Introduction to Image Acquisition for Quantitative Analysis

Top

00:00:07.29 Hi.
00:00:09.05 My name is Nico Stuurman,
00:00:10.18 and I'm a research scientist in the lab of Ron Vale at UCSF.
00:00:15.23 I co-developed an open-source software package called Micromanager
00:00:19.28 that is built to acquire images with microscopes.
00:00:26.00 And here in this image analysis course,
00:00:29.13 we really hope to give you an overview
00:00:32.29 of everything you can do and what you should pay attention to
00:00:36.21 with image analysis.
00:00:38.18 And in this particular lecture,
00:00:40.11 we're gonna talk about where images are born,
00:00:42.23 where they come from.
00:00:44.25 And so, we of course start with a microscope.
00:00:48.19 We have an object, a cell.
00:00:50.21 We have a magnifying device,
00:00:53.15 an objective.
00:00:55.05 And in the end, the goal here
00:00:57.09 is to acquire data...
00:01:01.22 shine light or collect light that comes from the sample,
00:01:05.08 and collect numbers.
00:01:09.04 And what I want you to go away with in the end is that
00:01:13.16 we are going to care much more about those numbers
00:01:16.08 than how good the picture looks.
00:01:18.25 So, the primary goal here
00:01:21.26 is to collect numbers in a good way
00:01:23.28 so that we can analyze them.
00:01:26.00 And as a tool to do that,
00:01:29.15 we do need to look at them
00:01:31.23 -- we need to visualize them on the screen.
00:01:34.07 But it is not so much about...
00:01:37.00 this is not the end goal.
00:01:38.18 The end goal is the analysis.
00:01:40.14 The visualization is just a tool to a means.
00:01:44.21 Okay, so image acquisition...
00:01:48.09 you will need software.
00:01:50.00 You will need a computer for it.
00:01:51.20 And part of the role that this computer and this computer software is gonna play...
00:01:57.16 our package Micromanager is just an example...
00:02:00.14 there are many, many different ones out there.
00:02:03.05 What they do in part is robotics.
00:02:06.16 It's moving things.
00:02:07.27 It's switching light sources on and off.
00:02:10.04 It's doing that all in an orchestrated manner,
00:02:13.16 where you, the user, are in control
00:02:15.28 and tell it what program to play.
00:02:21.04 So, for instance, you will want the software
00:02:24.02 to switch on a light source to start illuminating the sample.
00:02:27.20 At that point, you can start...
00:02:30.03 your detector can start collecting the data.
00:02:34.01 Once that collection of data is finished,
00:02:37.09 you want the software to transfer those numbers
00:02:40.26 back into the computer
00:02:43.04 so you get them in the computer's memory.
00:02:46.14 The software will also do things
00:02:49.07 such as moving a Z-stage
00:02:51.09 so that you can take a through-focus series.
00:02:53.23 You want it to change filters to change illumination.
00:02:57.27 You may want it to move the XY-stage
00:03:00.13 to take images at different positions.
00:03:04.06 So, the robotics is an important part,
00:03:07.06 but I'm not gonna talk much more about that.
00:03:09.10 What we really care about is now the detector,
00:03:12.14 and the numbers that it collects,
00:03:15.24 and how to do that in a faithful manner.
00:03:19.14 So, what kind of detectors are out there?
00:03:24.22 Very quickly and very shortly,
00:03:27.10 there are single point detectors,
00:03:30.29 things like PMTs,
00:03:33.20 avalanche photodiodes,
00:03:36.04 that only measure light at a single point.
00:03:39.23 So, those are usually used in confocal spot scanners,
00:03:42.17 and they collect data really fast.
00:03:46.21 You image many, many different points,
00:03:48.19 and then the computer builds up an image.
00:03:52.13 On the other hand,
00:03:54.07 there are things like cameras
00:03:56.13 that acquire lots of spatial elements all simultaneously.
00:04:00.23 So, they are more like parallel detectors.
00:04:02.29 And recently, we're seeing kind of things in between
00:04:05.18 that collect smaller areas,
00:04:07.07 really, very fast.
00:04:09.04 But in the end,
00:04:10.27 what all these detectors do is that they...
00:04:13.19 at a local spatial element,
00:04:16.07 they collect photons for a specified amount of time
