Leggi in italiano

Credit: Manjurul/Getty Images.
The analysis of bacterial cultures is the most common study in microbiology laboratories and guides decisions such as the choice of antibiotic therapies. In a new study1 a group of researchers led by Alberto Signoroni, at the University of Brescia, has proposed an AI algorithm, called Deep Colony, that can interpret culture plates and offer a first identification of the bacterial species they contain.

When tested on nearly 5,000 plates processed by a large US laboratory to diagnose urinary tract infections, Deep Colony matched the experts’ judgement in 95.4% instances. The algorithm design was inspired by the cognitive process normally followed by microbiologists.

A plate usually hosts several colonies, which can correspond to different bacterial strains. Using visual inspection, microbiologists normally formulate a first hypothesis about the strain in each colony. This hypothesis can be tested by regrowing samples from each colony separately and then employing mass spectroscopy techniques.

"We used images of nearly 26,000 isolated colonies of 32 bacterial species identified with mass spectrometry in various US laboratories, or included in the [public biology database] American Type Culture Collection,” Signoroni explains. This dataset was used to train the first neural network in Deep Colony to estimate the probability of a given colony containing each of the 32 strains.

The results were refined by comparing colonies in the same plate, something microbiologists usually do to contextualize their first guess. A second neural network clusters the colonies of a plate according to their similarity, disregarding the species identification. If colonies in the same cluster were estimated by the first network as probably belonging to different species, their identification was revised.

The results were then aggregated to label each plate as positive (if they showed significant growth of clearly identifiable bacterial species over the healthy flora), negative (no significant growth) or contaminated (too many species or not clearly identifiable). The agreement between Deep Colony and the manual interpretations was 99.2%, 95.6% and 77%, for positive, negative and contaminated plates, respectively. Deep Colony deemed positive many of the plates labelled as contaminated by microbiologists. “We maximized true negatives while allowing for some false positives, so that Deep Colony allows to focus on the most relevant or critical cases,” Signoroni concludes.