Top
SCIENCE. STORIES. IMPACT.

iBiology

Bringing the World's Best Biology to You

Introduction to Bioimage Analysis

00:00:13.13 Hello. Welcome.
00:00:14.22 My name is Christian Tischer.
00:00:18.01 I work in the Centre for Bioimage Analysis
00:00:20.06 at the EMBL in Heidelberg.
00:00:23.02 And I will give an overview of bioimage analysis.
00:00:27.11 So, what is bioimage analysis?
00:00:29.11 I think one can say it's an emerging field in computer science
00:00:32.29 that deals with extracting quantitative information
00:00:35.23 from biological image data.
00:00:37.23 So, why is it that such a field is emerging just now?
00:00:41.18 I think one reason is that
00:00:43.24 advances in microscopy automation
00:00:46.05 made it now possible to acquire so much data
00:00:48.11 that we cannot analyze it any more, manually.
00:00:51.07 So, I have some examples here.
00:00:53.24 For example, in screening microscopy,
00:00:57.04 you acquire data of many cells
00:01:00.01 treated with different compounds, sometimes thousands,
00:01:02.20 so that gives you, really, many images that need automated quantification.
00:01:07.14 Another example is electron microscopy.
00:01:09.25 Here, there is so much data
00:01:11.28 because they are acquired at high, nanometer resolution,
00:01:15.00 and also, now, volumetric,
00:01:16.26 so one single data set can also be several terabytes.
00:01:20.07 And then, finally, as a last example,
00:01:23.10 I put light-sheet microscopy.
00:01:24.23 So, here the large amount of data
00:01:27.22 mainly comes from the fact that you can acquire
00:01:29.24 really long time-lapse movies
00:01:31.28 of even some whole developmental processes.
00:01:35.10 So, obviously for any of these modalities,
00:01:37.16 it's really hard to analyze manually.
00:01:40.29 That's why we need bioimage analysis.
00:01:43.11 So, what can bioimage analysis give you concretely?
00:01:46.12 Here's one example.
00:01:48.13 It's just a very simple 2D image example.
00:01:51.14 Here on the right, we have a control sample
00:01:53.24 with normal-looking interphase cells, mostly.
00:01:57.08 Then we have a treated sample,
00:01:59.07 and we see many cells look different now.
00:02:02.11 Now, maybe 10-20 years ago,
00:02:04.14 it was still state of the art
00:02:06.19 that you just published these two images
00:02:08.06 and say these are representative,
00:02:09.29 and then you have your scientific statement.
00:02:12.08 I think this has changed quite a bit, now,
00:02:15.04 so you really need to quantify.
00:02:16.16 So, for example, you have to make such a plot
00:02:18.13 and compute the fraction of non-interphase cells
00:02:21.01 in these two samples.
00:02:23.07 Now, ideally, this should also be reproducible,
00:02:26.04 which means it should typically be not done manually
00:02:29.12 but by some automated computer algorithms,
00:02:31.20 so that you can run it again and you'll get the same value.
00:02:35.06 And then, finally, you shouldn't only do this for one image,
00:02:37.18 but for many, and do statistical testing
00:02:40.12 to see that your finding is really true.
00:02:42.23 And all of this is part of the discipline of bioimage analysis.
00:02:47.25 Another thing where bioimage analysis can really help you a lot
00:02:51.16 is if you want to find something in a bit data set.
00:02:53.13 So, as I already pointed out,
00:02:54.19 sometimes you not only have two treatments but maybe 10,000.
00:02:58.07 Now, to see that treatment number 8,756 induced an effect
00:03:03.26 can be very tedious,
00:03:06.29 and so it's nice if there's a computer that tells you,
00:03:08.29 look, this is an image you really should have a look at.
00:03:12.29 Then, last but not least,
00:03:15.02 we also want to extract really biophysically meaningful numbers.
00:03:19.12 So, in this I think very beautiful movie,
00:03:21.25 we see a molecular motor walking along microtubules.
00:03:25.23 And the natural thing that we want to know is,
00:03:27.24 how fast does it move?
00:03:29.16 And so, for example, here,
00:03:31.18 the scientists did a really good job.
00:03:34.02 So, they measured, for many of these motors, how fast they move,