00:04:20.15 -- your exposure time --
00:04:22.08 and convert that into electrons.
00:04:24.26 And that is always through the photoelectric effect.
00:04:29.15 One thing you care about there is the quantum yield
00:04:33.14 -- how well does your detector convert photons into electrons?
00:04:38.22 Every detector also has a noise aspect.
00:04:43.22 There usually is an amplifier
00:04:46.02 to amplify the signal.
00:04:47.19 There will be noise associated with that.
00:04:50.27 And eventually, there has to be some kind of conversion
00:04:54.29 from an analog signal to a digital signal.
00:04:58.05 So, we go from an analog number...
00:05:01.13 an analog entity to a digital fixed number
00:05:07.07 that now can be moved through computer memory
00:05:11.09 without any loss whatsoever.
00:05:14.01 In the AD converter,
00:05:16.05 what's important is the range.
00:05:18.12 So, what kind of range of numbers
00:05:21.11 does it need to deal with?
00:05:23.07 And so, in all these different detectors,
00:05:25.28 these principles are always at play.
00:05:28.05 But the important thing to remember is
00:05:30.12 what you're really measuring is photon flux.
00:05:33.00 You measure the number of photons over a time unit.
00:05:37.05 So, photon flux.
00:05:42.16 A couple of the important detector features.
00:05:45.16 That noise that we talked about.
00:05:47.13 It's kind of a fixed value
00:05:49.22 that will always be added at every intensity value.
00:05:54.19 And when you think about that a little bit,
00:05:57.06 it's that that noise, that noise in your detector,
00:06:00.22 will determine the minimal meaningful step
00:06:03.14 to go from one intensity level to the next.
00:06:08.05 Then there will always...
00:06:10.05 there's an issue of a zero.
00:06:11.29 So, you would think,
00:06:13.23 well, if I have zero photons
00:06:15.27 then my digital number should be zero.
00:06:18.08 And in a way, you're right.
00:06:20.03 But on the other hand,
00:06:21.27 there's a problem there,
00:06:23.16 because the detector adds noise.
00:06:25.07 And if you add noise over an average of zero,
00:06:27.29 you end up with negative values.
00:06:30.08 So, that is...
00:06:32.15 you can...
00:06:34.00 you could potentially accommodate those in computer memories,
00:06:36.14 but it's kind of very painful and wasteful,
00:06:40.16 so that is never done.
00:06:42.08 And the solution to this is
00:06:44.14 just to ramp up the zero a little bit
00:06:46.10 so that the noise is actually visible.
00:06:49.02 So, whenever you see an image like this
00:06:51.23 that is all black and is all zeros,
00:06:53.20 you know there's something not right here,
00:06:55.28 because someone is hiding the detector noise from me,
00:07:00.06 and I need to know that.
00:07:02.03 And so an image like this,
00:07:04.00 that may look like a lot of noise...
00:07:06.13 if that is your dark image, that is the good thing.
00:07:09.21 And this is what you're looking for.
00:07:14.05 The range, the digital range.
00:07:17.08 It of course should fit the range
00:07:19.21 that your detector actually delivers.
00:07:23.13 So, the maximum signal from your detector
00:07:25.22 has to be caught in that range.
00:07:28.12 But then the number of steps
00:07:30.19 is determined by the noise,
00:07:33.16 because that noise -- as we in the previous slide talked about --
00:07:36.28 determines the step size.
00:07:38.23 So, that total digital range
00:07:40.14 is going to be the maximum signal
00:07:42.22 divided by the step size
00:07:44.24 plus the offset.
00:07:47.06 And so the range that is useful
00:07:49.23 will depend on the detector.
00:07:51.09 And as very general rules,
00:07:54.10 you know, those single point detectors...
00:07:56.04 they're usually very fast.
00:07:57.25 They collect only a very few photons.
00:07:59.28 So often their dynamic range is relatively small.
00:08:03.21 And so you can get away with
00:08:06.03 a relatively low range of intensities.
00:08:09.16 Whereas cameras are often...
00:08:12.11 have a very high dynamic range.
00:08:17.01 And so you can sometimes discriminate
00:08:19.09 1,000 to 50,000 or even more
00:08:23.09 grayscale intensity values.
00:08:26.24 So, one thing that you would like to know,
