Cameras and Detectors II: Specifications and Performance • iBiology

01:00:11.15 In the second part about detectors, I'll be talking
01:00:16.28 exclusively about cameras. And I will now be talking about their
01:00:21.09 specifications and their performance. So basically this will
01:00:25.16 be very useful if you decide to buy a camera, what are you going to
01:00:29.29 pay attention to. But also when you already have a camera, it is
01:00:33.28 highly useful to understand how it works and when
01:00:38.27 your camera is actually limiting your signal noise ratio,
01:00:42.28 i.e. the quality of your experiments, or when there's
01:00:46.11 something else going on. So camera specifications
01:00:51.16 are often, and always should be, provided by the
01:00:55.28 manufacturer. And so I've pulled down here the specifications
01:01:00.29 for one of these scientific grade cameras, not an endorsement
01:01:06.11 for this specific one at all, but just to have an example.
01:01:09.01 And so at the end of the talk, we'll basically have gone
01:01:13.01 through everything that we see here on this spec sheet.
01:01:15.24 Also, as you will see, there are two basic types of
01:01:22.03 chips that are being used at the moment in scientificimaging. One is the so-called interline chip, as we discussed
01:01:31.02 in the first detector lecture. There's like a classic
01:01:35.04 Sony chip, the ICX285, that you will see in many, many
01:01:39.06 cameras. The other is this frame transfer chip that has
01:01:44.13 EM gain, and I'll explain that in a little bit. An this is
01:01:48.00 a chip made by e2V, the CCD97, that is in many cameras
01:01:53.20 that we like a lot. So first of all, pixel size and number.
01:02:00.17 So, of course you know, you want a lot of pixels and
01:02:05.27 in the end, you also want them to be a certain size.
01:02:08.29 But what is the right size for your microscope system?
01:02:14.12 Now this Sony Interline ICX285 -- although this particular
01:02:20.21 camera says it's not one of those, its specs are highly,
01:02:24.10 highly similar. It has 1040x1392 photosensitive elements, pixels.
01:02:32.11 And they are each 6.45 microns on a side, so 6.45 microns
01:02:39.22 square. So is that good or is that bad? So one thing to
01:02:46.13 keep in mind is when you project an image through your lens
01:02:52.02 onto the camera, there's often no additional magnification
01:02:56.00 there. And so that means you can now back calculate
01:02:59.29 what a 6.45 micron pixel corresponds to in your sample
01:03:05.14 plane. So if you had 100X objective, that means that you have
01:03:08.25 65 nanometers, 6.45 microns divided by 100, 65 nanometers
01:03:16.10 per pixel. So is that okay? Now, what we all kind of intuitively
01:03:27.06 know is that if you have few pixels, that the image looks
01:03:33.06 very blocky, and that is bad. And the more pixels you get,
01:03:37.02 the less blockier and better looking it becomes. But is
01:03:42.07 there some kind of limit to this? Should we just be
01:03:45.14 throwing more and more, smaller and smaller pixels
01:03:48.23 at it? Or is there a certain optimum? Well the answer to that
01:03:54.03 comes from a very old sampling theorem. And so the question
01:04:02.28 is, how many pixels do we need to reproduce the smallest
01:04:07.22 thing that we can discriminate in our microscope? And
01:04:12.15 this Nyquist-Shannon sampling theorem that actually came out of
01:04:18.11 an old signal processing work on much more linear signals,
01:04:24.17 but it states that to sample an analog signal, you need at least
01:04:32.26 two pixels per resolvable element. So what we tend to tell people is
01:04:39.24 that we really like a little bit more than that, so between
01:04:43.25 2-3 pixels. And that way, there's a very intuitive way of
01:04:48.19 thinking about it, that is if we have two points that are just
01:04:56.00 resolvable within our microscope, if we then digitally sample
01:05:00.09 those two points, we will need at least 3 pixels to
01:05:04.23 see them. So to have 2 bright pixels to represent
01:05:09.22 the points themselves, and then one in between so that we can
01:05:13.26 actually see that there is a dip and that these two things are separated.
01:05:17.01 So, two to three pixels per resolvable element is what we want.
01:05:22.11 So that then brings us to a resolution centered view of
01:05:29.15 pixel size and how to match your pixel size to your camera.