00:03:37.16 in micrometers per second.
00:03:40.25 So, that's also something that bioimage analysis is doing.
00:03:45.05 Now, how do you actually do this?
00:03:47.14 So, to know how bioimage analysis works,
00:03:50.04 more technically you first have to understand
00:03:53.01 what an image really is.
00:03:54.12 And my attempt, here, of defining an image
00:03:56.12 is to say, well,
00:03:58.20 it's a set of values at certain coordinates.
00:04:02.11 So, in the simplest example,
00:04:04.16 the coordinate system would be 2D,
00:04:05.20 so you have an x axis and a y axis,
00:04:08.28 as pointed out here,
00:04:10.02 and then at different coordinates
00:04:12.01 you have different values.
00:04:14.26 So, essentially, you have a 2D matrix in this case.
00:04:18.08 Now, this 2D matrix representation
00:04:20.19 is very good for the computer,
00:04:22.08 because it can compute something on it.
00:04:25.01 But for our visual system,
00:04:26.22 that's obviously not the best representation.
00:04:28.22 That's why what we normally do is
00:04:31.14 we apply a so-called lookup table,
00:04:33.04 where, for example, in this case we say,
00:04:35.07 okay, 0's should be painted black,
00:04:37.04 and values of 255 should be painted white,
00:04:40.04 and then we have a nice image that we can actually look at.
00:04:44.24 Another thing you have to know is
00:04:48.22 a bit about image formation
00:04:50.18 to interpret these values correctly.
00:04:52.04 So, there are two things that happen during image formation,
00:04:54.20 and that's blur and noise.
00:04:56.24 So, in this example, you see the bottom of a cell,
00:05:00.17 and it's expressing a growth factor receptor.
00:05:04.14 Now, this line here is a line profile,
00:05:10.25 and you see...
00:05:12.20 so, this is basically corresponding to this plot.
00:05:14.10 So, this is the coordinate in the image,
00:05:16.07 and here we have the intensity in this image.
00:05:18.25 So, you see the bright edge,
00:05:21.11 and then you see we are in the cell.
00:05:24.12 Now, the width of this peak, for example,
00:05:27.07 is on the order of a few hundred nanometers,
00:05:29.06 so that means that it's probably not true...
00:05:31.29 not biological, but it's just diffraction-limited.
00:05:35.10 Another thing you have to look at, too,
00:05:38.14 are these variations in intensity.
00:05:40.11 So, you have to wonder, are they really true?
00:05:42.13 Are those really different receptor concentrations, or is it something else?
00:05:45.19 So, it could be very well noise
00:05:47.22 that could be introduced by the detector,
00:05:49.28 or just by the stochastic emissions of the photons.
00:05:53.19 So, you have to be aware of these two effects
00:05:56.22 when you interpret your image data.
00:06:00.12 Another thing one can see here is...
00:06:05.10 so, typically the blur is worse in the Z direction
00:06:10.16 compared to the X-Y direction,
00:06:12.10 so I indicated it here by drawing these point spread functions,
00:06:15.16 and I also indicated the outline of the cell.
00:06:18.01 Now, you see there's something interesting happening.
00:06:20.25 If you go to the edge of the cell,
00:06:22.24 more of the cell membrane
00:06:25.29 is fitting into this point spread function than down here.
00:06:28.26 And that means pixel intensities
00:06:31.16 will be brighter here than here,
00:06:33.18 even though, biologically, nothing interesting happened.
00:06:36.08 So, the fact that we see these bright boundaries here
00:06:39.24 might be actually an artifact of
00:06:43.25 the asymmetric detection volume of our microscope.
00:06:47.25 So, that's another thing you always have to pay attention to.
00:06:50.11 So, I highly recommend always drawing these pictures,
00:06:52.24 especially if you work at the reflection...
00:06:55.28 diffraction limit,
00:06:57.14 to be sure that you interpret your data correctly.
00:07:01.01 Now, there is now a whole field
00:07:04.08 -- it's called image restoration --
00:07:06.11 that tries to counteract the blur and the noise.
00:07:08.18 So, here you see an example, again,
00:07:10.19 which obviously has some blur and some noise.
