with our laser power at 0.1%. To further reduce the sample brightness, I inserted an OD1 filter in the filter turret, so that it attenuated both excitation and emission light, resulting in a 100x attenuation of the signal. This let us reduce the signal to where we were unable to see an image on the EMCCD camera.In the next set of measurements, we also adjusted the exposure times to compensate for the different pixel sizes of the two cameras. The iXon 888 EMCCD has 13 µm pixels; the Prime 95B has 11 µm pixels. The different in area is (13/11)2 = 1.4, so we used a 1.4x longer exposure time for the Prime95B to collect the same number of photons per pixel on each camera.

EMCCD, attenuated signal, 50 ms exposure

EMCCD, attenuated signal, 50 ms exposure. Download the 16-bit tiff here.

Prime95B_70ms

Prime95B, attenuated signal, 70 ms exposure (= 50 ms * 1.4). Download the 16-bit tiff here.

Finally, we did one additional test. The 13 µm pixel size of the EMCCD is not quite Nyquist sampling with the 100x/1.4 lens we used to acquire these images, so we used a 2x magnifier to reduce the effective pixel size to 6.5 µm and then increased the exposure time on the EMCCD so that the photon flux per pixel was equivalent between the EMCCD and the Prime95B. This is probably the most favorable test case for the EMCCD.

EMCCD, 2x magnification, 143 ms exposure time. Click here to download the 16-bit tiff.

EMCCD, attenutated signal, 2x magnification, 143 ms exposure time. Download the 16-bit tiff here.

Prime 95B, 50 ms exposure. Click here to download the 16-bit tiff.

Prime 95B, attenuated signal, 50 ms exposure. Download the 16-bit tiff here.

Here it’s harder to decide which image is better; I would say they’re about equivalent. This was essentially what we found for all our testing, which included real experimental samples in addition to these test slides. The only times we could see the EMCCD outperform the Prime95B was when the image was barely detectable. This is consistent with the theoretical graph shown above, where the crossover between the EMCCD and Prime95B occurs at at SNR just greater than 1.

If you’d like to see more data from our demo, email me. I have a number of other demo images I can send you.

I’m very impressed with this camera. We did not see significant fixed pattern noise, and although the tests presented here are not quantitative, they are consistent with the theoretical expectations that this camera should outperform existing cameras for most signal levels. The improvement in camera performance over the last decade is quite striking, and there is not much more improvement that one can hope for. Manufacturers can continue to beat down the read noise, and perhaps increase QE by a percent or two, but that’s about it. The one thing I would like to see would be a version of this camera with more, smaller pixels for Nyquist sampling on low magnification objectives, and the ability to capture their full space-bandwidth product.

In conclusion, if you’re looking for a high performance camera, the Prime95B is definitely worth checking out. It lives up to the hype!

17 thoughts on “Testing the Prime95B – a back-illuminated sCMOS camera with 95% QE

  1. have you compare the zyla4.2plus with 95B at the same pixel size? I think there will be little difference ,there is only 10% QE difference , the zyla4.2PLUS or FLASH4.0V2 may be better in low light , because the read our noise is lower than 95B, and zyla4.2 and Flash4.0 will be better resolution in low power objective

    • Unfortunately, we don’t have a Zyla4.2Plus or Flash4.0V2 to test. The read noise for the two cameras is pretty similar – 1.4 e- RMS for the Flash4.0V2 vs. 1.5 e- for the Prime95B, so I don’t expect the Flash4.0V2 to outperform the Prime95B in low light, but I can’t be sure without testing. The smaller pixel size of the Flash4.0 / Zyla4.2 is a big benefit for imaging with low power objectives.

      • The shared script used to calculate SNR is said to ignore pixel size difference. But calculating SNR curves taking into account the pixel size of the different cameras changes significantly cross over points. Not doing normalization for pixel size is the same than taking a different photon flux for each camera.

        If you take a photon flux corresponding to:

        – 10 photons per iXon Ultra 888 pixel, ie per 13×13 = 169 um2 area

        The same photon flux gives:

        – 7 photons per Prime 95B pixel: 10*(11×11)/(13×13)
        – 2.5 photons per FI sCMOS pixel: 10*(6.5×6.5)/(13×13)

        It means that, using the graph here, SNR(EMCCD) at 10 photons/per pixel should be compared with SNR(Prime 95B) at 7 photons/pixel. Therefore, the cross over point is found to be at 10 photons instead of 2 photons/pixel. Similarly, the cross over point for EMCCD vs. 82% sCMOS appears to be at about 20 photons rather than 3-4.