01:05:33.05 So the first thing that we want to know is, what is the
01:05:36.02 actual resolution of the microscope itself? And so, you actually
01:05:40.08 have learned in other lectures is that resolution is a function
01:05:43.22 of the numerical aperture of the objective and it's a function
01:05:47.28 of the wavelength of the light. And so, if you have an
01:05:53.23 objective that has a numerical aperture of 1.4, so a nice
01:05:57.20 oil immersion lens, you have green light, you're imaging GFP
01:06:01.21 or something, that gives you a resolution of roughly 220nm.
01:06:07.21 Now, if you want two pixels per 220 nm, as the Nyquist-Shannon
01:06:13.26 theorem states, that means that you need 110nm per pixel.
01:06:19.15 So, instead of choosing a camera with a different pixel size,
01:06:26.26 you can now choose an objective that has a certain magnification
01:06:31.04 so that you end up with 110nm per pixel. So we can simply
01:06:37.05 calculate it for the 6.45um pixels, so we need 110nm
01:06:46.27 in the end, and that means we need a magnification of 58.6.
01:06:51.04 Or roughly a 60X magnification. So therefore, a 60X
01:06:58.09 lens would be fine with this. With a 40X lens, we would be undersampling,
01:07:03.17 we would be looking at a larger field, but we would not be
01:07:06.23 getting the full resolution out of it. And 100X lens would be
01:07:10.29 oversampling. But also note, 100X lens would give you
01:07:15.14 about 3 pixels per resolvable element. And we said previously
01:07:20.17 we kind of like that, too. So the 6.45um square pixels
01:07:25.18 really matches our 60X and 100X oil immersion
01:07:30.07 lenses very well. And that is one of the reasons why this particular chip
01:07:34.25 has become so popular in microscopy. Another important
01:07:41.06 aspect of cameras is how well they transform incoming photons
01:07:46.22 into charge that then can be measured out. And that is
01:07:51.07 called the quantum efficiency. So a maximum quantum
01:07:55.07 efficiency would be 100%, every photon in is being
01:07:59.14 transferred into an electron that we can then measure. You see here
01:08:03.13 two curves. One looks pretty good, goes up to 70% or so.
01:08:08.26 The other one looks less good, goes up to 40% or so.
01:08:12.25 Both of these are not as good as what we get in real life.
01:08:17.04 We can get from front illuminated CCDs something up to
01:08:21.09 70% in the green. And for these back illuminated CCDs,
01:08:27.16 we can actually go up to 95% or so. So pretty amazing
01:08:32.08 that almost every photon that hits it will be detected.
01:08:35.24 So what is the difference now between this back illuminated and
01:08:39.15 front illuminated? Well front illuminated means that we come from
01:08:43.20 the top of the chip, and we go through these transparent electrons
01:08:47.23 -- all the other stuff that the chip manufacturer need to
01:08:51.25 put there to make the chip actually work, all of that
01:08:57.07 absorbs a tiny little bit of the light and that reduces the
01:09:01.18 quantum efficiency. So the trick in a back illuminated CCD
01:09:06.02 is to put the chip upside down and then to carve out that
01:09:10.29 back, that backing of the chip, so that the photons hit
01:09:15.07 immediately the photosensitive material. So you can
01:09:19.22 already see that intuitively and understand that to do
01:09:24.16 the backing and making a chip very, very thin, that that's a
01:09:29.21 very difficult process. And a lot of them will actually break
01:09:33.05 or get lost in the process. And so that makes these backside
01:09:36.22 CCDs substantially more expensive than front illuminated
01:09:41.17 ones. But they bring us that higher quantum efficiency
01:09:45.06 that we very much like. Okay, I'll now talk about different
01:09:53.28 types of noise. Because in the end, what we care about is
01:09:58.02 our signal over noise. So understanding noise is important
01:10:03.16 in making that ratio, that signal noise ratio, as good as
01:10:08.19 possible. So we all intuitively understand that when you decrease
01:10:13.01 your exposure time, your image gets grainier and
01:10:17.05 grainier, worse and worse. So you get more and more
01:10:20.24 noise and less signal. But why is that really the case?
01:10:26.14 Why does your signal noise ratio decrease with decreasing
01:10:32.04 exposure time? There are 4 major types of noise that are
01:10:38.23 important to understanding our cameras. And the very first