00:08:31.08 but that is often not available,
00:08:33.24 is how to convert that digital number that you have
00:08:37.19 to the original number of photons
00:08:39.17 that was collected by the detector.
00:08:41.27 Whenever you can, find out.
00:08:46.00 That number will depend on things like gain of the detector.
00:08:49.24 But that really would give you a way
00:08:54.14 to go back to a physical, meaningful intensity value,
00:08:58.21 rather than something relative.
00:09:01.02 You will see in practice that you often work with relative changes.
00:09:04.29 This is not absolutely essential but desirable.
00:09:10.04 Now, we have numbers coming from the detector.
00:09:14.16 So, how are we going to display them?
00:09:16.09 So, we want to look at these numbers.
00:09:18.10 And often the simplest way
00:09:20.05 is to just put them on the display
00:09:23.07 as black, white, or gray images.
00:09:28.05 And so, of course, how you do that
00:09:30.13 depends on the range that is being delivered by the detector.
00:09:35.11 So, if you happen to have a detector
00:09:38.01 that only has 0's and 1's
00:09:40.06 -- so, it's a... 1 bit, in computer terms,
00:09:43.25 with a range of 2 detector --
00:09:46.13 then the most straightforward way
00:09:50.04 is to do blacks and whites,
00:09:52.01 and it will immediately be very obvious what is 0 and 1.
00:09:54.10 And this is what we call a binary image.
00:09:57.20 You will not often collect images this way,
00:10:00.04 but you will see later on that this is
00:10:02.19 a very important concept in analyzing images.
00:10:07.19 So, when you have more grayscale values...
00:10:10.04 and that... in computer lingo, we often go in bits.
00:10:15.20 So, 2 bits gives you a range of 4,
00:10:18.09 4 bits gives you a range of 16, etc.
00:10:23.19 So, you will notice, now,
00:10:25.14 that when you have more than 8 bits
00:10:27.17 we don't really see any difference anymore
00:10:29.29 in that grayscale value.
00:10:32.13 And that is because the monitors
00:10:35.17 in the computer screens that we work with
00:10:38.11 basically almost always have only 8 bits of dynamic range.
00:10:42.01 So, going from black to white,
00:10:44.27 there are only 256 grayscale values.
00:10:47.17 So, if your detector delivers more,
00:10:50.25 you cannot look at them at one and the same time.
00:10:54.10 Now... and that is a bit of an issue,
00:10:57.06 and an issue you will need to deal with.
00:11:00.26 And there are various ways of doing it.
00:11:03.15 And so, a very useful concept
00:11:06.18 is to think about lookup tables.
00:11:09.08 So, basically when an image
00:11:11.28 -- and this is the primary data --
00:11:13.24 has certain numbers,
00:11:16.05 we can now start assigning grayscale values
00:11:20.19 to those numbers.
00:11:22.09 And the most natural thing for this 2-bit image
00:11:25.03 is to just make a gradient from black to white,
00:11:28.07 and it gives us an image that looks like
00:11:31.07 this little spot here.
00:11:32.19 But now, let's for instance say that
00:11:35.07 we know that this is a little dot of
00:11:37.13 green fluorescent protein,
00:11:39.11 and we'd really like it to be green.
00:11:41.11 So then, you say, well, fine.
00:11:43.16 I'm just gonna make my colors, now,
00:11:45.25 green instead of black and white.
00:11:48.07 And then it looks green.
00:11:50.11 Nothing changes with the numbers here,
00:11:53.01 only the way you display it.
00:11:55.23 And you can go totally wild,
00:11:57.23 and you can make all kinds of other funky color schemes.
00:12:01.18 And these ones here don't look super useful to me.
00:12:05.11 There are definitely others
00:12:07.20 that are very useful.
00:12:09.08 But actually, when you think about it,
00:12:10.24 like, you know, it may be hard to see the black here,
00:12:13.16 and by making a 0 value red
00:12:17.08 you immediately see that there's... there's stuff there,
00:12:20.02 that there is... that there are pixels there.
00:12:22.18 Or, for instance, there's no number 2
00:12:24.16 in this set of numbers.
00:12:26.18 And with these three lookup tables,
00:12:29.02 that's immediately obvious,
00:12:31.06 because there's no blue or orange in any of these images.
00:12:35.25 So, you can really use these lookup tables to your advantage.
