 The user perspective is important here. How, as a user, is one going to efficiently analyze and share image data from the brain? And the perspective we have been taking is a system level perspective. And since brain systems are widely distributed, and since aging and brain disease affects all these systems at different levels in many regions, we have seen a need for looking brain-wide in analysis, that you look across the entire brain when you do your analysis in animal models and so on. And this brings on large collections of data that have to be managed somehow. This is also facilitated by emergence of new and exciting technologies that allow us to acquire image data more efficiently. Slide scanners, two-photon serial tomography instruments, high-resolution MRI, and so on. So we are looking at collection, or we are starting to amass large collections of data, and we need to organize them in an efficient way, and we need to analyze them in an efficient way. So how to navigate in rotten brain image data? That's more or less what I'll try to look at, and of course, the starting point here is that without appropriate assignment of location to the data, we are virtually lost at sea. And indeed, the perspective we take is that the most profound qualities of interpreting data is location, location, and location. So location and navigation is very important. So I wanted to start with a very short story, with one slide, summarizing what I will talk about and what the main point is, starting with the analog, the frequently used analog to brain attacing, which is cartography. When I traveled to Leiden yesterday, I had, of course, been assigned to a hotel. And since I studied in Leiden many, many years ago, 20 years ago, I considered myself an expert in the region. And I looked up the location of the hotel in an online atlas, very convenient, and found out that it is located between the central station and the academic hospital, which I know very well, so I thought, well, all is fine. Coming there last night, I found out, well, I couldn't relate to the nomenclature I was using. I was expecting a sign with hotel somewhere. I couldn't see it. Well, I had to re-consult my digital atlas, and I looked it up, and I couldn't really, well, I saw it, I saw the nomenclature, the name located there. So I had to switch from name to spatial coordinates and start to reorient myself. And looking closer, it turned out that we have the central station here, the marker was placed there, the hotel turned out was actually here. Again, illustrating the main points of what I wanted to convey to you, that correct annotation and assignment of location is extremely important. And very often errors are made, and it is important to have atlas resources that are updated in order to avoid errors in your navigation, and most importantly in data importation, because if you have such errors in data importation, they are actually very difficult to get rid of afterwards. All right, so then moving to the brain again. Many of the projects that we have have the where in the brain question as a common marker, or as a common thing. And this also goes for, for example, people doing electrophysiological recordings from the brain. Here we see DI labeled electrode tracked somewhere in the brain. If you get a dappy-stained image from the same section, you have some cytoarchitecture to navigate from, and then basically what you have to do is to look it up in an atlas. Brain atlases are, as you know, very important for our data importation, and it's very interesting to see that the rat-brain and stereotaxic coordinates has, or is, France as the most cited neuroscience publication with more than 30,000 citations over 25 years. But so you can look at your image, you can find the corresponding location in the atlas, and you can see if you can assign a name or spatial coordinates to it. It might not be all that simple. In particular, if it turns out that your section has a different angle than the atlas representation. So many researchers end up with a need to contact an expert to consult. In particular, if you're working with a campus and you're wondering, well, was my electrode in this or in this sub-region of this complex region, they usually want to consult someone with detailed knowledge of the anatomy there. And our collaborator, Mendovita in Trondheim, has indeed received a lot of such requests, which motivated him to start working on atlas in the field of atlasing. All right, so the challenges we face is that we have a large number of anatomical terms that are differently used by different researchers. The boundaries of these regions indicated by these terms are ambiguous. And if you look up in an atlas, it is extremely difficult to trace back the criteria used to draw the boundary, or the line, or the dotted line, or whatever you have. And that makes it very difficult. And of course, as I mentioned, differences in orientation. If your experimental material has a different orientation than the standard plane of the atlas, you very soon run into problems, at least in some regions. So it has been said that, well, what we need is a proper 3D digital brain atlas, and that will solve most of the problem, because it should be open access. It has to be volumetric. And if we have such a template, then we can assign standard coordinates. We can have the structure names in there, and voila. Of course, you do need tools and efficient workflows that actually allow users to take their data and their large collections of data into this space without having to study informatics or learning programming, which is something that is needed for many of the registration tools currently available. All right, so the solution is probably to have three-dimensional atlas. And over the last 10 years, many such spaces have been emerging, starting with reconstructions of the existing anatomical atlases of the rat brain