all issues mixed

6. All issues mixed: the less sense it makes – the better it sounds!

Interview statement:

“This extremely high resolution goes hand in hand with gigantic amounts of data. How do you intend to cope with these masses?

It’s a great challenge. At HIBALL, we talk about several petabytes. This can only be achieved by using completely new analytical methods – for example those that work with deep neural networks or machine learning – modern memory and communication technologies, and powerful computers. In addition, we want to analyse networks of nerve cells in the human brain and develop network models in order to deduce how the function of artificial neural networks can be improved.”

My Question 1: Where “several petabytes” came from?

My answer: Based on simple calculation, one-micrometer resolution data set should take approximately 180 TB. Calculations: 7400 files x 25 GB =~ 180,000 GB = 180 TB. Coincidentally, the biggest (AFAIK) hard drive available today has 18 TB capacity and costs between $600 and $800. So, storage of all files for 1 mk model will take 10 hard drives and cost between $6000 and $8000. Given all multi-million super-computers available in Juelich and multi-million dollar funding, this should not be a problem. Given many commercially available storage solutions (see, for example, Drobo), even a small laboratory can afford to have such a storage device. So, why data volume is such a big problem for sharing 1mk-resolution data, and where petabytes came from is not clear.

My Question 2: So, why is it such a problem to create 1-micron resolution model?

My answer: For some reason there is no clear explanation of this issue in the interview. Getting digitized image at 1 micrometer resolution takes time, but this is not a problem. Whole-slide scanners do this job rather well, and most of the time they work automatically day and night. Storage of thousands of such images takes a lot of space, but this is not a problem either. According to [3] there were at least two series of images of two brains created by 2016, and more I am sure, is done today. Still, after all efforts, the best resolution achieved with BigBrain-2 was 40 microns, not even 20. Why?

Actually, explanation of a problem come from Julich, and it is surprisingly simple (that is explanation which is simple, not the problem). The registration procedure, which is the essence of making the “model”, cannot use MRI picture a a reference for high-resolution images, because the resolution of available MRI image is too low. Accordingly, the alternative strategy should be used: registration with reference points, which have to be found in every image. However, the thickness of 20 microns section, used in the project, is relatively high compared to the size of an average neuron. The consequence is clear: finding the corresponding neuronal section on the next image becomes a very difficult task. The shape of the neuron changes from one section to another so significantly, that automatic algorithm used in the process can reliably locate very few of them (see image below).

Fig.6. Zoomed-In visualization of a cell that appears in 2 neighboring histological images and therefore is bisected. Source: M. Huysegoms personal page, INM-1.

Thus, it becomes evident that we need to try some other approaches to this problem, because throwing it blindly under the bus of Artificial Intelligence and Machine Learning would not help. One alternative possibility, which I personally tested, is to perform high-resolution registration by relatively small blocks (see, for example this demo on my YouTube channel). I presented these results to prof. K.Amunts and her colleagues as early as in 2015, but it was completely ignored.

<<Next Page>>