00:07:14.10 And down there you see
00:07:16.11 the restored version, which is looking much nicer.
00:07:21.07 So, the classical denoising method
00:07:22.29 would be probably just to average the image,
00:07:25.13 and the deblurring with what would be called
00:07:28.14 deconvolution.
00:07:30.17 Now, as we all know, we live in the decade of machine learning,
00:07:33.10 so this specific example actually has not done...
00:07:36.17 been done with conventional techniques,
00:07:38.24 but they used a deep convolutional neural network
00:07:41.13 to simultaneously denoise and deblur,
00:07:44.05 and they got this really nice result.
00:07:46.29 Now, if you do this,
00:07:48.20 there's always this discussion, yeah?
00:07:50.12 Is this really true, here,
00:07:52.13 or did we invent, maybe,
00:07:53.26 something during the process of image restoration?
00:07:56.04 I think there's no easy answer,
00:07:58.02 so it's really our job as scientists
00:08:00.10 to provide control experiments
00:08:02.07 that convince ourselves that what we did here
00:08:05.00 is the good thing.
00:08:07.05 Okay.
00:08:08.24 So, let's get a bit more, really,
00:08:10.23 into the math of bioimage analysis.
00:08:13.12 So, let's look at a simple averaging filter.
00:08:16.10 So, here we have a noisy image.
00:08:18.28 And after averaging, it looks,
00:08:21.19 obviously, much cleaner.
00:08:23.05 So... and the recipe in this case is...
00:08:26.26 so, for example,
00:08:29.20 here we have one pixel that had a value of 59.
00:08:32.08 Now, in the new image,
00:08:34.26 is has a value of 54.
00:08:37.25 And the recipe is we take all pixels in this 5x5 neighborhood
00:08:41.06 and we average them,
00:08:43.22 and that will be the new value,
00:08:46.08 which in this case is 54.
00:08:49.28 Another way of saying that is...
00:08:52.10 and that's where the name convolution filters come from...
00:08:54.29 is we convolve the image with a so-called kernel,
00:08:58.04 a 5x5 kernel, where in this case the weights are all 1,
00:09:02.23 and we get the result.
00:09:05.24 Now, you may say, hmm...
00:09:08.29 if I take all the numbers and I average them,
00:09:12.24 isn't it a big coincidence that I get 54
00:09:14.23 and not maybe 54.3?
00:09:17.00 And in fact, you are exactly right.
00:09:19.01 So, this is not entirely correct.
00:09:21.12 The correct result would be down here.
00:09:23.22 So, you get these floating point numbers.
00:09:26.14 So, why don't they show up here?
00:09:28.26 One additional thing you have to know about images...
00:09:32.06 they also have data types, yeah?
00:09:34.06 And they can be integer data types
00:09:36.12 -- that means you cannot have floating point --
00:09:38.26 or they can be a floating point data type.
00:09:41.08 So... and you can change between them.
00:09:43.22 That makes everything a bit complicated,
00:09:45.14 so in fact when you do image processing
00:09:47.10 you have to care about these technical details
00:09:50.14 and make sure that the data type of your output image
00:09:53.05 can actually represent
00:09:56.12 the result of the mathematical operation that you applied.
00:10:01.06 Here's just another convolutional filter,
00:10:04.05 a very classical one,
00:10:05.28 an edge detection filter.
00:10:07.07 So, we first see if we apply it on the right, actually,
00:10:10.00 all the right edges of the objects have been enhanced.
00:10:13.17 So, how does that work?
00:10:16.20 I computed here a little example.
00:10:18.19 Again, we want to replace the center pixel
00:10:21.07 by some value that's computed
00:10:24.05 based on the neighborhood values of the pixels.
00:10:26.18 So, in this case, we take these three values,
00:10:29.06 we add them
00:10:30.23 -- that's basically this recipe --
00:10:32.11 then we take those three values
00:10:34.16 and we subtract them from the other ones,
00:10:37.25 and then we get 116,
00:10:39.18 which is the value in the new image.
00:10:44.06 Why do we do all this image processing?
00:10:46.10 It's not only to make images look nicer,
00:10:48.25 but it's mostly to help image segmentation.
00:10:51.23 And image segmentation is going from a gray-value image,
