fbpx
Bez kategorii

Kaggle RSNA – Intracranial Hemorrhage AI Challenge

By 06/12/2019 No Comments

BrainScan was founded in mid-2017. From the beginning, its founders were guided by the idea of ​​rapid implementation of machine learning methods, commonly referred to as artificial intelligence, in the field of medicine, and more explicitly supporting the analysis of medical imaging research. So, starting from 2018, BrainScan began work on a system that can automatically find pathological changes on head CT imaging. As part of cooperation with a team of radiologists, a list of 28 detected lesions was created. The system is not only able to mark the type of change, but also indicate the place of its occurrence.

 

The current management of the company is experienced managers – Robert Kitłowski and Marek Trojanowicz. The first is a graduate of the Warsaw School of Economics, he is responsible for commercialization, formal and legal matters, and the certification process, the second is a graduate of the IT department of the Gdańsk University of Technology, he is responsible for technology, acquisition and the process of data normalization for the needs of machine learning. Dariusz Wiśniewski, acting as the operational director, is responsible for representing the company at conferences, meetings, and events organized for start-ups, whose energy often allows to overcome seemingly insurmountable obstacles.

 

The critical year for BrainScan was 2019 when the company obtained funding for research and development works. The most important ones include funding under the Fast Track program implemented by the National Center for Research and Development for PLN 3 million and the EIT Health Headstart Funding Program for PLN 200,000. Golden. At the end of 2019, the Life Science Innovations Fund investment fund joined the company’s shareholders. Currently, the company has an amount of EUR 1 million, which over the years, 2020-2021, will be fully used for research and development. Also, Brainscan won the „Innolabs” accelerator award as one of the most promising start-ups in Europe.

 

Question: Did you win the 8th place in the competition announced by RSNA on the Kaggle platform?

 

Yes, in one of the most prestigious competitions in the world, we took 8th place and received a gold medal, which awarded all teams that fit in the first 12 – says Marek Trojanowicz. Kaggle is an online platform where machine learning competitions are organized, in which teams of specialists and scientists from around the world take part in trying to solve tasks in the field of data analysis and artificial intelligence. 1345 teams took part in the competition.

Question: What was the work like during these two months?

 

It was not easy – says Marek Trojanowicz – these were long hours often not only in front of a computer screen but also in front of a pen with a pen in hand. It should be emphasized that with such a large number of teams, the final result was determined by the details. Adam, Tomek, Maciek, and Szymon gave their best. We tested approaches based on the latest research in the field of machine learning – says Tomek Gilewicz, one of the specialists working on the BrainScan system. Some of the experiments were unsuccessful, but they enriched us with additional knowledge. Without many trials, we would not be able to reach the final solution.

Question: You made it to the final. What is it like to be at the conference organized by the RSNA (Radiological Society of North America), to be on the stage among the ten finalists?

 

This is an unforgettable feeling – comments Robert Kitłowski, CEO of BrainScan. We went to the conference in a two-person team – Adam Brzeski and I. I took place in the audience, and Adam was asked to present our approach and took the stage. Presenting the solution in just five minutes was a challenge – says Adam – but challenges are what we like the most. We received the official diploma – certificate from the RSNA organizer. Interest was very high – lots of questions, meetings, and new contacts that will pay off in the future – ends Robert.

 

Question: Large amounts of data are needed to teach neural networks. How do you handle data collection? What makes up the whole process?

 

The algorithm submitted for the competition was taught exclusively on the data provided by the organizers. However, we also have our own set of tests used to teach the algorithms that are the core of the BrainScan system – says Darek Wiśniewski, who is responsible for coordinating all activities related to obtaining head CT images. Accessing data is a lengthy process. From the first contact with a representative of a medical unit with data until the moment when anonymous imaging tests appear on BrainScan servers, it often takes up to 6 months. Why? This process is associated with the separation in the IT infrastructure of a temporary server unit to which data from the production server can be copied. At the same time, the process of signing contracts is underway, in which the critical issue is to get rid of any sensitive data from imaging tests, i.e., data enabling the identification of the patient or doctor describing the test – this is the process of anonymization. Ultimately, the data goes to our hard drives – presents Darek.