01:10:42.08 one is the most fundamental one, and that is this photon
01:10:46.18 shot noise. And basically, this is due to the fact that light
01:10:51.16 is not only a wave but also a particle. And so to measure
01:10:57.07 light, you have to imagine that there's this continuous bombardment
01:11:00.26 happening with particles. And if you do a measurement
01:11:05.08 during a given time period, that measurement is never
01:11:10.07 going to be exact for the rate with which these photons hit
01:11:14.03 you. And that is simply given by nature, that there's no
01:11:20.05 continuous stream of photons. Photons arrive with a certain
01:11:27.28 irregularity. And that's what we call photon shot noise.
01:11:33.11 So it's given by nature, our camera cannot help it.
01:11:38.01 All we know is that it scales with the square root of the
01:11:42.26 number of photons. So already here you see that if I
01:11:46.02 get more photons, since this only goes with the square root,
01:11:50.26 I'll improve my signal noise ratio. Then another important type
01:11:58.02 of noise, and this is coming from our camera, is the
01:12:03.07 read noise. So, remember in that first lecture, we had a
01:12:07.22 little measuring cup that was measuring how much water
01:12:11.05 was in these buckets. Well, here in the CCD, we're also
01:12:15.00 measuring. And there's simply an error in measuring
01:12:19.11 how much charge is in each of these pixels. And you can also
01:12:23.22 understand that the faster you start to read out, the
01:12:28.01 larger that error becomes. So this error is independent of
01:12:32.03 the number of photons hitting the chip. It's for every pixel
01:12:35.20 there's a certain error added on by our CCD camera.
01:12:39.12 And so this is a number that you can measure, and our camera
01:12:43.27 manufacturer will tell us. And you will see here that
01:12:47.28 it gives us the read noise in electrons, photoelectrons,
01:12:53.16 and it also tells us that at a lower speed, we have a lower
01:12:58.03 read noise than at a higher speed. So this is a property
01:13:05.12 of the camera and the speed with which you read
01:13:08.03 out the camera. Another type of noise that we often don't think
01:13:13.25 about too much is fixed pattern noise. So if we illuminate
01:13:16.17 the chip with a very homogeneous light, that illuminates
01:13:21.29 everything evenly, you will see that not all the pixels
01:13:25.22 in that chip respond exactly the same. Some of them will
01:13:29.05 be a bit dimmer, others will be a bit brighter. There
01:13:32.15 will be patterns depending on how the chip was manufactured.
01:13:36.15 So this fixed pattern noise scales linearly with the signal,
01:13:41.03 and what you also can do is take a control image,
01:13:44.05 a flat field image where you do this illumination of the whole chip
01:13:50.05 in one go. And then use that to correct the image that you
01:13:55.19 measured. So you take out this fixed pattern noise.
01:13:58.17 Another type of noise is dark current. So the chip, even when
01:14:07.00 there's no light hitting it, will accumulate a tiny bit of charge.
01:14:11.00 And the longer you expose, and also the warmer the camera
01:14:16.03 itself is, the more of that charge will show up. So by cooling
01:14:21.22 the camera, the manufacturers reduce this dark current
01:14:26.00 dramatically. And you will see here, for this specific chip,
01:14:29.29 we have a dark current of something like 0.0005 electrons
01:14:36.28 per second, and that means you would need to expose
01:14:41.19 2000 seconds before, on average, you would get one
01:14:45.25 electron darker per pixel. And so in practice, our exposure
01:14:51.12 times are way lower. And that means that in normal
01:14:55.02 use, this dark current is not really an issue. So that leaves us
01:14:59.23 with the two main important noise sources, the photon
01:15:04.16 shot noise and the read out noise. Now if we put that
01:15:10.21 semi-mathematically, so what we're interested in is
01:15:15.13 measuring the signal noise ratio. So we know that our signalscales
01:15:21.18 with the number of photons. So for the noise, we have a
01:15:25.24 contribution of the read noise and a contribution from the
01:15:29.15 photon shot noise, the square root of the number of photons.
01:15:33.16 And we can express that then as the square root of the
01:15:37.10 read noise squared plus the number of photons.
