As mentioned in the previous newsletter, MI4People has participated at the TUM.ai AI4SocialGood Makeathon that took please from 22nd to 24th of May. We were contributing to this great event as one of the challenge setters in the MedTech track and it was a highly rewarding experience for us from multiple angles! Therefore, we have decided to dedicate this newsletter exclusively to this event.
Enjoy the newsletter, put your capabilities to help Public Good delivery into action, and let us together make the world a better place for all of us!
Your MI4People Team
TUM.ai Makeathon is an event that is organized twice a year by TUM.ai – one of Europe's largest AI-based student initiatives. During this event students and young professionals from various domains have a chance to compete on real-life AI challenges that are provided by the 'challenge setters' from large enterprises, startups, VCs, governmental organizations, and NPOs. It is a great chance for students to prove their skills, get new connections, compete on real-life problems, get attention of potential employers and VCs, and – in case they win the Makeathon – secure first seed capital for their ideas and solutions.
This year the event was focused on use cases from the area of AI4SocialGood and got a great attention from the community. Around 400 students and young professionals applied for the Makeathon from which more than 250 were selected as participants. They came from 57 countries and 6 continents - the event was truly international! There was a total of 9 challenges set by the challenge setters spread across 3 tracks: AI for Healthcare (MedTech), AI for Education, and AI for Environment. 38 teams of ca. 5 people each have created working solutions – some of them have defined their own AI4SocialGood challenges – and presented them to a jury consisted of many industry leaders and domain experts.
TUM.ai had also requested MI4People to set a challenge as well as be part of the jury, and we were happy to be able to contribute to this great event.
MI4People’s Makeathon Challenge
The MI4People’s challenge was focused on identifying diseases from X-ray images of chest region of humans. This challenge was a prologue for our second big project “General Computer Vision for Healthcare in Developing World” that will be kicked-off on 21st of May. This project aims at supporting primary care physicians in poorer developing countries by providing them with a free Computer Vision System for diagnosis of diseases and injuries using various medical images, e.g., photos, X-ray images, CT and MRI scans.
For this challenge we decided to use a very prominent and easy-to-access open source ChestX-ray dataset from NIH Clinical Center (original paper) that contains fourteen typical diseases detectable from chest region X-rays.
This Makeathon challenge can also be viewed as an important first step for MI4People's General Computer Vision for Healthcare in Developing World project because we could identify techniques that might be helpful to tackle many challenges that our future AI system will likely be confronted with in developing countries, e.g.:
Poor access to existing medical IT tools / less available computing power / restricted access to the Internet and cloud computing / old smartphones
Old analogue X-ray devices, so that the quality of images might be much poorer than what is in the machine learning training data set
Digitization of analogue X-ray images, e.g., using smartphone cameras, and how different photo quality and light conditions can be considered
Challenges of integrating the MI4People’s Computer Vision system into existing processes in the clinics
The accuracy of the created underlying Computer Vision model was, of course, also relevant, but the focus of the challenge was more on how to apply our future system in practice.
Since our Makeathon challenge was in fact one of the most difficult one, both from technical as well as from domain expertise points of view, we decided to award our own special prizes in form of cool goodie bags to the best performing team in our challenge.
Solutions for our Challenge
Four teams decided to tackle our challenge and every one of them has provided valuable insights that will help us in our further research. In fact, all solutions were so great that it was very difficult to choose the winning team. Below, we mention some of the highlight insights from each team.
Has experimented a lot with balancing our very disbalanced data and did a great job in testing multi-stage classifiers. It means that they first built an AI model that distinguish between “no disease” and “there is a disease” and then started to build the second classifier for only the second category to identify a particular disease.
Has tested some promising AI model candidates that were developed specially for medical images like TorchXRayVision and PYLON. In our MI4People research project we will also be evaluating these models. Further, the team has discussed the option of connecting our future system with some sort of Clinical Decision Support (CDS) in order to integrate it into typical clinical workflows.
Has created a great working MVP (video demo, open-source code) and also did a great job on data augmentation to simulate bad image quality and on creation of light-weighted models, so, that their solution is very suitable for the special conditions in the developing world. In fact, dAIgnose has really impressed us and was very close to win our challenge.
DS@LMU – the winners!
DS@LMU (Data Science at Ludwig-Maximilians-Universität München) has provided the most balanced solution from our point of view (pitch deck, open-source code). It has incorporated balancing the data, data augmentation, light-weighted models and also considered usability and integration into clinical workflows. The team has also shown very good understanding of the underlying problem and could define a sound roadmap for the development of the product.
Besides the actual challenges the TUM.ai Makeathon has also provided us a great platform to spread the idea of the Machine Intelligence for Public Good and create new connections. So, we could win several new volunteers who all will contribute to our project General Computer Vision for Healthcare in Developing World and are in touch with many further potential candidates. We could also meet several students from TUM.ai who are interested in creating socially oriented and responsible AI4Good startups, non-profits, or open-source projects and are discussing with them how we could cooperate. Further, we got connected with InnovationLab of City of Munich and Bavarian Federation of the Blind and Visually Impaired and are discussing how Machine Intelligence (MI) could help track environmental situation in cities in a better and more transparent way for the citizens, and how MI could support visually impaired and blind people.
The TUM.ai Makeathon was a big success and a major step forward for MI4People! We would like to say thank you to all the TUM.ai-students for the organization of this very important event and for helping us spread the idea of Machine Intelligence for Public Good across students and other organizations. It was a great pleasure for us to work with TUM.ai and we are looking forward to further cooperation!