[Internship] NAIST – Imaging-based Computational Biomedicine


Hello dear Gentlemen,

Currently in a 3-months internship in Japan, you might have seen that I didn’t spend much time either on my website or my EasyLD project nor 3D project. As expected, 90% of my time is dedicated to my internship, especially during the last month (this one). Quite a huge report to write, but also try to get the best result I can have with the work I am currently in.

So, what’s exactly my internship?

Good question. First of all, this internship is “Confidential”. I will try to explain my work but I will then be limited by the amount of data I can shared.

Where is your internship?

Currently at NAIST, near Nara (or Osaka) in the Kansai area of Japan. NAIST is one of the best University of Japan. I am working in the Imaging-based Computational Biomedicine Laboratory (ICB Lab) (Official website).

What about the organization?

The starting hour was 9h, until 20h approximately, depending on what you have to do. And this, 5 day a week. You was able to take some pauses (for example to buy some foods, or do a jogging). Each other week, a meeting was planned to see how the work is going on and discussed about the new possibilities and the future work. Each Monday and Wednesday, a meeting with all the members (Master 1 as well) of 2 hours and half explained some classic papers of imaging-based algorithm as well as bio-medicine. Some times, we had some business trips, like the visit of an endoscopic surgery using 3D glasses. We were welcomed by a welcome party, and each Friday night, the master 1 and some of the staffs lead a Nabe party to share happiness and conviviality. The labs got its first foreign student with our arrival. Following by another French master 1 and a doctoral student from America. The discussion were mainly in English regarding the meetings, but outside of them, Japanese was the first language. So, I had to learn as fast as possible the remaining missing words! (I already started Japanese courses in France, so it was sufficient enough at the beginning, but not enough to read a newspaper).

What about your work?

I am working on the “Hierarchical Multi-atlas Method” created by Prof. Yokota. My goal is to expand his method and try to improve it as much as possible. I will not speak about how this method works but only on what I am working because of confidentiality purpose. This method is applied to the femus/hip area. The goal is to segment 20 muscles of the hip with as less as possible manual segmentation. For this, we get 20 CT data-sets with their ground-truth of the 20 muscles. Here is an example of a launch:

1) We want to segment the 20 muscles of the target. We do not have any information about the target, except the CT images.
2) We got a multi-atlas of 19 atlases with their ground-truth of the 20 muscles. We will try to registered each of them to the target.
3) We do some kind of stuffs (confidential) to get the 20 muscles of the target.

Hierarchical Multi-atlas method (Available on the icb lab website)

My first work was, obviously, to understand his method, try to launch it and get some results. Then update the CT data-sets because 2 years ago, only one CT data-sets got his 20 muscles segmented. I did some manual segmentation because some muscles (3) were missing, using “Vincent” a software by Fuji-film. First time for me, because my university courses are not directly related to medical imaging. So, a great new experience.
During this work, I was able to do a study about how each step improves the final accuracy. This was really interesting because we saw that the use of every data-sets doesn’t mean that we will get the best segmentation of the 20 muscles of the target. In fact, some atlases can be a bad representation and will not help to be more accurate. This will lead my future works.

My second work was to improve this method as best as possible regarding different possibilities. Firstly regarding the last stage with “How to select the best atlas” question. Because of I am working in a laboratory, this step was a research step with some directions like SFS (Sequential Forward Selection).
This work was directly followed by a big work on an use of SFS algorithm for the 2nd step. The work consisted in the programming of the algorithm, but also on how to improve it and a study about the efficiency. I had to compare the standard SFS they used with my own sort of SFS totally different from them. My SFS was based on the target to get a result directly linked to the target when we are computing the similarities. Their SFS was about guessing a best combination of atlases inside the multi-atlas, without the target. My thought was this will not help enough to get the best final segmentation because we chose the best combination of atlases for the multi-atlas and not for the target. My SFS was a choice of a best combination for the target. The main idea was to combined both SFS to get a best segmentation. This is where I am currently working on.

The second work as you see seems to be quite long, especially because of “Research”. We don’t know which algorithm will be the best, and we have to do quite a lot of study, and experiments some new ways. Seeing the remaining time (3 weeks), I will be in a rush to end this as best as possible, and writing my report as well. I am really passionate by this subject, so I am feeling exciting by the result I will get, but also stressed because of the remaining time. I am quite sad to not be able to explain my research more than that because of confidentiality, but I hope this document showed you enough on what I am involved, and what kind of stuff I am doing.

The research domain is finally something that I can really think about it but I will wait next year to really decide where I will go. I really like to develop games, new stories, new way of thinking and new game-play as well that working on the medical field cannot be the best idea, even though there is innovation. I hope my next internship will be in the game development field! But before this, I will create another games for my next LD! And guess what? September will be a programing month, and I hope this month will not be a solo programing month ;) Maybe a cooperation with friends to sell a game at this end of the month (or two month? I don’t know, depending on the idea (I already have quite a lot)). Why selling? To see if I am able to end a game. Because every time I create a game,I stop after some times or don’t try to improve it. I have to push myself ahead because next year is my last year!

Moreover, I am now able to speak as well as English than Japanese even though my Japanese remains a little bit too low. Currently working on improving my basics of German because of shame. I hope I will be able to speak in German as good as my current level of Japanese and I hope that my level in Japanese will catch my level in English a day! Fluency, fluency everywhere! (English, Japanese, German and French, let’s try Korean!)