Ex vivo MRI scanning (a.k.a. scanning dead brains)

Imagine you were taking a photograph in a room with very dim lightning. You could still take a picture if you use a very long exposure: by keeping the shutter open for a long time, you can collect enough light to create your image. A problem with this approach is that it only works if the object we’re imaging remains completely still. If the object (or person) or the camera moves, we obtain a blurry picture.

A blurry photograph

A similar thing happens with MRI. Obtaining MRI images al ultra-high resolution requires very long scanning times (tens of hours), so it is inevitable that the subject being scanned moves during that period of time. For that reason, practical MRI scanning protocols are usually shorter than 10 minutes. However, if we want to build accurate, high resolution models of brain anatomy, there is a way of overcoming the restrictions imposed by subject motion: using ex vivo brains from cadavers. 

The idea is as simple as this: dead brains don’t move, so we can scan them as long as we want without motion artifacts. Unfortunately, scanning ex vivo brains the same way we scan in vivo (meaning, living brains in living people) does not work. The reason is that the fixation process (immersion in formalin for preservation of the sample) changes the magnetic properties of the tissue. Moreover, the formalin and air bubbles introduce image artifacts that degrade the quality of the scan. 

An ex vivo brain scanned in formalin. The red arrows point at artifacts created by air bubbles

What can we do to fix this? One way is to replace the formalin by a fluid that is transparent to MRI. Since MRI is based on detecting and measuring protons, we can use a proton-free fluid. Many studies use a lubricant called Fomblin. We use a cheaper (and less slippery) alternative called Fluorinert. 

Slice of a ex vivo brain scanned in Fluorinert

Once we have fixed the problem with the artifacts, we have yet another problem. The hardware of clinical scanners like the one we have at the BCBL is not meant to acquire images at ultra-high resolution. For this reason, when we try to acquire a big 3D scan, the machine runs out of memory during the reconstruction (the process of transforming MRI measurements into images). This problem can be overcome by acquiring different parts of the image separately and then stitching them together. The strategy is normally to acquire several slabs that can subsequently be stacked to create our final scan. 

Stacking slabs to create a MRI scan

 Stacking slabs enables us to bypass the memory limitations of the scanner, but it introduces yet another artifact: the Venetian blind. This is because the sensitivity of the scanner is not uniform across each slab; instead, it is lower in the first and last couple of slices of each slab. If we acquire the whole 3D scan in one shot, this is not a problem, because the first and last slabs do not cover the brain anyway. But when we stack slabs, we get patterns that resemble a Venetian blind.

Venetian blind artifact

As bad as this might look, there is a bunch of image analysis algorithms you can use to correct for this artifact. Here I’ll show you the output from an algorithm that we have developed ourselves, and which also corrects for intensity inhomogeneities (more on that another day; for now, let’s just say that it’s an artifact that makes some regions brighter than others).

Before (left) and after correction with our method (right)

These are pretty pictures, right? Let’s look at a close-up of the hippocampus, and compare with a standard resolution in vivo scan.

Left: standard resolution (1 mm). Right: ex vivo scan (0.25 mm)

What will we do with these beautiful scans? That remains for a future post ;-)

The THALAMODEL project: visiting Dr. Insausti

A couple of days ago, we went to Albacete to visit Dr. Ricardo Insausti, full professor in human anatomy and embryology at the University of Castilla - La Mancha (UCLM). Dr. Insausti is going to be an instrumental part of the THALAMODEL project. He is going to find the donors whose brains we will use to build our models; and then he is going to extract the brains and carry out their fixation with formalin.

Brain fixation is critical to study the brain ex vivo (meaning, outside a living body). If a brain sample is not fixated soon after death, it deteriorates quickly due to the blood supply. After the fixation, the brain can be preserved for a long time.

Dr. Insausti, holding a fixed human brain

 Dr. Insausti is also going to carry the histological study. Such study consists of two different phases: slicing the brain, and staining the slices. In order to slice the brain (or in our case, a block of tissue around the thalamus) into very thin sections, one first freezes the sample with dry ice and then sections it with a machine called "microtome". As opposed to thicker blocks of tissue, these thin slices can be examined under the microscope.

Rather than looking at the slices directly, one enriches their contrast first through a staining process. Different types of stains and dyes can be used to enhance different properties of the tissue. The most popular technique is arguably Nissl staining, invented by Franz Nissl in the late 19th century. After staining, the samples are mounted on slides that protect them, and can then be examined with a microscope.

Nissl-stained slice of human thalamus

Looking at mounted slides with samples of a human hippocampus

 Finally, Dr. Insausti will use the stained slices to manually trace the boundaries of the thalamic nuclei. This information will be critical for us to build the thalamic atlas which is at the core of our project. How we go from slices of the thalamus to a 3D atlas of the thalamic nuclei will be discussed in future posts.

