A colony counter that uses artificial intelligence is a true counting aid and counts many times quicker than an analyst; this article outlines the current status of this technology. Dutch version.
A new generation of colony counters now makes use of artificial intelligence (AI). The software is trained to recognise the colonies visually and to count them, just as a human analyst would do. Except that the neural network used for this is many times quicker than an analyst. This means that these colony counters can actually count all the colonies in a petri dish in a few seconds and distinguish the colonies from an air bubble or other contamination. This makes them very accurate.
Peter Krul is the designer, developer and builder of such a colony counter and explains that the AI-trained software is now here to stay in laboratory equipment: ‘AI really is the future. You can see now that this development is ongoing and that the AI equipment differs significantly from its predecessors. For example, an AI colony counter is faster, more accurate and also simple to operate. This all means that these AI counters deliver robust data and that problems in laboratories such as space, shortage of personnel and growth can now be dealt with very simply.
‘People sometimes think that AI means that a device can think for itself and even carry out independent actions that are perhaps undesirable’ says Krul. But in the case of the colony counter, AI means that the neural network is trained visually, just as you would train an analyst. The principle is actually very simple; it sees in the same way as an analyst, but it can count more quickly.’
If the software is specifically trained once, it will not learn automatically by itself. ‘Nor is that desirable,’ explains Krul; ‘you always want to work with a standardised method and not with a fluid method, under ever-changing circumstances. You draw a line and subsequently work with that version. Correct standardisation and validated methods are essential in a laboratory.’ Because an AI colony counter is connected to LIMS, it is also important for a software version number to be linked to the data.
A standard neural network can be specifically trained for a certain combination of agar and micro-organisms. For this combination, the upper and lower lighting of the colony counter are calibrated, just like the HD camera, which allows reflection-less photography through the cover of the petri dish. The supplier of the equipment then trains the required neural network, so that the correct colonies are counted. In principle, the colony counter is then suitable for use and the software is custom-made for the client.
An HD camera creates a photograph of the petri dish and the software then ‘sees’ the colonies present, for which it has been trained. Within a few seconds, the connected hardware, for example a tablet, displays the photograph, clearly indicating which colonies have been counted. Following any potential additional count, this count is easy to validate - by ticking the screen with a pen - by analysts. In most cases however an additional count is not necessary. This enables swift analysis, dish by dish.
It is also possible to photograph the dishes one after another without direct assessment and to be displayed as a series dilution on a tablet. At a later time, this series can be analysed by an analyst. Following this analysis, the data can be stored directly in LIMS.
The storage of the data and photograph for each petri dish provides the opportunity to re-assess the petri dishes at a later stage, whereas normally the dishes would be disposed of. This storage also provides valuable input, for example for business intelligence (BI) and trend analyses. Krul: “The advantage of an AI colony counter is that it is very accurate and counts consistently. It therefore obviates the differences between counts that exists between different analysts. Because additional counts are made by analysts, you can also easily develop a trend analysis between analysts. One will count more accurately or quickly than another.
Another advantage of the data storage is that changes to the process are identified. So it’s also possible to remove unusual counts directly or to hang a so-called flag on them, so that they can be examined in detail at a later stage. The colony counters also have a filter tool, with which various data can be filtered out. Using this filter tool, you can also isolate the dishes showing high additional counts; these dishes can then be used for training.
In the case of changing circumstances, or for example the use of a different type of supplier of the growth medium, it is also possible to re-train the AI colony counter. “For example, for our own AI colony counter, Iris, a specific training set has been developed to create data sets for the training of neural networks,” explains Krul. “The analysts encircle a number of colonies from several dishes as accurately as possible on the screen or tablet. Using the training set, we can then re-train the software remotely. For example, one of our machines is located in Taiwan, and we can easily re-train the software from within the Netherlands.”
“We must be realistic,” continues Krul, “and recognise that even an AI colony counter is not 100% accurate. What is annoying is that people tend to think that it’s only human for an analyst to make mistakes, but a device must always be completely accurate. The data that these counters provide are however very consistent and robust, and the results are delivered exceptionally quickly. With accurate additional counting by analysts, you achieve a high degree of accuracy quickly and easily.
Another potential disadvantage is that the photograph is less suitable for counting if there are labels, adhesive tape or pen markings on the underside of the petri dish. Krul: “Condensation can also be a limiting factor when we have to photograph through the top lid, but these circumstances are naturally best avoided. This requires an only slightly adapted dish protocol and of course, in the event of an occasional ‘dish contamination’, an analyst can always carry out a manual count.
Because an AI colony counter is easy to train, it can count almost everything in a petri dish. This could include glass fibres or bugs or various types of shapes, colours or sprouting stages of seeds. Krul: “An AI colony counter counts what you teach it to count; if a human can tell the difference between colonies visually, AI can also do that.”
What this means is that there is a huge amount of growth potential for AI laboratory equipment. This equipment can eradicate a huge number of problems encountered by laboratories from day to day.
The neural networks only have to be trained on a single occasion, in contrast to continually having to train new analysts. Or that it prevents repeated turnover of analysts and that different actions can be carried out by different people, depending on the level of the analyst. “In addition to this, also consider that in the case of an AI colony counter, you can also expand it into an automated input and output unit and you can see that the AI equipment can deliver huge time savings, which make it an investment with a very swift payback period,” concludes Krul.
This article has been jointly created with VMT. The VMT editorial staff consists of specialized subject editors and operates in a network of legislative and food safety experts, food technologists and sustainability experts.