and moving towards using high-resolution MRI templates. As you probably all know, one of the really powerful resources at the moment is the Allen mouse brain atlas, which is based on series of nissle sections that have been reconstructed, it is constructed as a three-dimensional volumetric atlas that can be sliced in different orientations, giving you access to all these segmentations in different planes. The INCF atlasing task force also some years ago defined a new standard reference space for the mouse and for the rat brain, the voxel space, which is based on internal landmarks that can be found in isolated sections where the skull has been removed, and of course that makes a lot of help. Interestingly, when you look at the available MRI voxel templates, it is actually also possible to identify skull landmarks in the MRI images, which allows you to also have the serotaxic coordinate system to get with a voxel space coordinate system, and it is possible to translate. Very useful. So at the moment we have both a voxel space template and an atlas for the mouse brain based on MRI, DTI, and with nissle sections and 37 delineated structures. Recently, since earlier this year, we also have the rat brain voxel space atlas based on MRI, DTI, high resolution images where at the moment 76 structures have been delineated and in the next phase additional structures will emerge. Just to show you a little bit more about the rat voxel space, high resolution structural MRI data, and diffusion tensor imaging data, it has been possible to actually find a high, or to delineate anatomical structures with a high degree of precision, combining information from these modalities. So if we just take a very quick look at just to illustrate how this atlas looks, here we go through the atlas from anterior to posterior looking at coronal images, here switching to DTI, here to the delineation, and you see that, well, it has a very high resolution. Here we tilt the angle of view and you see that, well, you can actually, since the MRI volume has such a high resolution, it is possible to cut it in any plane you want so that you can generate custom atlas diagrams fitting whatever section you have. So that's a very powerful tool since the resolution is so high. Okay, so then I wanted to move on. So we have several useful volumetric 3D atlases and then how should one go from the collection of images one has to atlas-based. There are many ways to do it and I'll just quickly review some of the examples that we have used to give you some neuroscience use cases because the method you choose depends on the tools you have, the needs you have and some of the registration methods available at the moment require expertise. Okay, so what we have done in several of our projects recently is to say, well, we need to collect brain-wide section images and these are important to share, to make the raw data available and high-resolution digital images have been organized in atlas repositories where the sections have been assigned a coordinate relative to the skull coordinate system. So here it is possible to look up a lot, 200, 300 images in a standard plane that have been lined up where one can read off a positional value merely saying, well, this section has been found to be located at this particular level. It's actually quite useful because it allows you to look up different experiments to say, well, I want to look at the thalamus, I want to compare labeling in two different experiments and then you basically have to look up corresponding coordinates, bring up the images, zoom in on whatever region you want to do and then superimpose an atlas. In this case, it's just a simple conventional Poxenos and Watson diagram but it turns out, well, using local boundaries, merging images, it actually gives you a nice view. But again, it is a tedious process and it doesn't really scale to 300 images but for certain analysis it's very useful. An other approach would be to say, well, if you have 200, 300 histological images, why not just reconstruct them, make a 3D volume and then merge the volume with the 3D atlas? That would probably give you a tool that could allow you to do brain-wide analysis very efficiently. So in collaboration with Daniel Wojcik and Piotr Majka in Poland, we have tried that using a set of images which has been assembled into 3D so that you can cut your coronal sections in different planes, giving you very interesting perspectives on the labelling patterns which has been very useful and it can be registered to an atlas. In this case, it's a 3D reconstructed atlas, not the voxel space, but it does illustrate it. And you can see that, well, it is possible to read off something, at least at the regional and subregional level that gives you an indication of where your signal is located. But it's not perfect. It doesn't match perfectly. Here we used a fine registration, so it gives you a good indication of where your signal is, but it doesn't give you the final answer without error. Okay, and then another example, moving back to this electrode location, recording site, and I'll try to show you how we can work with this using the voxel space or the rat brain, where you say that, well, the contrast you have here, it can be made into grayscale and it matches quite well the T2 star-weighted MRI image. So if you then map in the electrode track, it's a matter of looking up your image and saying, well, here we have the position that we are interested in. We can look up the voxel space atlas, find the corresponding level, and then start playing with rotation until we have found the position and angle of plane that matches. And then you can transfer your image by a fine registration. And if you then bring in the segmentation, you do get an answer of where this position is, and you see here how the angle of sectioning has been matched. Again, it gives quite good results, but it requires