01:15:40.27 And from this simple equation, you can see that there are
01:15:45.01 basically two different regiments. So when we have a lot of light
01:15:50.05 or enough light so that we get out of that read noise
01:15:53.20 squared, when we're higher than the read noise squared,
01:15:57.13 then we're in shot noise limited. Whereas when the number of
01:16:02.09 photons is smaller than the read noise squared, then we're
01:16:06.00 limited by our read noise. And those are two different regiments
01:16:10.15 because in this case, we really care about the camera,
01:16:13.26 whereas here we really don't. Now this also, once you have
01:16:21.13 a shot noise limited image -- so the camera is really fine, it's just
01:16:26.14 nature that limits our noise. So what can you do to
01:16:31.04 improve the signal noise ratio? Well, if you want to double your
01:16:36.18 signal noise ratio, you'll need to increase your number of photons
01:16:41.20 four times, so that square root of four is two, so you double the
01:16:47.00 signal noise ratio. So you can do that by exposing four times
01:16:52.01 as long, and this is why longer exposures look better.
01:16:55.14 Alternative strategy is to use 2x2 binning. So what is
01:17:02.13 binning? In binning, we're now reading out multiple
01:17:08.03 pixels at the same time. So we're using these pixel
01:17:13.24 shifting technologies of the CCD to shift twice into this
01:17:19.04 serial readout register. So that you accumulate the
01:17:23.02 charge of these two pixels into this one here. And in that
01:17:27.23 serial readout register, we do the same trick again. We shift
01:17:31.14 twice into the readout amplifier. So that in the end, the readout
01:17:36.16 amplifier now has a charge of these four pixels here
01:17:40.01 and reads that out in one go. And that means that we now
01:17:45.14 have the charge of four pixels, we have only one
01:17:49.19 read out noise contribution. So we increase the signal noise ratio
01:17:54.27 by doing this two fold. And of course, we also are faster
01:18:01.07 because the readout takes most of the time, so this means
01:18:05.20 that we lower the total read out time by 2 or 4 fold. But
01:18:09.12 of course, we're giving up spatial resolution, so we now have like
01:18:14.01 instead of a 6.5 micron pixel, we have kind of a 30 micron
01:18:18.16 virtual pixel. Okay, so back to read out noise and photon
01:18:26.05 shot noise. Now, what we saw is that when the photon
01:18:31.27 rate is somewhere on the order of the square of the read out noise,
01:18:40.13 that is where the read out noise becomes important. And regretfully,
01:18:44.20 since we don't want to use much light in our microscopes,
01:18:49.05 and we don't want to kill our live samples, sometimes we just don't have
01:18:54.08 much light coming out of it, we are often working in this regimen
01:18:58.19 where we're limited by read noise. So this example shows
01:19:03.07 a sample that has on average a photon flux of 10 photons
01:19:09.20 per pixel. Now, when we add to that the photon shot noise,
01:19:15.21 you already see that the image becomes a lot grainier,
01:19:20.25 but on the other hand, this is the best that you can do in nature.
01:19:25.06 This is just a simulation, this would be the real image
01:19:29.19 that you get. Now, if we have a camera with 5 electron
01:19:34.07 read noise, and that is actually pretty good, it becomes more or less
01:19:38.23 unusable. So, what you should remember here is that
01:19:44.13 we have here 10 photons per pixel on average, square root of
01:19:49.13 10 is roughly 3, so we have here the photon shot noise
01:19:54.27 is about 60% of the read out noise. And that is really
01:19:59.05 you know, an area that you don't want to be. So
01:20:02.14 what you'd like to do is reduce the read out noise of the camera.
01:20:06.22 Now there are multiple technologies that accomplish that
01:20:11.29 to an extent. And one of them that has become really popular is
01:20:16.08 this electron multiplication process. So what happens here
01:20:20.28 is that we amplify the signal on the chip before it goes
01:20:26.20 to the readout amplifier. So these chips have a special
01:20:30.12 register, they have a special electron multiplication register.
01:20:36.07 And in that register, the charge is being clocked into the next
01:20:41.03 well with a very high voltage, much higher than normal.
01:20:44.28 And just as in this average photodiode that I talked about