00:10:54.28 which we see here of two nuclei,
00:10:57.02 to another kind of image,
00:10:59.06 which is basically an image of two objects, yeah?
00:11:02.10 So, that's a main task of bioimage analysis,
00:11:04.22 to segment images.
00:11:06.02 And image processing is an important tool.
00:11:09.17 So, in the next slide I have maybe
00:11:12.06 the simplest image segmentation workflow
00:11:14.10 that you can do.
00:11:15.18 So, again, we start with our gray-value image.
00:11:18.24 So, you see here all the numbers.
00:11:20.16 And we are at the edge of the nucleus,
00:11:22.00 so you see here they are a bit brighter.
00:11:23.23 They when we're inside of the nucleus.
00:11:25.19 Down here we are starting to get outside,
00:11:27.19 so we have lower values.
00:11:29.04 Now we simply say,
00:11:31.16 all the pixels that are brighter than 73
00:11:34.02 should become 1 in the new image,
00:11:36.04 and all the ones that are darker should be 0.
00:11:38.11 And then we get this new image,
00:11:40.05 and that's a so-called binary image,
00:11:41.29 because it has only two values.
00:11:44.18 Now, maybe for you, you think we might be done, right?
00:11:47.02 I mean, I can already very clearly see two objects.
00:11:50.07 But the computer has to do one other step,
00:11:52.23 and it's called a connected component analysis.
00:11:56.02 So, basically, it has to simply look...
00:11:58.00 which of these 1's are connected to each other.
00:12:01.00 And then, for example, those 1's are all connected,
00:12:04.01 and now we have a new image where all the values are 2
00:12:06.19 if they belong to nucleus #2,
00:12:08.23 and they would be 1 if they belong to nucleus number #1.
00:12:11.26 So, another thing you see here is a very typical thing.
00:12:15.13 I mean, we have three images,
00:12:17.26 but the meaning of the value in these images
00:12:19.20 is very, very different.
00:12:21.14 That's a typical thing for image analysis
00:12:23.19 that you have to pay attention to
00:12:25.20 when you look at these numbers.
00:12:29.06 Now, how does image processing
00:12:31.08 help us for image segmentation?
00:12:32.26 Again, here's an example.
00:12:35.06 So, we have now an image of a bit more noisy nuclei.
00:12:38.00 If we want to try to threshold this image,
00:12:40.25 we have a problem,
00:12:42.18 because some of the noise pixels outside of the nuclei
00:12:46.02 will be as bright as the pixels inside of the nuclei,
00:12:48.25 and that's why it doesn't really work.
00:12:50.25 But if we now first apply the averaging filter
00:12:54.22 that we have seen a few slides before,
00:12:56.17 and then threshold, then everything works beautifully, yeah?
00:12:59.00 So, in this case, for example,
00:13:00.18 we used image processing to reduce noise,
00:13:02.07 and that's how we facilitate image segmentation.
00:13:06.05 Now, I want to show one more example like this,
00:13:10.13 that's a bit more complicated,
00:13:12.02 but it's one of my favorite filters for biology,
00:13:14.11 so I have to show it.
00:13:15.23 It's the so-called tophat filter,
00:13:17.12 and it addresses a very typical problem in biology.
00:13:20.26 So, we want to count, now,
00:13:22.24 how many of these vesicles we have in the cell.
00:13:24.28 But there is also background fluorescence
00:13:27.14 of a protein that's not bound to the vesicles,
00:13:30.16 and that makes it very difficult to find the threshold.
00:13:33.00 So, for example, if we want to find this dot,
00:13:35.07 we will also get the whole nucleus,
00:13:37.19 because it's so bright.
00:13:39.23 So, what this tophat filter is doing,
00:13:41.20 it's actually producing a good estimation of the background
00:13:44.17 that's not vesicle,
00:13:46.12 and we can subtract it from the original,
00:13:48.28 and then we beautifully only segment the bright vesicles.
00:13:54.07 So, that's very useful.
00:13:55.28 And how it works, more technically, is
00:13:58.12 there's two so-called rank filters applied, where intensity values of individual pixels
00:14:02.26 are replaced by the minimum intensity
00:14:06.07 that you find in a certain neighborhood.
00:14:08.08 That's called an erosion.
