10 Years of the International Cell Tracking Challenge and IT4Innovations' Involvement

The organisers of this initiative maintain and update a collection of heterogeneous datasets. These include datasets showing heavily clustered and single cells, fast-moving cells, and, for example, image data of deliberately poor quality (e.g., blurred images). More than 50 research teams worldwide are currently participating in the Cell Tracking Challenge. Computer scientists are trying to process specific datasets. They submit their results to the competition for review, and the competition organisers then evaluate them for accuracy and reliability. The results are then published on the initiative's website. This way, they build a ranking of the best competitors or the best algorithms.

The Cell Tracking Challenge also acts as a catalogue. If scientists need to process their data, they can access this competition, find out, and choose the best method currently available for the given data type. Therefore, as a condition of the competition, participants must make their programs publicly available.

Vladimír Ulman from IT4Innovations, one of the organisers of the Cell Tracking Challenge, explains the idea of cell segmentation and tracking: "Cell segmentation is a task in which the computer has to find all the cells in a given image - it is the same task as, for example, wanting to outline every elephant in a picture of a herd of elephants (that would be elephant segmentation). And it is precisely the ability to outline that is crucial because then the computer can calculate the size of the cell (or elephant), the shape category, etc. And if it was not an image of elephants but a video of a herd of elephants, a time-lapse series of images, then tracking means finding the same elephant in every video frame. After successful segmentation and tracking, the computer can identify a particular elephant or cell and see how, for example, its shape changes over time or the distances between them. Biologists are often interested in how cells change their volume or number over time, whether they all move in the same direction or chaotically, etc. If a computer is to detect and summarise this autonomously, it needs to be able to perform both segmentation and tracking well. For it is impossible to do without that."

After ten years of the Cell Tracking Challenge and five years, since the first results were published in the prestigious Nature Methods journal, the competition's organisers have again published the latest research results in Nature Methods. The paper "The Cell Tracking Challenge: 10 Years of Objective Benchmarking" describes what has happened in the field over the last five years and where the competition has gone. Vladimir Ulman adds: "The main underlying theme is that these five years have seen a small revolution in the form of the emergence of artificial intelligence (AI). In 2017, we had a single AI method; today, we get virtually no results other than through AI. Our latest paper describes what methods are most used and which are suitable for which type of data. We also provide analyses of the performance of the methods and how they relate to the nature of the input data, etc. We explain how we have gradually added new data types to this competition, expanding the catalogue portfolio and the potential applicability of the underlying effort."

IT4Innovations was partly responsible for the data dissemination. For example, IT4Innovations scientists are involved in developing a data annotation tool and a tool for collaborative cell tracking. Vladimír Ulman explains the need for data annotation: “The need for data annotation arose precisely because of AI methods. Annotation means manually outlining and tracking a part of the data from the catalogue to train computer image-processing models. Traditionally, annotation is done manually, which is inappropriate when large amounts of data need to be processed. IT4Innovations has come up with a solution. It enabled the creation of supercomputer-annotated data known as the silver standard corpus, where the computer could annotate 99.8 percent of all cells in the image. This gives AI plenty of examples of how to train itself well. This result is freely available on the Cell Tracking Challenge website. AI methods trained on this data are also appearing in the competition. We hope to hear more about computer-annotated data from IT4Innovations as we develop this area actively."

For the paper, see: https://rdcu.be/dcrUp

This work was supported by the ERDF in the IT4Innovations national supercomputing center - path to exascale project (CZ.02.1.01/0.0/0.0/16_013/0001791) within the OPRDE.