[NOT WEBRTC] – Making a difference – Deep Learning and Cancer

For once, a blog post about something else than webRTC.

In a previous life, I was a scientist working mainly with biologists and medical doctors on finding ways to cure cancer. I’m actually still active on the subject and am helping a small start up in France developing corresponding software, more oriented to cancer drug development (NovaDiscovery). One of the 2 latest PhD candidates I helped supervise, humayun Irshad, while developing a webRTC platform, was focussing on automated diagnostic of breast cancer.

I helped him first on automatically evaluating the aspect (discrete geometry) of the cells, which in turn allowed for detection of  the mitosis of the cell (cancerous cell never stop duplicating …). Eventually we worked on the color (multispectral) analysis of the images obtained from staining the biological samples with some markers (Histo-Pathology). Microscope can acquire different color spectrums, and by choosing the right ones as input, the detection algorithm was providing much better results. At that stage, we could detect cancerous cells in 76% of the cases. 

My colleague Ludovic Roux was helping him as well on the mitosis detection, and worked a lot on an open scientific challenge which opposed in 2012 public labs and private labs from all around the world on that breast cancer detection. The IPAL team, finished 2nd, with a detection rate of (if I remember correctly) 78%+.

Ludovic and myself were also trying to teach the future researcher proper software development process (source code repository, tags, branches, automated compilation, cross-OS compilation, automated testing, ….). For that part, we’ve been less successful 🙂

He eventually graduated and went on to move to Harvard Medical., joining a deep learning team that recently developed a second version of  algorithm showing up to 99+ % true tumor detection rate! Actually, it only reaches 92%, so using deep learning, the same form of artificial intelligence used by google for beating humans at GO, you gain an extra 14% (don’t believe that it s all due to AI, there was 6 years of research before that). It’s still less accurate in a way than humans who can reach 96%.

Where it becomes remarkable, is that the algorithm will do mistakes that the humans will almost never do, and vice versa, meaning that, the human can detect the remaining errors the algorithms didn’t (and vice-versa), leading to 99+ detection! Of course, the professors leading the labs have right away set up a start up to commercialize the results (pathAI.com).

I wanted to write a blog post to congratulate him. A PhD can be a cruel and demanding experience. His was more than others, with visa difficulties, travel challenges, and all the life trials like growing a family, not being a student anymore but not being employed just yet. Bravo! Ludovic, Daniel and I are immensely proud of your hard work, and hope we have helped make a difference.

Leave a Reply

Your email address will not be published.

Time limit is exhausted. Please reload CAPTCHA.