00:12:40.13 Now, things get more interesting
00:12:42.27 when you have images
00:12:46.15 that have a much wider range of possible values
00:12:49.23 than your display.
00:12:52.07 So, the natural thing is to do what we've done so far.
00:12:55.15 You set 0 to black
00:12:57.19 and you set the brightest value to white,
00:13:01.05 256 on the monitor.
00:13:03.17 And then you will see,
00:13:05.20 in a real image that is collected with a 16-bit sensor,
00:13:09.23 it can very well happen that you see
00:13:12.20 barely anything.
00:13:14.10 Now, the tool that most software packages should have
00:13:18.08 is shown here.
00:13:20.18 So, there are... and it's kind of a combination.
00:13:23.22 It's a combination of a histogram.
00:13:26.10 So, this histogram has on the x-axis
00:13:29.04 the intensity values as delivered by the detector,
00:13:34.08 and on the y-axis the number of pixels
00:13:36.24 with that given intensity value.
00:13:39.27 So, that histogram is an incredibly important tool
00:13:42.25 because it lets you look very quickly
00:13:45.12 at what the intensity values in your image
00:13:48.01 actually are.
00:13:49.15 It also lets you very quickly adjust the display,
00:13:53.24 because often you will want to set
00:13:57.03 the blacks to be the lowest value in the image,
00:13:59.23 rather than 0,
00:14:01.21 and it wants you to set the whites,
00:14:04.01 the brightest, to the highest value in the image.
00:14:08.00 Also note that actually the minimum and the maximum,
00:14:11.06 and even the mean,
00:14:12.29 are displayed here.
00:14:14.09 And this is something that many software packages will do for you.
00:14:17.24 And there's a trick here,
00:14:19.18 because since it's software it can do
00:14:21.18 boring stuff for you directly.
00:14:23.04 So, we could move these sliders
00:14:25.00 and change the brightness and contrast, basically, of the image.
00:14:28.18 But we can also say, autostretch.
00:14:30.27 And when I do that,
00:14:32.25 it will automatically set these little triangles
00:14:36.00 -- the white and the black point in the image --
00:14:39.28 and give me an... you know,
00:14:41.24 what usually is a pretty optimal display
00:14:44.13 onto the image.
00:14:46.00 Note here that almost all the possible pixel values
00:14:51.15 that the detector could deliver
00:14:53.27 are actually being shown as white.
00:14:56.16 So, the full dynamic range that we use
00:14:59.19 is pretty small.
00:15:01.13 Yet this is the dynamic range
00:15:04.19 that is present in our pixel values.
00:15:06.15 That's the stuff we care about.
00:15:10.25 Now, there are all kinds of tricks you can play
00:15:13.01 with these tools.
00:15:15.04 So, often it is useful to look at rare values
00:15:22.07 in your histogram,
00:15:24.08 so you often want to use a log scale,
00:15:25.28 a logarithmic scale,
00:15:27.20 for your y-axis,
00:15:29.11 so that the rare bright pixels actually stand out
00:15:31.11 and you can see them.
00:15:33.06 You may want to scale the x-axis
00:15:36.24 so that you actually get a little bit of a better view
00:15:39.18 on what pixel values you have there.
00:15:43.18 Again, we're not changing any of the original values here.
00:15:47.06 We're only changing how we look at them.
00:15:52.26 Then there are interesting lookup tables
00:15:56.23 that you can assign,
00:15:58.14 and often a very useful one
00:16:01.23 is to highlight pixels that are bright
00:16:04.10 and pixels that are dark.
00:16:06.13 So, here we're setting the top 0.1% of the pixels to red,
00:16:11.23 and we see some kind of structure in the cell, here,
00:16:14.08 that is red.
00:16:16.11 And we see that the top 1% of the dark pixels
00:16:19.01 are set in blue,
00:16:20.27 so we kind of see the noise, here,
00:16:23.06 in the dark areas.
00:16:25.27 And one important lesson when you're acquiring images
00:16:29.24 is that to maximize the information content,
00:16:33.08 you want to maximize the use of that histogram.
00:16:36.18 So, in the previous image
00:16:39.14 I used a 5 millisecond exposure.
00:16:41.28 For this one, I switched to 150 milliseconds,
00:16:45.22 and I got a way wider dynamic range.
00:16:48.26 And even though the image
