Rapid counting of Kazachstania humilis and Saccharomyces cerevisiae in sourdough by deep learning-based classifier

When maintaining sourdough through backslopping, bakers must ensure that the yeast mycobiota remains stable. By introducing two-staged incubation temperatures for cultivation, we found that the colonies of Kazachstania humilis and Saccharomyces cerevisiae could be differentiated by size and color. We then developed a classifier that used the deep-learning method, YOLO, to automatically count these colonies. For sourdough isolates of K. humilis and S. cerevisiae, the classifier had accuracies of 0.99 and 0.98, respectively. This classifier also showed accuracies greater than 0.95 for S. cerevisiae strains used in bread, sake, and wine. To investigate the practical feasibility, the sourdough was repeatedly refreshed by backslopping at 25 °C, 30 °C, and 35 °C, with the goal of artificially fluctuating the yeast mycobiota. At 25 °C, K. humilis and S. cerevisiae accounted for proportions of approximately 25 % and 75 %, respectively, whereas at 30 °C and 35 °C, K. humilis comprised less than 1 % of the mycobiota. The accuracy of this classifier was 0.98 for K. humilis and 0.99 for S. cerevisiae; this was very close to the accuracy obtained with manual counting, indicating that the classifier could detect changes in the yeast mycobiota. The classifier took approximately 126 milliseconds to count colonies on one Petri dish. The use of our novel classifier can enable fast, less-laborious, and objective judgement, potentially facilitating the ability of small-scale artisan bakeries to manage fermentation on a daily basis.