The THALAMODEL project: thalamus and dyslexia

Location of the thalamus (in red)

I have recently been awarded a Marie Skłodowska-Curie individual fellowship to build a probabilistic atlas of the thalamus, as well as a set of tools that enable us to use the atlas to automatically analyze the thalamic nuclei brain MRI scans from neuroimaging studies. The title of the project is “Multimodal, high-resolution modeling of the thalamus for neuroimaging studies: application to dyslexia”, and the funding is from the European Union’s Horizon 2020 research and innovation program under the Marie Sklodowska-Curie grant agreement number 654911.

Why the thalamus? Located between the cortex and the midbrain, this cerebral structure is the main relay station of the brain: it is connected by fibers to virtually the entire cerebral cortex. In addition, the thalamus is related to the regulation of consciousness and sleep, and it is also closely related to language. The relationship with language has been proven in case studies involving thalamic lesions and electrical stimulation of the thalamus. However, it is still unclear whether this relationship is due to the connection between the thalamus with cortical regions related to language, or due to its involvement in the integration of language functions via memory.

Moreover, the thalamus has been linked to some of the most common language disorders, including dyslexia, a condition in which individuals with normal intelligence have serious problems learning to read. Dyslexia is the most common neurobehavioural disorder in children, on which it has a terrible impact: it causes severe disadvantages in school development, education and self-esteem that greatly increases the risk of social marginalization later in their lives.

This project aims for building a detailed atlas of the thalamic nuclei using multimodal imaging data from autopsy brain samples, and for creating companion image analysis tools that can use the atlas to analyze the thalamic nuclei in in vivo brain MRI data (that is, from living people) from neuroimaging studies. The project consists of four major steps. First, we plan to acquire ultra-high resolution MRI data of the autopsy samples. Since these samples don’t move (for obvious reasons), we can make the MRI acquisition very long therefore with very high resolution. Second, we will slice the brains to perform a microscopic anatomical study (known as “histology”) of the samples. The ultra-high resolution MRI and the histology images will allow us to build a thalamic atlas with very high level of detail. Third, we will create the tools that enable us to use the atlas in the analysis of in vivo data. And fourth, we will make the tools publicly available and apply them to a dyslexia study at the BCBL. Hopefully, the new tools will allow us to better understand this disorder, while enabling researchers at other institutions to improve their understanding of this cerebral structure by carrying out analyses at a higher level of detail - compared with current tools.

I will regularly post here with the progress of the project; stay tuned!

High-resolution brain imaging for Alzheimer's disease studies

Disclaimer: this post is largely based on an article I wrote for "The Global Scientist", which you can find here.

Figure 1: location of the hippocampae (in red) in the human brain.

The hippocampus is a brain component that plays a very important role in memory. We have two of them – as shown in Figure 1; one on the left side of the brain, and another on the right side. The hippocampus is a structure of much interest for the neuroscience community, partly due to its connection to Alzheimer’s disease. Alzheimer’s is the most common form of dementia, a disorder that is expected to affect over half a billion people in the next 40 years, and whose estimated total cost was over a hundred billion euros in 2010 in Europe alone.


Hippocampus means “seahorse” in Greek, which is not surprising if we look at its shape. 

Figure 2: photographs of a human hippocampus (left) and a seahorse (right), illustrating their similarity. The plane overlaid on the left-hand image illustrates the orientation of Figure 3.

Even though it is hard to tell from Figure 2, the hippocampus consists of a number of distinct, interacting subregions, called the hippocampal subfields. If we slice the sample in Figure 2 in the direction of the red plane, it turns out that the hippocampus looks pretty much like a Swiss roll, as shown in Figure 3. If we follow the subfields from the outside to the center, we first have the subiculum, where the hippocampus connects with the brain cortex, then the CA fields – 1 through 4, where “Cornu ammonis” stands for the horn of Amun, the Egyptian god – and finally the dentate gyrus.

Figure 3: analogy between a Swiss roll (left) and the internal structure of the hippocampus (right). These pictures illustrate a cross-section indicated by the red plane in Figure 2. The location of the subfields is indicated in the sketch on the right.

Different studies [1,2] have shown that these subfields are involved in different memory subsystems, and that they interact with different regions of the cortex to make the brain work. Moreover, autopsy studies have shown that normal aging and Alzheimer’s disease affect these subfields very differently: whereas normal aging introduces just minor changes in the hippocampal structure, Alzheimer’s disease produces loss of volume (atrophy) in the hippocampus, with different subfields being affected to different extents.


Magnetic resonance imaging (MRI) is an innocuous technique that allows us to study the human hippocampus in vivo, that is, of living persons rather than autopsy samples. MRI allows us to track the morphology of the brain through normal aging or the development of neurological diseases. Unfortunately, the resolution of MRI scans has traditionally been insufficient to identify the subfields in the MRI images (see the image to the left of Figure 4 below). This eliminated the possibility of studying the individual subfields, forcing researchers to carry out studies of the hippocampus as a whole and to disregard valuable information on localized volume changes. However, with progress in MRI technology we are beginning to recognize subfields of the hippocampus, as seen on the right of Figure 4.