00:16:50.23 looks almost the same as the previous one,
00:16:53.00 this is the one I want to do my analysis on,
00:16:55.19 because the information content
00:16:57.26 is much larger than in the previous one.
00:17:02.22 Now, to...
00:17:06.27 of course, this has the downside
00:17:09.22 that if your sample photobleaches
00:17:12.15 then you don't want to be doing this.
00:17:15.03 So, you always... you know, you have to live with your sample
00:17:19.05 and you have to live with the limitations that are imposed
00:17:22.05 by your biological object.
00:17:24.13 But if you can,
00:17:27.02 you want to optimize the content of your image.
00:17:31.23 And so... to maximize your information content.
00:17:39.04 One thing that we often do
00:17:42.02 when we are acquiring multiple images
00:17:44.18 under different illumination settings...
00:17:47.16 so multiple channels.
00:17:49.19 So, here, for instance,
00:17:51.07 we had an image from a fixed slide with...
00:17:53.22 where the nuclei were stained with DAPI in blue,
00:17:57.06 and actin stress fibers were stained with a green dye,
00:18:02.14 and we had mitochondria stained in red.
00:18:05.00 So, often you will want to look at those overlaid.
00:18:08.17 So, the first thing would then be to use
00:18:12.07 these colored lookup tables on them.
00:18:14.07 So, we now get the blue nuclei,
00:18:15.28 and the green actin,
00:18:17.11 and the red mitochondria.
00:18:19.14 And then you can overlay those in one image.
00:18:23.05 And then you get things that often look very pretty.
00:18:26.17 But that also lets you see
00:18:28.23 how things are located with respect to each other.
00:18:31.15 So, this will be a very common technique
00:18:33.12 in all acquisition software.
00:18:35.21 There's one thing to remember,
00:18:37.15 and that's that there are a certain percentage of people
00:18:40.22 who can't... who are red-green color blind.
00:18:44.07 And so if you're sharing these images,
00:18:46.27 it really is useful to use color blind-friendly colors,
00:18:51.18 and there are all kinds of resources out there
00:18:54.29 to help you select those.
00:18:58.08 Now, when you're acquiring things
00:19:00.26 like multiple Z-positions,
00:19:03.14 multiple XY-positions,
00:19:06.11 or time-lapses,
00:19:08.24 the way that is often represented in computer memory
00:19:11.17 is that these are stacks of images.
00:19:14.01 And you will have some kind of tool, a viewer,
00:19:17.07 that lets you go through those.
00:19:19.08 So, here I have a time-lapse image
00:19:23.17 taken at different Z-positions.
00:19:26.01 And when we now scroll through the Z,
00:19:29.19 you see that with changing the focal position
00:19:35.03 we look... we get an idea
00:19:37.27 about the three-dimensional organization.
00:19:40.03 And then we can play a time-lapse image of this.
00:19:42.19 This is actually a drosophila S2 cell
00:19:45.19 with an ER label
00:19:48.03 in mitosis.
00:19:50.00 And you saw in the time-lapse how that behaves.
00:19:55.29 One... so, in general,
00:19:58.13 I hope you got the message
00:20:00.25 that it is not good to change
00:20:03.23 those pixel values that you acquired.
00:20:05.21 There's one... before you start analyzing them.
00:20:09.05 There's one possible exception to that,
00:20:12.13 and that is correcting for your illumination.
00:20:16.10 So, this is an example of what
00:20:20.00 the so-called flatfield image looks like,
00:20:22.08 if we just look only at the way
00:20:24.24 we illuminate our sample in the microscope.
00:20:28.06 And you can see that points in the center, here,
00:20:31.21 are much brighter than points in the corners, here.
00:20:36.07 So, if you have cells and you want to compare
00:20:39.00 their intensities,
00:20:41.00 and you have cells here and cells here...
00:20:43.12 just by virtue of the illumination,
00:20:45.20 the ones here are going to be brighter than the ones here.
00:20:48.06 So comparing their intensities wouldn't make any sense,
00:20:51.11 because you're comparing illumination intensity
00:20:54.04 rather than the endogenous level
00:20:56.15 of whatever you are interested in.
00:21:00.03 Likewise, there is a background
00:21:02.21 coming from your sen... your sensor.
00:21:05.13 Usually, that is much more homogeneous,