manipulation and at the moment there has not been a very efficient tool to do so. So we have been working on creating such a tool and we are starting to make some progress and I'll show you some very preliminary data of this. So for registering a 2D section, and here we have a series of them, we pick one of them, it is possible to then take it into a tool where you have a reference template. In this case, it's the voxel space atlas or rat. And then you can bring in your section image and then there are tools allowing you to match the atlas to your section image by manipulating the position and the rotation of the atlas slice in a tool. So you can here create an overlay where things start to match up. So still it's a bit trial and error to make it match, but it turns out to be quite user-friendly, giving you a reasonable segmentation of your image quite efficiently. Again, what you can see is that, well, having the voxel space in the custom, it's slightly off the coronal plane, you have it onto your tuning section, you might want to ask, well, I can see where the hippocampus is, but really where is the boundary between CA2 and CA3, for example? So then you need to look up the criteria for defining these boundaries. In this case, we can look up in a tool called the hippocampus atlas where sections from or an atlas containing different stained sections through the hippocampus that have been annotated. And most importantly, when you find the corresponding section, it's possible to go to an associated text library giving you the criteria. So you can see, well, the drop in the size of neurons indicates the boundary of CA2 and CA1, and then there is a difference in the thickness of the cell layer between CA2 and CA3, and you can look up more closely in the atlas and you can find the corresponding site, and you can start observing, well, here I see a difference in the thickness in the cell layer, and if you look more closely and you bring in the segmentation, you can start playing with this in the atlas using the criteria, going back to your experimental section and actually say, well, here it is. Then, so this is a way where you can say that the atlas and the registration of image to atlas gives you a first indication of where you are, but then you still need to look at the criteria for defining boundaries. So it's very important to have access to such criteria. Okay, and then moving back to the atlas, to the voxm space atlas, we have been, so the first version is out there and we have then continued to delineate in more detail the hippocampal region in collaboration with MenoVitter, and there, again, combining information from multiple MRI modalities, it turned out to be possible to identify most of the boundaries that have been identified in the histology, and most importantly, it is possible to give explicit criteria giving the boundaries, for example, for the CA2 region, which is detected as a thin zone of dark labelling, the cell layer, which is absent in CA2. Again, illustrating how the criteria for delineating are very important. So to conclude, we have now several volumetric 3-atlas templates that are available and that seem to be very useful in order to allow or to facilitate population of such atlas, which I think would be my suggestion for what is the most powerful user case to exemplify here, requires that workflows and tools are developed that allow anyone to register the data. And then it is important to realize that regardless of the registration method you use, it is, as far as I've seen, never very perfect. So the accuracy in different regions need to be measured and somehow shared when you share your data and the delineation criteria remain very, very important and they should be coupled to the atlas resource. Yes, then I'd like to thank all the people who have contributed, Daniel Wojcik and Piotr Majka at the Nenski Institute to reconstructing histology and of course, Meno Witter for his valuable input to delineating the hippocampus. Yes, thank you. So I think we have time for questions while Doug Bowden is coming up. The order was slightly changed from the original. So while Doug is getting ready, please use the microphone. Thank you. I was wondering if you had any thoughts about the tricky issue of how you represent variability between individuals in these atlas? I haven't talked about that here because the atlas we have are not probabilistic, but of course that is something that has to be taken into account as well. So I think what is needed is that the atlas templates continue to evolve and they should indeed incorporate individual variability and then comes aging as a parameter and of course, what also will be very important is, as Marion mentioned earlier, the different interpretations, the different delineations need to be combined so at least you can see how they vary. Did you find the need for non-linear mapping of the section on to the... The histological image to atlas mapping that we currently have been using is purely based on affine methods. But of course, non-linear methods will improve it. But again, probably not, depending on how much you work with it, it will probably not give you a perfect result, but of course it will reduce error and that will be very important. Do we have time? I'm sorry, do I have one more? Wait for the microphone. Like how many more years will you have to wait for any neuroanonymous or any scientist with a set of sections to be able to reference those sections to a template atlas? Are we there yet today? And if not... Well, in terms of the tools that we have, we are having something that allows us at least to sit down and register sections online and how long it will take until that can be made fully available. I wouldn't dare committing it to a given date, but it's moving along and it looks promising. So there is reason to be hopeful, at least. But I can't give you a specific date. Sorry.