00:14:10.01 And then we do here the opposite;
00:14:11.13 we do a dilation.
00:14:13.07 And then we get, if we subtract this, the result.
00:14:16.04 So, I know I didn't explain it,
00:14:19.11 like, in all detail,
00:14:20.29 but I really highly recommend looking at this filter.
00:14:22.29 And I also just wanted to make the point that
00:14:26.16 one filter can be, actually,
00:14:28.08 a sequence of filters that leads to the desired result.
00:14:33.07 Now, where it becomes really difficult
00:14:36.08 is electron microscopy data.
00:14:37.20 So, for example,
00:14:40.00 here maybe we want to find four different classes of objects,
00:14:42.24 actually not only two.
00:14:44.08 So, for example, the endoplasmic, the cytoplasm,
00:14:48.10 here we have mitochondria,
00:14:49.23 and we have some other class of unknown structure.
00:14:53.05 Now, how do we do this?
00:14:55.12 So, in principle, we have now find for all of these four classes,
00:15:00.19 like, strategies of image processing and thresholding
00:15:03.25 that will then highlight them and segment them.
00:15:07.12 So, this can become very, very difficult
00:15:09.14 and, like, close to impossible
00:15:12.14 to achieve in a finite amount of time.
00:15:14.08 So... and that really is where machine learning helps.
00:15:17.14 Basically, the computer, for us,
00:15:21.29 tries all kinds of different processing filters
00:15:24.14 and picks the ones that make sense, yeah?
00:15:26.23 So... and how you help the computer
00:15:28.22 is by painting a few annotations.
00:15:30.19 So, you label some pixels in the endoplasmic reticulum,
00:15:33.24 some in the cytoplasm,
00:15:35.11 and some in the mitochondria,
00:15:36.23 to tell it what it should find.
00:15:40.01 And then the computer computes all kinds of filters,
00:15:42.23 for example an edge filter and averaging filter,
00:15:44.24 but also many, many others.
00:15:46.24 Then it tries very many different thresholds,
00:15:49.07 and then basically finds the right combination
00:15:52.01 that gives you the good result.
00:15:54.07 And there have been quite some advancements in the recent years,
00:15:56.29 and so the keyword is
00:16:00.02 deep convolutional neural networks,
00:16:02.05 and one of the main ideas is
00:16:04.23 you don't only compute filters once,
00:16:07.17 but you do this actually several times in a row.
00:16:09.26 So, this direction is basically the depth...
00:16:12.10 so, that's why it's called deep convolution...
00:16:14.12 we don't do it only one time but several times...
00:16:16.18 and that really makes these algorithms
00:16:19.04 much, much more powerful.
00:16:20.12 And I highly recommend having a look at the publications down here.
00:16:24.22 Umm... now that we have segmented the image,
00:16:27.19 we actually have objects,
00:16:29.05 so we can measure something.
00:16:31.06 So, for example, we can measure how big they are.
00:16:33.23 So, we could just measure the area in pixels.
00:16:36.22 Then there is the so-called shape parameter,
00:16:39.29 so we can for example measure how round they are.
00:16:43.06 And we can also measure their intensity.
00:16:45.23 And I would like to point this out
00:16:48.00 in a bit more detail.
00:16:49.10 So, for example, this cell has
00:16:52.16 the brightest maximum intensity value,
00:16:54.10 and that makes total sense,
00:16:56.07 because it's in metaphase,
00:16:57.26 so the DNA is condensed,
00:16:59.29 and that's why there is a very bright pixel.
00:17:02.17 Now, what you can also measure is you can
00:17:05.12 measure the sum intensity in these cells.
00:17:06.26 And what's interesting is they don't actually differ so much.
00:17:11.05 And again, if you think about it,
00:17:13.06 that makes total sense,
00:17:14.20 because why should there be more DNA in metaphase
00:17:17.07 than, for example, in prometaphase.
00:17:19.06 So, I think this is a good example
00:17:21.12 to also show if you measure different aspects of intensity
00:17:25.20 you can get really different information.
00:17:27.24 So, it's very powerful,
00:17:29.18 but also you have to be a bit careful
00:17:31.11 to measure the right thing to get the right answer.