00:21:07.23 so you're pretty good in that.
00:21:10.08 But if you can one way or another
00:21:13.15 measure the illumination profile,
00:21:16.09 and know your background,
00:21:18.07 you can correct for it.
00:21:19.21 And the way to do that is...
00:21:22.01 well, your real image is going to be
00:21:25.04 the object that you're interested in
00:21:27.23 multiplied by the illumination
00:21:30.13 -- the illumination intensity at every pixel --
00:21:33.22 with the background added.
00:21:35.25 And so, to now get your object out of this,
00:21:40.20 it is simply swapping this equation.
00:21:44.28 Your object will be the image that you measured
00:21:48.08 minus the background
00:21:50.10 divided by the illumination intensity.
00:21:52.22 And so this is something that
00:21:56.08 you want to be doing right away,
00:21:57.28 although even then there are some
00:22:00.01 rounding issues.
00:22:02.05 Were it not for those rounding issues,
00:22:04.03 you could always revert it
00:22:06.03 and get back to your image.
00:22:07.19 This is also something you can do after the fact,
00:22:09.27 but you'll definitely...
00:22:11.22 will be wanting to do this
00:22:13.21 before doing any further quantification of your images.
00:22:18.21 We've talked a lot about pixels,
00:22:21.19 about intensity data,
00:22:23.25 and I want to shortly introduce to you
00:22:26.27 the concept of metadata.
00:22:29.09 It's of course not only those intensity measurements.
00:22:32.17 Those intensity measurements
00:22:34.17 are the most important part -- I would say.
00:22:36.29 But there are of course...
00:22:39.11 you know, for your computer to do anything with this,
00:22:41.22 it will need to know...
00:22:43.18 first of all, okay, I have this string of bytes.
00:22:46.04 What does this represent?
00:22:48.26 So, the very first thing is, like,
00:22:51.17 well, this is the way I store my numbers.
00:22:54.09 I have, like,
00:22:55.28 1 byte per pixel or 2 bytes per pixel.
00:22:58.04 These are unsigned integers
00:23:01.10 or floating-point numbers.
00:23:02.28 d So, the computer has to settle on
00:23:06.09 how to interpret that byte stream.
00:23:08.15 It has to know simple stuff like, okay,
00:23:11.09 what is the width and height,
00:23:13.01 because otherwise it can never
00:23:15.17 reconstruct that image for you again.
00:23:18.03 Then there are metadata.
00:23:20.26 So, for instance, in our time-lapse,
00:23:22.25 we want to know, okay,
00:23:24.23 which plane is which time point?
00:23:26.15 What was the interval between these time points?
00:23:29.19 Which plane is what position?
00:23:31.07 What illumination did you use?
00:23:33.12 Etc, etc.
00:23:34.26 So, there's the kind of logistics
00:23:36.29 of what this all meant.
00:23:38.22 And that has to travel with your image,
00:23:42.12 or you're gonna be lost.
00:23:44.12 Then there's a whole lot of information
00:23:46.05 about the equipment that you used
00:23:48.18 to acquire this data.
00:23:50.16 And so, for instance,
00:23:52.06 in the Micromanager software,
00:23:54.06 we record all that information,
00:23:56.07 a lot of it very boring,
00:23:57.26 that you will never look at it again.
00:23:59.19 But some of them are very important,
00:24:01.08 like, you know, the pixel size.
00:24:03.23 What did this pixel physically represent
00:24:09.01 in the sample space?
00:24:11.02 So, it is important,
00:24:13.15 and it's important to keep that with your images.
00:24:17.10 And then there are...
00:24:19.03 what I call here experiment metadata.
00:24:21.04 What was this sample?
00:24:22.19 What drugs did I use on it?
00:24:26.17 What do I actually want to achieve
00:24:30.01 with this experiment?
00:24:31.10 Etc, etc.
00:24:32.22 And we will talk about metadata,
00:24:34.29 and how to keep and store and use them
00:24:39.09 in another lecture in this series.
00:24:41.23 One important thing to consider is that these datasets of images
00:24:46.21 tend to grow.
00:24:48.23 They can get very, very, very big.
00:24:52.12 So, we are now routinely in the lab
00:24:55.06 using cameras that have 2000 x 2000 pixels,
00:24:58.24 2 bytes per pixels,
00:25:00.15 so every plane, every image,
00:25:02.17 has 8 megabytes.
00:25:05.08 Now, when you, you know,