        • I think the main reason not to include pixel size is that you can always change the magnification before the camera to get the same effective pixel size on the camera. How easy this is to do depends on the system, but for instance, we have replaced the projection lens on our CSU-22 spinning disk confocal with a Thorlabs achromat and we can achieve any pixel size you want. Alternatively, you can use a 60x / 1.4 NA objective instead of a 100x / 1.4 NA lens to increase the effective pixel size of a small pixel CCD.
          As always, there’s no single right way to do this calculation. You need to understand what your requirements and constraints are and appropriately include those in the calculation.

    • Hi,

      We had the chance to compare the Zyla 4.2 and the prime95B (in fact we lately tested the Zyla4.2 Plus, Flash4 LT and V2, and Prime). The Prime 95B was more sensitive at low light levels even when we binned the Zyla4.2. We reason this to the higher QE and the almost absent fix pattern noise of the Prime95B.

      Mohammed

      • sCMOS binning will increase noise of sCMOS, so you can’t compare Flash or Zyla with 95B by 2 X 2 binning, you can compare it by the same method as compare with EMCCD as this paper mentioned

        • Dear Giant,
          Although the 95B should have a readout noise of 1.5e (rms), the demo unit we got had a Readout noise of 2.5e. The Zyla had a readout noise (@ 12 bit high gain) of 1.4e. Thus when binning the Zyla, the readout noise goes up to 2.8e which is comparable to the readout noise of the 95B. Keep in mind that the effective pixel area of the binned Zyla is 40% larger than the 95B which favors the Zyla in this scenario. Still, the image of the 95B was better.

  2. I wonder about applications like single molecule imaging where EMCCDs are still very important. Do you think the quirky noise characteristics of sCMOS (each pixel has unique gain and has to be calibrated) make it impractical to use the 95B for measuring super dim or small signals even when its apparent sensitivity is comparable to a EMCCD? I guess the corrections used in Huang et al., Nature Methods (2013) could help.

    • We didn’t try any single molecule imaging, although I know Photometrics has demoed the camera with labs that are doing single molecule imaging. My sense is that the fixed pattern noise on the Prime95B is pretty small, but without testing it on single molecules it’s hard to know for sure what the effect will be.

  3. thank you for this valuable comparison.

    however, if you further take into account
    1) v2+ has a slow scan mode with lower read noise:
    2) median read noise is significantly lower than rms for the v2+ -> so for single molecule localization, with proper weighting of the pixels the median read noise might be the proper metric to compare:

    the flash v2+ slow speed read out at 2048×1200 ~ 70 fps. which is comparable to the 95b 1200×1200 82fps at 12bit high gain readout. the flash has nearly twice the area, but the 95b has a bit more speed.

    the snr curves with median noise parameter is as follows: (just modified kurt thorns script)

    https://dl.dropboxusercontent.com/u/56869298/public/effective_scmos_comparison.jpg

    code:
    https://dl.dropboxusercontent.com/u/56869298/public/CameraModeling_fc.m

    • Hi – Quick question – in your calculation, you set the QE values for the cameras as:

      %Flash4.0v2
      QE = 0.82;

      %Prime95B
      QE = 0.96*0.95;

      Why do you multiply the 95% QE of the Prime 95B with 0.96?

      Additionally – with equal magnification – the field of view with the cameras will be different.

      Flash 4 v2 – 6.5 micron pixels
      2048 x 1200 = 13.3mm x 7.8mm =

      Prime 95B – 11 micron pixels
      1200 x 1200 = 13.2mm x 13.2mm

  4. Great post. Do you have an idea of the range of prices for the Prime 95 B ? Will it be more like a EMCCD or a SCMOS camera ?

    Also do you think we could expect larger sensors in the future (say 4-5 MPixels ?)

    Thanks

  5. Dear all,
    I recently tested a Dhyana95 which use the same back illuminated sCMOS sensor of GPIXEL. The readoutnoise was ~3e- rms but in your application with low light flux, could you compare the dark noise ? In fact, the Dhyane dark noise measured ~5e-/s/px @-15°C is very importe compare to the 0.05e-/s/px also measured in ORCA Flash 4 v2.

  6. Currrently comparing these 3, rough specs and quoted prices(April, 2017):

    Zyla 4.2, 82% QE, 53 fps, $14,500
    Ixon 888, 93% QE, 26 fps, $41,000
    Prime 95b, 93% QE, 83 fps, $26,500

    Zyla and Prime have Rolling Shutters
    All just FYI, no judgements

Leave a Reply

Your email address will not be published. Required fields are marked *