01:20:49.19 in the first lecture, we now get an amplification of the number
01:20:54.04 of electrons. So on average, with every pixel that you move,
01:20:59.25 you get like, depending on the voltage you put over them,
01:21:04.03 you get like up to 2% more electrons. But if you now do this
01:21:10.13 512 times, you get really an enormous amplification of
01:21:15.08 the signal. So now what the net effect is here is that we amplify
01:21:21.14 these one or few electrons before it goes through the amplifier.
01:21:27.00 so the amplifier sees like, let's say 100 to 1000 electrons,
01:21:31.12 and then adds the read out noise, but since the signal is much higher than the
01:21:36.26 read out noise, we don't really care anymore. So to summarize,
01:21:40.17 the EMCCDs have a fast, relatively noisy CCD, so they
01:21:45.23 -- you know, often the chips at these fast speeds can have pretty high
01:21:50.15 read noise itself. But since we amplify the signal, by 100 fold,
01:21:57.00 that 50 electron read noise now looks like 0.5 electron.
01:22:01.06 Well that's great! So the downside of this is that the amplification
01:22:07.16 process itself is also stochastic. And when we go from 1
01:22:12.28 well to the other, where we have the high voltage and
01:22:17.24 impact ionization, that is a stochastic process as well.
01:22:22.11 And that kind of looks the same as that Poisson shot noise,
01:22:26.17 and so some people think about EM gain as if you are
01:22:32.22 halving the quantum efficiency of your chip. You know, the real good
01:22:38.08 thing is that you get your image really fast and you don't have
01:22:42.04 to worry about read out noise anymore. Now this used to be
01:22:45.29 the major game in town for a couple of years. But very recently,
01:22:50.14 the camera manufacturers realized that they could actually
01:22:53.28 use the CMOS architecture. So this is the photoelectric
01:23:00.27 cells that have a few transistors built around them, so when they
01:23:05.03 transfer the charge into a voltage right there are the pixel,
01:23:08.21 these things can run very, very fast. And it so turned out
01:23:13.20 that they can actually make them such that they have
01:23:16.22 very low read noise. So the scientific CMOS cameras
01:23:21.29 claim, some of them claim, to have a read noise that
01:23:26.16 is less than 1.5 electrons. And that is really in a very nice
01:23:31.07 area, where this basically becomes so low that you don't care about it.
01:23:38.08 So the current architectures look like these cameras have
01:23:43.24 often something like 2000x2000 pixels, that's 6.5 microns.
01:23:48.16 So it's huge. It's really like 3-4 times larger than the
01:23:53.11 interline chip that I've been talking about. They also can
01:23:57.19 run extremely fast. So out of the box, they can go 100 frames
01:24:04.24 per second, at full chip. And you can even take subregions
01:24:09.17 of those cameras up to like 25000 frames per second.
01:24:13.17 So that opens up a whole new realm of imaging. They do
01:24:18.25 have quite a bit of fixed pattern noise, but that's something
01:24:21.21 you can deal with. These cameras tend to have a lot of
01:24:24.22 logic onboard, so that a lot of these corrections are already
01:24:28.25 being taken care of inside the camera. And one thing to
01:24:32.28 remember is when you use a CMOS camera, since you now
01:24:36.03 have the analog to voltage conversion right at the
01:24:40.01 pixel itself, so binning does not -- as in a CCD, reduce
01:24:46.05 the read out noise. Because each of these pixels causes its own
01:24:49.26 read out noise. And another interesting feature of these
01:24:54.28 CMOS cameras is that they don't read out the whole chip in one
01:24:59.19 go. The read out, the start of the exposure and the read out
01:25:04.06 often goes in kind of a wave. And this is called the rolling shutter.
01:25:09.04 Some of them also can operate in this global shutter
01:25:13.07 mode, where you expose in one go. But then they have a
01:25:17.26 higher read out noise again. But this reading out in a wave
01:25:21.17 is something, this rolling shutter, is something to be aware of.
01:25:26.03 Since our specimen often don't move very fast, it's not like a
01:25:31.05 big issue. But the following image that was taken from a
01:25:37.08 propeller plane, that was taken with a camera that has this
01:25:43.11 rolling shutter. You can see that there are beautiful artifacts
01:25:47.00 happening because we're now exposing the camera and