00:25:07.14 do a simple time lapse that we may be doing,
00:25:11.10 where you have 2 channels, 80 slices,
00:25:15.05 you get 1.3, roughly, gigabytes
00:25:18.27 for every time point.
00:25:20.08 And if you'd run that for a day,
00:25:21.29 you get close to 2 terabytes of data.
00:25:24.23 So, that is a lot to deal with,
00:25:27.03 to move it around,
00:25:28.25 to store it.
00:25:32.05 Usually, I'm the one in the lab who's dealing with it,
00:25:34.13 and the graduate students and postdocs
00:25:37.17 don't know what is going on.
00:25:39.07 But, you know, just to help people like me,
00:25:41.05 plus yourself later on,
00:25:43.03 because you will deal with this at one point or another,
00:25:46.12 you know, think about this,
00:25:48.12 and try to reduce the size of the images you're taking
00:25:52.19 to the absolute minimum necessary
00:25:54.22 to do your analysis.
00:25:56.28 So, if you can just use a low resolution
00:26:00.15 and still get the results out
00:26:02.27 that you're actually interested in,
00:26:04.27 please do it.
00:26:06.20 Don't go to the beautiful high-resolution images
00:26:08.24 just because they look cool.
00:26:10.27 You will regret that later.
00:26:13.27 And so, just as an example,
00:26:16.00 this is an image taken with a 40x objective.
00:26:20.01 If I were to be just interested
00:26:22.21 in the size of these cells,
00:26:24.06 kind of the outline of it,
00:26:26.07 I would do way better using a 5x objective,
00:26:30.24 where I get a 64-fold larger field of view
00:26:36.13 in one image of the same size
00:26:38.17 as with the 40x,
00:26:40.03 but see how many more data points I get in one shot.
00:26:43.09 So, use the lowest resolution possible.
00:26:48.29 The end of all of this is that
00:26:51.03 you're going to save your data
00:26:53.11 one way or another.
00:26:55.01 So, you're gonna save your pixels and your metadata
00:26:57.18 for further analysis.
00:27:00.10 There are many different file formats,
00:27:02.10 and we will be talking about some of this later on
00:27:05.07 in other recordings in this series.
00:27:11.22 An important concept to remember
00:27:14.03 is lossy and lossless saving.
00:27:20.26 So, lossy basically means
00:27:24.02 that you are throwing away data
00:27:26.14 to reduce file size.
00:27:28.02 So, all our cameras that we have in our phones
00:27:30.12 or SLRs,
00:27:31.29 they tend to save in this file format called JPEG
00:27:35.11 that has been optimized to preserve what we,
00:27:39.09 with our eyes, consider as important.
00:27:42.06 But that may not necessarily be what our analysis
00:27:45.11 considers important.
00:27:46.01 So, once you start saving in one of these formats,
00:27:48.27 you basically lost your pixel data.
00:27:51.24 And in a way, that's the end of it for analysis.
00:27:54.17 You're never gonna get them back.
00:27:56.01 So don't do that if you want to do
00:27:58.24 further analysis.
00:28:00.26 Now, there are many different formats for lossless saving.
00:28:06.29 You will always get to these TIFF kind of formats,
00:28:11.13 which is kind of more or less the universal format
00:28:14.04 at the moment
00:28:15.28 to exchange data.
00:28:17.20 And this all will be touched upon in later...
00:28:21.24 in other recordings.
00:28:23.28 For now, what's important is to know
00:28:27.23 how your software saves the data.
00:28:29.22 And you want to be absolutely sure that it is lossless.
00:28:32.11 And you want to be sure that it is in a way
00:28:34.15 that you can work with it again
00:28:36.16 in your analysis software.
00:28:38.20 And that brings me to the end.
00:28:41.01 So, I hope that you now are
00:28:44.29 convinced that your digital images are numbers with metadata.
00:28:49.27 And that you are responsible for keeping those intact,
00:28:53.01 and keeping them in a way that actually
00:28:56.15 faithfully represents what you were recording.
00:28:59.21 And do it in a way so that you can analyze them further.
00:29:03.11 And I want to do a short shout-out to Kurt Thorn.
00:29:06.10 He and I have done several versions
00:29:09.20 of similar kinds of presentations,
00:29:12.14 and we've really helped each other
00:29:16.02 a lot with this.
00:29:17.04 So, thank you, Kurt.