00:17:35.08 Now, you can do even better.
00:17:38.05 You can try to calibrate these measurements.
00:17:40.07 So, for example,
00:17:41.25 here we measured the area in pixels.
00:17:43.13 It can be much nicer if you have it in micrometers squared.
00:17:46.19 And that's usually relatively easy to get,
00:17:49.05 because when you are on the microscope
00:17:51.10 you know how big one pixel is,
00:17:53.06 so you can just apply this to get the...
00:17:56.05 the right units.
00:17:59.17 Now, it's also interesting... there is this circularity parameter,
00:18:02.22 where you actually divide the area
00:18:04.28 by the perimeter squared.
00:18:06.12 And you see it has no unit,
00:18:09.13 and that's because you divide two spatial units.
00:18:12.07 So, these kinds of numbers are nice,
00:18:14.08 because they're in some sense self-calibrated.
00:18:16.16 And usually if you do a ratio of something against something else,
00:18:19.07 it's a good idea,
00:18:20.17 because you cancel out all kinds of things, yeah?
00:18:23.05 So, that can be a good idea.
00:18:27.20 Then for intensities,
00:18:29.20 in fact it's also possible to calibrate them,
00:18:31.20 but it's a bit more work.
00:18:33.11 So, instead of just measuring gray values,
00:18:36.02 you actually measure the concentration, for example,
00:18:39.06 or for the sum intensity you really measure the amount of molecules
00:18:42.23 that you have in a certain region in the cell.
00:18:44.25 Obviously that's extremely nice to have,
00:18:47.04 but it's a bit more work.
00:18:49.00 So, one thing you can do is
00:18:51.05 you can combine the conventional microscopy
00:18:53.11 with a technology called fluorescence correlation spectroscopy,
00:18:57.22 which allows you to calibrate the intensity to numbers.
00:19:01.05 And again, there is a publication here
00:19:03.01 that you may want to have a look at
00:19:05.01 if you're interested.
00:19:09.00 Now, what bioimage analysis does for you
00:19:14.02 is it converts images to numbers, essentially, right?
00:19:17.01 So, we have all this data,
00:19:18.28 and then we maybe say, okay,
00:19:20.17 how many cells are in one of the images?
00:19:22.23 Or how bright are they?
00:19:24.12 Or similar.
00:19:25.24 And then we may look at all these numbers
00:19:27.04 and see, oh, this looks very interesting.
00:19:29.04 Something happened, yeah?
00:19:31.02 But then, what you naturally next want to know is
00:19:33.17 how does the image look like that corresponds to this number.
00:19:36.13 That's why you really also have to look for softwares
00:19:39.26 that allow you to go back from the numbers
00:19:42.16 to the images.
00:19:44.17 So, to give a more concrete example,
00:19:47.00 here the y axis is how fragmented the Golgi was in a certain cell,
00:19:53.07 so each dot is actually an individual cell,
00:19:55.26 and we have three different treatments.
00:19:59.00 Now you see this cell...
00:20:01.13 wow, it really had 80 Golgi fragments inside...
00:20:04.04 is that really true,
00:20:05.19 or maybe was it just noise in the cell
00:20:07.18 that has been wrongly segmented?
00:20:09.00 So, you really need to check.
00:20:10.23 So, what you really need is a tool
00:20:13.23 where you can literally click on this dot
00:20:15.13 and it shows you the image
00:20:18.00 with the segmentation that the computer did.
00:20:20.04 And in this case, it was really correct.
00:20:22.06 This is really a highly fragmented cell.
00:20:25.06 In contrast, if you click on one of these control dots,
00:20:28.00 you see, okay, everything looks like...
00:20:30.18 normal-looking Golgi.
00:20:32.18 Now, what you also should do for a few example images, at least,
00:20:37.11 is characterize how correct the automated computer segmentation actually was.
00:20:43.25 So, what you should do is...
00:20:46.08 you can count how many of these dots
00:20:49.01 were correctly segmented,
00:20:50.04 so that is the number of true positive dots.
00:20:52.19 Then you should also count
00:20:54.16 how many of them were wrong,
00:20:56.17 so this is false positive.
00:20:57.25 Then you can computer the precision.
00:20:59.23 And similarly, you can also compute