01:25:50.06 not in one go. So you get kind of very weird looking
01:25:56.20 pictures. So be aware of that rolling shutter issue.
01:25:59.26 Another specification of the camera is its dynamic range.
01:26:04.28 And that is, how many different intensity levels can it
01:26:10.07 discriminate? Well that depends on the full well capacity.
01:26:14.22 So how charge can I pack into a well before it starts overflowing?
01:26:20.10 It also depends on what is the smallest step that I can
01:26:25.21 discriminate? Now the full well capacity is something
01:26:29.21 you or the manufacturer can measure. The smallest step
01:26:32.19 is given by the read out noise. That means that we can
01:26:37.02 simply divide the full well capacity by the read out noise,
01:26:40.18 and that gives us the dynamic range. This is a little bit
01:26:45.01 of an optimistic measure. It might be a little bit better
01:26:48.06 if it's something like 90% of the full well capacity, and then we
01:26:52.26 also say our steps are 3 times the read out noise, that's
01:26:57.15 a lot more conservative. And that still would give us about
01:27:02.02 1000 different gray levels that we can really
01:27:06.23 reliably discriminate from each other. Well, our
01:27:10.05 computer screens cannot display more than 256 different
01:27:14.07 gray scale levels, and in part that is because our human
01:27:17.24 eyes really, once they're accommodated to a certain light
01:27:21.24 intensity, really cannot discriminate more than about 100 or so
01:27:26.16 different intensities. So, this number was also given on our data
01:27:31.26 sheet. We can see here that we had a full well capacity
01:27:35.21 of something like 18000 electrons, then with a 6 electron
01:27:40.29 read out noise, that gave us a dynamic range of 3000.
01:27:45.10 So, the manufacturer used this calculation here, where
01:27:48.26 they divided full well capacity straight by the read out
01:27:51.25 noise. That also leads us to another important factor, and that is
01:27:58.13 so-called bit depth. And this comes more so from our computer
01:28:03.20 because our computer likes to work in bytes that consist
01:28:07.10 of 8-bits. So a camera now has 3000 different grayscale
01:28:13.06 levels. Well, how is it going to send that to our computer?
01:28:18.02 If it would pack that in one byte, then it would be a lost cause,
01:28:23.27 because you can only stick 256 gray scale levels in that.
01:28:28.00 So you would lose a lot of information. So that means
01:28:32.00 that our camera should be using at least 12 bits or
01:28:36.05 4096 gray scale levels to maintain all that information.
01:28:42.25 So, you really want to be sure that your camera is
01:28:48.28 spitting out enough bits per pixel to the computer, to
01:28:54.17 really reflect the full dynamic range that it can provide you.
01:29:00.03 So this is also often listed by the manufacturers, so we see
01:29:05.18 that this camera has two different modes. And in one of them,
01:29:09.01 it can give you 16 bits or 65000 or so grayscale levels.
01:29:14.10 Which then, by the way, are oversampled because we saw
01:29:19.03 that it cannot discriminate more than 3000 different
01:29:23.02 intensity levels. One important thing to remember is
01:29:29.20 that the number that you get in the computer does not directly
01:29:33.22 relate to the number of photons hitting that pixel. That is what you
01:29:39.09 in the end really want to know. So first of all, when there are
01:29:44.05 no photons hitting the chip, you do not get the number zero
01:29:49.11 in your computer. And that is because we have this read out
01:29:53.11 noise, and so if you would now clip all that, you would really
01:29:59.24 be clipping information. And there would be information missing.
01:30:03.13 So the manufacturers often set a certain offset, something
01:30:07.16 like you know, 100 or whatever. And in some cameras, you
01:30:11.29 can actually change that offset yourself in the software.
01:30:15.07 Always when there is 0 pixel values, it's highly suspicious
01:30:20.29 that something weird is going on, and don't trust those
01:30:24.22 measurements. Also, an increase by 1 does not mean
01:30:29.19 that there was one more photon hitting the pixel.
01:30:33.05 So, the so-called electron conversion factor, that is
01:30:39.10 the factor that relates these digital numbers in the computer
01:30:42.26 to the number of photons that hit the pixel. It depends on certain
