Medical Imaging: EPITA research rewarded in China!
In October, an EPITA delegation traveled to Shenzhen China to participate in the 2019 edition of MICCAI, the leading international conference on medical image processing. The event also hosted a competition dedicated to researchers: the Left Ventricle Full quantification challenge (or LVQuan) allowed Elodie Puybareau, co-manager of the IMAGE major and member of LRDE, as well as Zhou Zhao, her doctoral student, to finish third place, hence rewarding EPITA research in medical imaging related to the heart. Élodie Puybareau speaks about her experience, which highlights EPITA’s increasing involvement in the health sector.
Zhou Zhao and Élodie Puybareau
Why is it important to participate in this type of competition?
Élodie Puybareau: This allows us to verify the methods we are developing in the laboratory and see where we rank in terms of other research teams around the world working on the same subject matter. The advantage of a competition is to obtain a direct assessment regarding our methods.
How was the LVQuan challenge?
Generally speaking, this kind of competition takes place in the following manner: the organizers give participants a batch of images to process, while indicating the specific problem. For LVQuan, the problem was to find a particular heart structure, the left ventricle. The idea was to identify the ventricle in the images of the heart provided by the organizers. We began by developing a method that allowed us to find this structure and take measurements, for example, of its thickness or how it changes over time. Then, in these types of competitions, the organizers ask that participants directly submit the codes used for their working methods or they send out new images to process that must then be returned, so that they can evaluate the work.
Finally, how long does a competition such as LVQuan last?
Several months! For example, we received the images to be processed in April for LVQuan. Most of the time, the competitions end between July and September. The organizers contacted us after the competition and asked that we present our method orally. In general, when the judges ask that you give a presentation, this often means that you have very good results or that the method used is very interesting. However, we only learned how we ranked on October 13, the first day of the Conference.
>How did you react when you learned that you won third place?
I was particularly pleased, especially for my doctoral student Zhou Zhao who worked extensively on this research. He has been a doctoral student at EPITA for a year and comes from China. This was his very first conference and presentation. He was a little stressed in the beginning but did an excellent job. At the start of the competition, there were approximately 30 participants in all. Subsequently, the number of participants dropped, with only ten who submitted a scientific publication. Some competitions, such as LVQuan, require participants to write a research article in addition to submitting results. Finally, only four publications (including that from EPITA, called “A Two-Stages Temporal-Like Fully Convolutional Network Framework for Left Ventricle Segmentation”) were accepted to be published in the MICCAI Conference proceedings via a specialized workshop on cardiac imaging.
Did the third place award that you received for this publication allow EPITA researchers to gain recognition on an international level?
Absolutely! Simply participating and being well ranked in a MICCAI competition is enough to show that you produce high quality work, and the fact that our scientific article was published led to additional visibility.
What technology did you use in your methods?
We used deep learning methods. Simply put, deep learning means teaching a computer or program to find a solution for us. You tell the program “here is an image”, then “here is the result that I wish to obtain” and finally “it is up to you to find the solution”. Behind all of this are major mathematical models that help determine the solutions. And today, in medical imaging, the majority of the methods that work best are based on deep learning.
For example, how many images and hours of work were needed to develop the method rewarded by MICCAI 2019?
There were not so many – we were given a database with 50 images. However, we spent endless hours working on the project: Zhou Zhao worked on the project full-time from April to July. It required a great deal of research, with extensive tests and experiments before we were able to find the best solution. Training a computer takes time. When we enter new information into the computer, it will work alone and sometimes this can take several days!
How long has EPITA been developing medical imaging?
In fact, EPITA has always worked on medical imaging, but the majority of research has been focused on the brain, particularly through partnerships with Telecom ParisTech. When I arrived at LRDE, with my doctoral work and passion for medical imaging, I tried to “contaminate” more researchers in the laboratory to add greater dynamics to the team. We started participating in the MICCAI competitions 3 years ago, but our diversification really began last year, with a competition that allowed us to determine whether our methods could also work with images of the heart. It turned out very well – we won 3rd place – and this reinforced our thinking that our potential thesis subjects on the heart were extremely legitimate. This is how Zhou Zhao began working on his thesis.
Is this why the school recently created the IMAGE major?
Yes, but not only! With Guillaume Tochon (the other manager of the Major), we have chosen more general subjects for this major, not simply medical. The image processing component of LRDE motivated us the most. However, given that EPITA has participated in the MICCAI competition for the last three years and always obtains good results, we also felt it was essential to offer courses related to medical imaging. For example, employees from General Electric Healthcare will give classes on the subject this semester to our 3rd year students.