This Talk

Speaker: Nico Stuurman

Audience:

Educators of Adv. Undergrad / Grad
Researcher
Educators

Recorded: October 2018

Talk Overview

How do we visualize biological samples? In this talk, Dr. Nico Stuurman provides an overview of the different tools, equipment, and software available to acquire an image of a biological sample using a light microscope, and the considerations one needs to take when using these tools. This lecture will allow scientists to understand the principles behind image acquisition in order to improve and optimize the analysis of their sample.

Concepts:
detectors, detector noise, image display, binary image, grayscale values, illumination correction, metadata, file format

Questions

Lookup Tables (LUTS)
1. make it easier to visualize intensity measurements
2. can be used to artificially color an image similarly as you would see it through the microscope by eye.
3. make it possible to overlay images acquired at different wavelengths
4. all of the above
To acquire high quality image data
1. make sure that black and white are clearly showing in the image
2. use a 12-bit camera
3. use the histogram to ensure that >80% of the available intensity values are represented
4. make sure the image looks good on the display
Preferred file format to save images:
1. JPEG
2. GIF
3. AVI
4. TIFF
The digital numbers in an image represent (choose the most correct answer):
1. CCD pixel
2. photons emitted from a spatial element in the sample
3. the quantum yield of the detector
4. color in the sample

Answers

View Answers

Speaker Bio

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