00:21:02.10 another quantity called recall,
00:21:04.24 that tells you how efficiently you actually found them.
00:21:07.14 So, it's good practice to do this
00:21:09.13 for a few images and basically,
00:21:11.16 like this, characterize how well the segmentation works.
00:21:15.22 Now, to do all of this,
00:21:18.00 you obviously need some software,
00:21:19.29 and I tried here to show some software packages
00:21:23.19 that I am aware of.
00:21:25.02 And actually, in orange are painted
00:21:28.19 software packages that will be covered
00:21:30.25 in subsequent talks here on iBiology.
00:21:32.22 So, probably many of you have heard about
00:21:34.22 ImageJ, which is one of the oldest
00:21:38.11 and most used softwares for bioimage analysis.
00:21:41.29 Then, another very important library
00:21:44.01 that you will learn more about here on iBiology
00:21:46.02 is Bio-Formats.
00:21:48.00 You might not see it much,
00:21:49.23 but you will use it every day,
00:21:51.00 because it allows you to open images
00:21:52.23 that are now saved in all these different file formats.
00:21:56.00 So, another very important thing.
00:21:57.22 Then you will also hear about CellProfiler.
00:21:59.23 That's a software that's very good
00:22:01.25 for analyzing this high-throughput screening data.
00:22:07.01 Another thing that you realize
00:22:09.05 when you look at that slide is that it's very full.
00:22:12.00 So, there's a lot of tools,
00:22:13.11 and they're also nicely distributed
00:22:16.08 across all major programming languages.
00:22:18.04 So, in one way, that's nice,
00:22:20.22 because it means you probably will for sure find the tool you need,
00:22:23.24 but it's also a bit of a problem,
00:22:25.07 because how do you find it,
00:22:26.25 and also how do you combine all of them?
00:22:28.24 So, here I would like to point out something,
00:22:31.25 a software that's called KNIME.
00:22:33.12 And I think it represents a recent trend.
00:22:36.02 And the trend is that you basically
00:22:39.02 have these image analysis components,
00:22:40.22 even in different programming languages,
00:22:42.12 and you can still chain them together to generate a workflow.
00:22:45.19 So, for example, in KNIME,
00:22:47.25 you could have a Java component,
00:22:49.08 and then you have a Python component,
00:22:50.26 and then a MATLAB component.
00:22:53.02 So, I think this is a good trend,
00:22:55.12 and I hope it continues and makes it easier to construct workflows
00:22:59.19 using the best tools of the respective languages.
00:23:03.14 Now, finally, as I said,
00:23:05.22 bioimage analysis is a rather new field,
00:23:08.10 so it's in fact still finding itself.
00:23:10.24 So, what's really a great thing is I think now that
00:23:15.00 we have this common forum,
00:23:16.23 and that's basically where almost all of the relevant
00:23:19.14 open source platforms came together,
00:23:21.02 and that's where you can ask questions,
00:23:22.24 and I really recommend it.
00:23:24.17 Don't be shy.
00:23:25.29 For example, if you would now like to know,
00:23:27.16 what's a good book that I can read
00:23:29.18 when I want to learn more about bioimage analysis?,
00:23:32.12 go to the forum,
00:23:34.00 and if you don't find it just ask.
00:23:35.19 There will be many people answering you happily.
00:23:39.17 Another noteworthy initiative, here,
00:23:42.22 is this Network of European Bioimage Analysts.
00:23:45.17 So, this, as the name says, happened in Europe,
00:23:48.28 where really bioimage analysts came together
00:23:51.26 and tried also to define the profession of a bioimage analyst.
00:23:55.18 So, kind of elevated from a postdoc that can do coding in the lab
00:24:00.10 to a real profession,
00:24:02.07 and that was also a quite successful initiative.
00:24:05.04 And there's more communities forming,
00:24:07.07 so I recommend just looking in your country
00:24:09.11 if you can find something.
00:24:11.24 Alright.
00:24:13.13 With this, I am at the end,
00:24:15.02 and I would like to thank you for watching.
00:24:17.07 And I wish you very happy imaging
00:24:19.28 and also very successful image analysis.
00:24:23.13 Thank you.

