Digital technology may make microscopes obsolete
The second step involved figuring out how to get the programme to identify the things experts would be looking for in the images. In order to accomplish this, the researchers had to put the programme through a learning phase. “At first the programme has no way of classifying or segmenting images. So we asked the experts to show us cells representing the different categories they wanted to distinguish. Healthy cells, for example, or tumour cells. Basically we got them to repeat what they were used to doing with every cell in every sample”. Once the experts had “instructed” the programme, it was able through statistical inference and on the basis of recognition of colorimetric variations in pixels to classify cells rapidly, and in an exhaustive manner.
“So, this is a tool that with the help of experts can automatize tasks,” according to Benjamin Stévens. “At some point the experts, by annotating the images, shape the algorithm in a circular relationship. According to the first annotations the algorithm offers an initial prediction that the experts may correct. The algorithm remembers the corrections, recalibrates its values, becomes more exact, and eventually the model becomes robust”. The classification that is suggested automatically to the expert is thus at least as accurate as the manual activity of classification, and as exhaustive, because all the pixels in the image are taken into account. The algorithm can develop into a routine calculation, and the amount of time saved is then quite considerable. “The quantity of information that is sent to doctors is dramatically reduced,” Stévens said. “A doctor no longer has to hunt through thousands of cells, there are just a few hundred. The information is concentrated, made more reliable, and it makes diagnosis faster to help doctors. There’s no thought of replacing the diagnostic ability of a doctor, or the doctor’s expertise.” Another benefit: since the algorithm is for general use, and because it learns its task from experts, it is assumed that it can classify any kind of image. It only has to be given enough examples, and it performs the task automatically. Research into image analysis that is going on now is studying the validity of these algorithms with a wide variety of kinds of images.
… that answers primary needs.
That which functioned properly in a single programme of research on a specific cancer was assumed to work as well for an entire range of pathologies, whether this had to do with cytology or histology. This early collaboration gave the team the idea of developing Cytomine as a generic software platform for the analysis of histological and cytological images. With the help of Loïc Rollus, informatics specialist at the University of Liège, Cytomine would now take on a whole range of “intelligent” algorithms, as well as tools and interfaces that would be required before and after using the programme. “In addition, we gradually realized that the simple fact of being able to store and share these images filled a primary need, which allowed us to produce several different versions of the programme,” Stévens said. “But that was not out initial aim. We just wanted to create a tool that would make analyzing images easier”.