01:30:47.18 parameters. For instance, it will depend on that electron
01:30:51.16 multiplication gain that we talked about earlier. Or also
01:30:54.25 when the camera has an analog gain in its analog to voltage
01:31:01.05 convertor, it will change depending on the setting of that gain, as well.
01:31:05.07 Now, that photon shot noise that we talked about before,
01:31:12.04 happens to give you an amazing way to measure what that
01:31:18.17 electron conversion factor is. And this is also how the manufacturers
01:31:23.22 themselves measure this. So basically, we saw that
01:31:28.09 photon shot noise goes with the square root of the number
01:31:32.29 of the photoelectrons on the camera. Electron conversion
01:31:39.26 factor, photoelectron conversion factor itself, is given
01:31:44.21 by the number of photons times this factor, times
01:31:50.25 the digital number. The digital number that we measured from this pixel.
01:31:55.11 And likewise, the noise in this also goes with the electron
01:32:04.11 conversion factor. And when we solve those equations for each other,
01:32:08.11 you'll see that the electron conversion factor can be calculated
01:32:13.06 by dividing the digital number of a given measurement by
01:32:20.22 the variants, the square of the noise, at that given pixel.
01:32:25.18 And so, this equation is valid when we're limited only
01:32:30.22 by photon shot noise. So if we can set up an experiment
01:32:35.03 where we're purely photon shot noise limited, we can
01:32:39.00 by measuring intensity and by measuring the noise, standard
01:32:45.26 deviation, we can backtrack the electron conversion factor.
01:32:50.02 So in general, measurement of signal versus noise is
01:32:57.27 extremely useful. We call these photon transfer curves.
01:33:02.15 And so you see here a kind of archetypical photon transfer
01:33:06.29 curve. So the way this is done, there are two different ways, one is
01:33:11.25 that you provide constant illumination. And then at every
01:33:17.15 pixel, you take many, many measurements. And you calculate
01:33:21.02 the average pixel value, and you calculate the noise
01:33:24.11 on that pixel. Alternatively, if you can really illuminate your
01:33:28.27 chip very homogeneously, you can measure the
01:33:33.00 average over all these different pixels. And you can measure
01:33:36.13 the standard deviation, the noise, over all these pixels.
01:33:39.18 You do this at different intensity values, and then you plot
01:33:43.17 on a log log scale, the average versus the noise.
01:33:48.11 So in these plots, what you will see is that at low
01:33:51.25 intensities, that the noise will converge to a certain
01:33:59.09 constant. And that constant is of course, the read out noise.
01:34:02.11 Because no lights, the only noise you see is the read out noise.
01:34:06.11 Then when we have very high intensities, we'll get to a
01:34:10.26 regimen where we're overfilling the well. And since they're all
01:34:15.11 clipped, the noise in the measurement will also drop dramatically.
01:34:18.26 And so this point will give you the full well capacity.
01:34:23.06 Then, on these linear areas on these curves, where
01:34:28.08 the intensities scale as the square of the noise, this is where
01:34:39.26 we can read out the electron conversion factor. And
01:34:44.02 it's really kind of amazing how simple these measurements
01:34:47.21 are. And I made a couple of scripts in our Micro-Manager
01:34:54.20 software that can do these measurements. And here you see a
01:34:58.20 photon transfer curve that was actually measured on one of our cameras.
01:35:02.09 And that gives directly the electron conversion factor.
01:35:08.02 Okay, this is where I'm going to leave this lecture about
01:35:13.15 camera specifications. I do want to thank Kurt, Kurt Thorn.
01:35:18.04 He and I work both on this lecture quite a bit at various points
01:35:24.08 in time. Some illustrations came from the micro.magnet.fsu.edu
01:35:29.25 website, which is a wonderful resource for anything
01:35:33.01 microscopy related. James Pawley's Handbook of Biological
01:35:38.02 Confocal Microscopy has a great chapter about everything
01:35:42.00 you want to know about CCD cameras, and then a little more.
01:35:45.22 And for photon transfer curves, there's a whole book on
01:35:50.03 about them by James Janesick, who also has another very good
01:35:54.04 book about CCDs and how CCDs work. Thank you.
01:36:00.10

