08.12.2017 change 08.12.2017

Artificial intelligence will recognize bacteria by their appearance

Perhaps someday an image from an optical microscope will suffice for artificial intelligence to identify the species of bacteria. The image shows the bacteria analysed in the research of scientists from the Jagiellonian University and Cracow University of Technology. Each species is illustrated by three examples. Source: PLOS One, Bartosz Zieliński et al. Perhaps someday an image from an optical microscope will suffice for artificial intelligence to identify the species of bacteria. The image shows the bacteria analysed in the research of scientists from the Jagiellonian University and Cracow University of Technology. Each species is illustrated by three examples. Source: PLOS One, Bartosz Zieliński et al.

An image from an optical microscope will suffice to identify the species of bacteria. And very quickly - in a few seconds - believe scientists from Kraków. They employed artificial intelligence algorithms to identify the bacteria.

A patient comes to the gynaecologist with symptoms of genital tract infection. The doctor takes a swab - and instead of sending it to the laboratory, where the analysis will take up to several days - takes photos of the sample under an optical microscope, and then uploads them to the program. In a few seconds the program returns gets the identification result and the doctor knows what microorganisms he is dealing with. He can suggest treatment during the same consultation.

This is what microbiological diagnostics could look like if the ideas of researchers from the Jagiellonian University and the Cracow University of Technology could be realised. Scientists want bacteria to be identified by their appearance thanks to deep learning algorithms. Their research appeared in September in the journal PLOS ONE.

LONG ROAD TO DIAGNOSIS

Today, the process is much longer. Co-author of the study, Dr. Monika Brzychczy-Włoch from the Chair of Microbiology at the Jagiellonian University Medical College described that a sample taken from a patient - such as a swab from a skin lesion - is cultured (microorganisms are multiplied). Then pathogenic strains are isolated. From them, in turn, a Gram stain preparation is made, which a laboratory diagnostician evaluates with an optical microscope. This is just the first step in identifying the species of isolated bacteria.

Today the evaluation of preparation by a diagnostician takes about a dozen or so minutes, and the information thus obtained is very limited. Only two characteristics are determined: the shape of the microorganism (for example its spherical or cylindrical form), as well as the type of staining - whether it is a Gram-positive or Gram-negative bacteria (they are stained differently). "This is where the evaluation of the preparation ends. We are not able to determine the species or even the type of isolated bacteria" - said Monika Brzychczy-Włoch.

Meanwhile, a computer algorithm from a microscopic picture of bacteria may obtain much more information than a man watching the same image. However, the computer must be "taught" to distinguish between bacteria.

A NETWORK THAT LEARNS FROM ITS OWN ERRORS

"We use the latest image analysis mechanisms - deep convolutional networks. Such networks require a lot of images so that they can learn to solve a given problem" - sais Dr. Bartosz Zieliński from the Faculty of Mathematics and Computer Science of the Jagiellonian University. According to Zieliński, such a network would need about a million images to accurately identify bacterial species. "And we had only 660 photos - 20 for each of the 33 analysed species" - said Dr. Krzysztof Misztal from from the Faculty of Mathematics and Computer Science of the Jagiellonian University.

But the researchers made a clever move. In their research, they used a network that was already preliminarily trained to recognize images. The network was previously trained with pictures of objects - such as flowers or cars. "So the network was able to extract certain features necessary for the classification of images. Having such a network, we were able to train it to recognize images of a new type - bacteria. For the time being this is the only possible approach due to the fact that obtaining a million images of bacteria with confirmed identification is currently beyond our reach" - said Dr. Zieliński.

It worked. The algorithm distinguishes even between very similar bacterial species - ones that a human would not be able to tell apart based on microscope images.

PLEASE FEED THE ALGORITHM!

The network currently recognizes 33 species, including aerobic and anaerobic bacteria, as well as individual species of yeast-like fungi. However, this is only an introduction to further research. "It all depends on the databases that we work on. If the collection will be large, the network will be wrong very rarely, even with an increased number of analysed species" - said Dr. Krzysztof Misztal.

The algorithm is now able to determine the species when analysing an image containing only one type of microorganism. The stages of multiplication and isolation are therefore still necessary. The next step, however, will be to teach the algorithm to recognise bacteria much faster - directly in the material collected from the patient. According to scientists from the Jagiellonian University, it is feasible. However, it requires expanding the database and further work on the proposed solution.

Dr Bartosz Zieliński added: "For now there is no such a global image database of different species of bacteria. Our dream would be to create an internet system that would allow teams from around the world to send us images of microorganisms, and we would classify them. We would then use these images to further teach our network" - said Dr. Zieliński. Researchers from the Jagiellonian University want to fight for a research grant that would allow them to collect a large number of images and improve the proposed algorithm.

Author: Ludwika Tomala

PAP - Science in Poland

lt/ zan/ kap/

tr. RL

Copyright © Foundation PAP 2019