Experts in a medical challenge with AI developed at the JKU had no chance.
Medical Competition: AI developed at the JKU beat human experts hands down
Artificial Intelligence developed at the Johannes Kepler University Linz competed with humans. The goal was to recognize protein in cells. The results: The program developed at the JKU beat the experts hands down.
Artificial intelligence has long beaten human champions in the game of chess. When it comes to the everyday practice of medicine, AI is becoming increasingly indispensable. The objective at the competition was not randomly selected. Elisabeth Rumetshofer, MSc (Institute for Machine Learning, dept. head Prof. Sepp Hochreiter) explained, “In order to better understand biological processes, it is important to know where a cell in protein is located. This is an important indication of a protein’s function and helps us better understand diseases.”
The AI learned to recognize patterns with the help of thousands of sample images. This information helped the AI to allocate proteins in new images. Elisabeth Rumetshofer added, “The hardest part was teaching the AI what cell structures look like.” The algorithms require tens of thousands of images to gradually learn how to recognize a cell. Humans have the advantage in that our brains can recognize clear patterns after just looking at a few images.
“We were pretty excited and expected a neck-in-neck competition.” said the JKU researcher.
Specially-Trained Human Opponents
The AI developed at the JKU competed against three human opponents: other object identification AI, a group of 3 experts, and a group of 25 Life Sciences students. They were all given special pre-training in order to really challenge the AI program. On competition day, all groups were asked to analyze 200 images that neither the human teams nor the machine had seen before. Elisabeth Rumetshofer remarked, “The high resolution and very fine cell structures were particularly difficult for the AI.”
Game, Set and Match for the AI
It took human experts from the MedUni in Vienna and Kepler University Hospital in Linz about 5 hours to solve the task. The AI analysis was completed in 26 seconds. And while the best expert was able to allocate 72% of the proteins, AI at the JKU got 91% correct – a clear difference.
“In practice, AI is relevant in several areas, including pathobiology when it comes to recognizing genetic mutations but also in drug development when it comes to examining where and how proteins actually work.”
AI in the Starting Gates
The AI developed at the JKU would actually be ready to apply to medical procedures today. “The obstacle lies in the different biotechnologies.” Rumetshofer explained. The problem would be solved by having uniform standards. “If necessary, the AI would just have to learn different parameters.”
In the medium-term, the AI technology developed at the JKU will be an integral, indispensable part of medical research.
Photo credit: JKU
Photo 1: JKU Researchers
Photo 2 and 3: Cells
Institute for Machine Learning
Ph.: 0732 2468 4537