This Talk

Speaker: Nico Stuurman

Audience:

Researcher

Recorded: April 2013

Talk Overview

This second lecture on detectors discusses how to match the pixel size of your camera to the optical resolution of the microscope, and it touches on quantum efficiency and its improvement by back-illumination. It then discusses various noise sources (such as photon shot noise, read-noise and fixed pattern noise), ways to reduce noise in the image (such as binning and electron multiplication gain), and shows how measurement of noise can be used to measure the camera’s photon conversion factor.

Questions

You are always using a 20x 0.75na objective and green (500nm) light on your microscope, which only has a 1x tube lens. Which pixel size should your camera have to make full use of the optical resolution of your microscope (assume that all other factors are equal, and that the optical resolution is given by 0.61 * lambda / NA)?
1. 27 micron
2. 16 micron
3. 6.5 micron
4. 4.0 micron
When selecting a camera, you generally want:
1. High quantum efficiency and high read-noise
2. High quantum efficiency and low read-noise
3. Low quantum efficiency and high read-noise
4. Low quantum efficiency and low read-noise
Back-illuminating a camera sensor improves
1. Readout speed
2. Readout noise
3. Quantum efficiency
4. Longevity
Binning pixels of a camera will
1. Increase spatial resolution and increase temporal resolution
2. Increase spatial resolution and decrease temporal resolution
3. Decrease spatial resolution and decrease temporal resolution
How many gray levels can an 11-bit camera record?
1. 11
2. 1111
3. 256
4. 2048

Answers

Speaker Bio

Nico Stuurman

Nico Stuurman

Nico Stuurman is a Research Specialist at the University of California, San Francisco, in the lab of Ron Vale. Nico combines his expertise in computer programming and microscopy to advance many projects including the Open Source software, Micro-Manager. Continue Reading