 

Question: What disk sizes are we talking about? How much space do the data you have?

 

We have over 100,000 imaging examinations in computed tomography of the head. Each study is saved in a special DICOM (Digital Imaging and Communications in Medicine) format. It contains all the results of digital imaging tests. DICOM is not only a format used to save image data, but also the entire standard, which is used, among others, to exchange images between various devices in medicine. One survey takes approximately 100 MB (megabytes) on average. Our collection of 100,000 tests takes about 40 TB (terabytes) on disks. We currently have a disk array with a capacity of 96 TB with the possibility of expansion if needed. In 2020, we would like to come to a collection of studies reaching 250,000 imaging studies – concludes the statement by Trojanowicz.

 

Question: How many computers – computing power do you need to bring BrainScan artificial intelligence to life?

 

The company participates in the „Nvidia Inception Program” program, created by one of the largest graphics processor manufacturers in the world, thanks to which the ongoing revolution in the field of artificial intelligence has become possible. Ten Titan Xp cards are the main driving force of our computing servers. Also, our arsenal has recently fed a new server, equipped with four Titan RTX cards. On the other hand, development computers of engineers are equipped with Nvidia GeForce GTX 1060 cards – ends Marek Trojanowicz.

 

Question: Can the BrainScan system be used for production? Is your product already certified?

 

We are still working on the development of algorithms for the classification of pathological changes in imaging studies of computed tomography of the head. We are currently acquiring more image databases that will open the way for us to add another pathological change recognized by the system – tumors. BrainScan plans its first implementations in mid-2020. To make them possible, we went through a painstaking certification process of the first iteration of the product, which allows the detection of two classes of pathological changes – strokes and bleeding. This is confirmed by the CE2A-PN-EN-ISO13485: 2016-04 certification document.

 

The place where the company – Tri-City, and Gdańsk is located is also not without significance – it is increasingly identified in the artificial intelligence basin in Poland, and we all also hope in Europe. Thanks to people like Professor Jacek Rumiński, who launched the „AI Bay – Artificial Intelligence Bay” initiative, Jacek Kawalc, who runs VoiceLab,  Łukasz Osowski, who after selling his first company IVO Software to Amazon, created a new Lab4Life project. Also, many other companies, such as AI Factory, or BrainScan – Gdańsk described in this article, attracts the best specialists in the field of artificial intelligence. Interestingly – it seems that all these people and companies have overcome the resistance usually found in Poland regarding joint action – they compete. However, because in most cases they specialize in various areas – they cooperate very much, exchange information – and this is very important – he says Marek Trojanowicz.

 

Materials and references:

 

RSNA press

https://press.rsna.org/timssnet/media/pressreleases/14_pr_target.cfm?ID=2146 

 

Kaggle leaderboard

https://www.kaggle.com/c/rsna-intracranial-hemorrhage-detection/leaderboard 

 

Photos

https://www.dropbox.com/sh/bya6bq1sp9mvmkd/AAD7lNx44x0USf6f5_NNG0h1a?dl=0 

 

Profesor Jacek Rumiński – Co to jest Sztuczna Inteligencja?

https://www.dropbox.com/s/z118pmfwc8eyljj/PNT_Gdynia_2019_AI.pdf?dl=0

 

List of NCBiR funding https://www.ncbr.gov.pl/fileadmin/POIR/4_1.1.1_2018/new/lista_rankingowa_I_runda_2018.pdf 

 

EIT Health

https://brainscan.ai/brainscan-got-eit-health-funding/ 

 

Innolabs

https://brainscan.ai/innolabs-final-event-we-won/