This Talk
Speaker: Christian Tischer
Audience:
  • Educators of Adv. Undergrad / Grad
  • Researcher
Recorded: March 2019
More Talks in Imaging in Biology
All Talks in Imaging in Biology

Talk Overview

Microscopy is a key technology driving biological discovery. Nowadays, microscopy based scientific findings must be substantiated by quantitative image analysis. The discipline concerned with such quantification of biological microscopy images is called bioimage analysis. Dr. Christian Tischer walks us through the main concepts of a typical bioimage analysis workflow. He explains how to quantitatively interpret the content of microscopy images and how to automatically detect objects in images and derive object based measurements. He also emphasizes the importance of visual inspection and quality control of automated image analysis. Finally, he presents an overview of current bioimage analysis tools and communities.

Concepts:
Bioimage analysis (scope and necessity); Image content; Image visualization; Image formation (blur, noise); Image processing (filters); Image restoration (denoising, deblurring); Image segmentation; Object based measurements; Calibrated measurements; Analysis results exploration and quality control; Bioimage analysis tools and communities.

Questions

Please read the text and select all the correct answers. Multiple answers can be correct.

  1. Bioimage analysis can help you to
    1. automatically analyze large amounts of image data.
    2. reproducibly extract quantitative information from images.
    3. quantify the form and structure of cells and organisms.
  2. The process of detecting objects in images is called
    1. image restoration.
    2. image segmentation.
    3. image registration.
  3. The discipline of image restoration is concerned with
    1. measuring intensities in images.
    2. measuring the shapes of objects in images.
    3. counteracting limitations inherent to microscopy, such as the diffraction limit.
  4. The recommended place to ask questions about bioimage analysis is
    1. www.image.sc
    2. https://www.flickr.com/help/forum/en-us/
    3. https://forums.adobe.com/community/photoshop

Answers

View Answers

Speaker Bio

Christian Tischer

Christian Tischer

Christian Tischer studied physics in Heidelberg, followed by Ph.D. at the European Molecular Biology Laboratory (EMBL) with Dr. Philippe Bastiaens working on microscope development and signalling in mammalian cells. Subsequently, he did a postdoc with Dr. Marileen Dogertom at AMOLF in the Netherlands, where he worked on microtubule dynamics. From 2009 until 2018, he worked… Continue Reading

More Talks in Techniques

Related Resources

Related Links:
https://forum.image.sc/t/bioimage-analysis-recommended-reading-and-viewing/28051

Reader Interactions

Comments

  1. Awala, Fortune O. says

    The presentation on bioimage analysis was insightful, please how do you quantify a complete chloroplast image.

    Reply

  2. Kanupriya says

    Thank you Christian Tische! That was a great briefing of the image analysis process. I have a question about the Results Exploration and Quality Control slide – What are the Softwares that can be used to achieve that mapping of data to the cell directly? Can you state if CellProfiler Analyst does that? It would be helpful.

    I look forward to watching the CellProfiler detailed followup!

    Thanks once again!

    Reply

Leave a Reply

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

iBiology and iBiology Courses are part of the Science Communication Lab (SCL). Our mission remains the same, to connect people to science. However, our focus has shifted to producing and evaluating cinematic films for education and the public, which you can find on the Science Communication Lab website. For more information, please see this blog post!