Cell-tracking algorithm keeps AI in check

3D cell-tracking software gives researchers the ability to automatically monitor the development of cells in tissues. Using neural networks, these programmes identify in series of microscope images where cells are moving and from which new cells have originated. Unfortunately, this method frequently makes errors in tissues with large numbers of cells, necessitating time-consuming manual correction.

‘In our group we study the development of intestinal organoids, small pieces of intestinal tissue that we grow in the lab’, says biophysicist Max Betjes, a doctoral candidate at the physics research institute AMOLF. ‘Tracking is extremely challenging for this, because the intestinal cells are packed very closely together and divide very rapidly. The manual correction work involved can easily consume half a PhD trajectory.’ Betjes therefore developed a new algorithm that not only tracks cells more accurately, but crucially prompts the underlying neural network to indicate where potential errors have occurred. This makes it far quicker to identify which parts of the cell tracking need to be corrected. Betjes: ‘In this way, our algorithm increases the throughput of experiments with organoids.’

Political philosophy

The fact that neural networks assign a certain probability to whether their prediction is correct or incorrect is nothing new. A classic example is a neural network presented with an image of a dog or a cat that must determine which of the two it is. In doing so, the network calculates an underlying probability distribution to indicate how confident it is in the given answer. ‘But our question is far more complex than that single decision’, says Betjes. ‘With organoids, the neural network must look at thousands of cells all at once. That is often still too difficult.’

For this reason, the researchers’ algorithm has the neural network examine each individual cell in an image, before combining all the results within a statistical framework. This yields a margin of error indicating the degree of certainty with which the cell tracking prediction was made. Betjes: ‘The innovation of our algorithm lies in the rigorous statistical approach we use to achieve this.’

Although Betjes drew his primary inspiration for this statistics from statistical physics, he draws an analogy with political philosophy, where he encountered a similar statistical approach. ‘Suppose you have many experts who all say something about reality — how do you then combine all those opinions into a single consensus? The underlying statistics that allow you to do that are also what we use in our algorithm.’

Winning a competition

The researchers demonstrated the performance of their algorithm by testing it on data from intestinal organoids from their own laboratory. They also made use of publicly available data from early-stage mouse embryos and from C. elegans, a small worm whose embryonic development always follows an identical course and has been known since the late 1990s through manual tracking.

For the latter, Betjes even entered the so-called ‘Cell Tracking Challenge’, a competition that ranks the data analysis of cell-tracking algorithms by measuring them against the manually tracked C. elegans data. He came first. ‘Though this only reflects how accurately our algorithm made its predictions. More important to us is that the algorithm correctly indicates the degree of certainty with which it made those predictions.’

Getting started

To further validate the performance of their algorithm, Betjes and colleagues from the groups of Sander Tans and Jeroen van Zon are now hoping to enthuse other groups to start working with this method. To this end, they have built a website where other researchers can test whether the algorithm works on their data, without needing to download the software.

‘Our algorithm will probably be one of the best available for tracking cells in complex tissues over the coming two years’, says Betjes. ‘We are therefore now focusing on reaching people who may also benefit from this method.’

Those interested can try out the algorithm at OrganoidTracker.org.