Figure 4: cross-section of the hippocampus using MRI imaging. Left: 1 mm resolution (standard). Right: 0.4 mm resolution (high). Certain subfields are marked to allow them to be related to the sketch in Figure 3.

Pictures like the one on the right will allow us to localize the effect of Alzheimer’s and better characterize the progression of the disease. The eventual goal is to find a link between the early stages of Alzheimer’s and the size of subfields that are measured with MRI. This would potentially allow us to recognize the onset of Alzheimer’s and begin disease-modifying treatment before the disease has caused too much damage to the brain. Moreover, being able to visualize the subfields in vivo would also allow us to determine the effectiveness of treatment. Currently, clinical trials rely largely on neuro-psychological scores, that is, the results from interviews and pencil and paper tests. In-vivo imaging of the subfields would provide much more reliable metrics of success for drugs being tested and potentially decrease the time it takes to bring such drugs to market.

Figure 5: cross-section of a high-resolution brain MRI scan, showing both hippocampae (top) and subfieldsthat have been delineated by hand (bottom). The MRI data used to generate this image are described in [3].

One remaining challenge in carrying out this type of analysis is the extraction of information from the images. A simple, albeit time-consuming, method is for an expert neuroanatomist to manually label the subfields on the image. He or she would have to trace the boundaries of each subfield on each slice, as in Figure 5. Due to the 3D nature of MRI, where many 2D slices need to be analyzed, human inspection of the images is time-consuming. For instance, in the MRI scan of Figure 5, an expert would have to examine the hippocampus in approximately 150 cross-sectional slices in order to cover its anatomy from head to tail, a process that could take up to a full week for a single subject. Since a study typically consists of tens of subjects, manual analysis is limited to a handful of very specialized sites that have the expertise and manpower to carry out such an excruciating task. Even at these sites, the number of subjects in studies must be kept relatively low to maintain reasonable times.


An alternative means of analyzing MRI images is to use computer programs that extract the hippocampal subfields automatically from the images. Even though automated annotations made by such programs are less reliable than those made by human experts, they allow us to carry out experiments at a much larger scale, which greatly increases our ability to detect the effects of Alzheimer’s or aging within large sample sizes. How do such programs work? I'll leave that for a future post, but designing such software is the focus of my research. The curious (or impatient) reader can read references [4] and  [5].


In the next 2-3 years, a vast number of high-resolution hippocampal MRI scans will have been collected. The enhanced resolution of these scans, together with newly developed image analysis software, will allow us to link morphological changes in the hippocampus to the onset of Alzheimer’s disease, which will be a great step towards combating this disease.


References

J.D.E. Gabrieli, J.B. Brewer, J.E. Desmond and G.H. Glover. Separate Neural Bases of Two Fundamental Memory Processes in the Human Medial Temporal Lobe. Science, 276:264–266, 1997.

M.M. Zeineh, S.A. Engel, P.M. Thompson and S.Y. Bookheimer. Dynamics of the Hippocampus During Encoding and Retrieval of Face-Name Pairs. Science, 299:577–580, 2003.

J.L. Winterburn, J.C. Pruessner, S. Chavez, M.M. Schira, N.J. Lobaugh, A.N. Voineskos and M.M. Chakravarty. A novel in vivo atlas of the human hippocampal subfields using high-resolution 3T magnetic resonance imaging. Neuroimage, 74:254-265, 2013.

K. Van Leemput, A. Bakkour, T. Benner, G. Wiggins, L.L. Wald, J. Augustinack, B.C. Dickerson, P. Golland and B. Fischl. Automated segmentation of hippocampal subfields from ultra-high resolution in vivo MRI. Hippocampus, 19:549–557, 2009.

P.A. Yushkevich, H. Wang, J. Pluta, S.R. Das, C. Craige, B.B. Avants, M.W. Weiner, and S. Mueller. Nearly Automatic Segmentation of Hippocampal Subfields in In Vivo Focal T2-Weighted MRI. Neuroimage, 53:1208-1224, 2010.

Why this blog

My family and friends know that I'm a researcher. Most of them know that I do something related to medical images. Few of them know that currently I'm focused on studies of the hippocampus. And pretty much no one, other than my boss and a couple of coworkers, really knows what I do.

This wouldn't really be a problem if it wasn't because... well, because these people support me through their tax dollars. And it's kind of hard for us scientists to make a point of why they should keep providing for our daily bread when they have no idea of what we do and of how it can impact their lives.

A big part of the scientific community (definitely including myself) has traditionally struggled communicating their science. So, "better late than never", here's my humble contribution to ameliorating this problem. In this blog, I'll try to post about challenges, debates and advances in the field of Medical Image Analysis. Being a community with thousands of researchers around the World, it shouldn't be hard to find cool stuff to post about here and there. I will also try to post a bit about my own research, for the sake of self-advertising ;-)

Stay tuned!

Greetings from Boston