Students' Research Circle    
 
 
2022
2021
Call for papers
The conference
» Veterinary Session
Veterinary Jury
Biology Session
Sponsors
Awards-list
2020
2019
2018
2017
2016
2015
2014
2013
2012
2011
2010
2009
2008
2007
2006
2005
2004
2003
2002
Home » Archive » 2021 » Veterinary Session

Veterinary session

Counting of Colony Forming Units using deep neural networks
Fodróczy Bettina - year 5
University of Veterinary Medicine Budapest, Department of Microbiology and Infectious Diseases
Supervisors: Dr. László Makrai, Dr. Norbert Solymosi

Abstract:

The identification and quantification of microorganisms per unit mass or volume is a common task in the field of microbiology, for example in food hygiene, infectology or biological research. In the case of solid medium testing, the viable cell count of the sample is determined by counting colony forming units. To this day, this task has been done manually, however it is time-consuming and labour-intensive process, which entails considerable costs as well. Automation of the counting of the colony forming units has been a long-standing request.

In our research, we have developed a convolutional neural network (CNN) algorithm that can estimate the number of the colony forming units from an image taken by a smartphone. We cultured 28 species of microorganisms, using the medium and incubation time adopted to the needs of each microbe. Digital photographs of the cultures were taken using a smartphone and then annotated using a web-based image annotation tool called COCO Annotator (https://github.com/jsbroks/coco-annotator). Using our annotated images, we trained the R101-DC5 Faster R-CNN neural network pretrained on the train2017 COCO data.

Based on the weights obtained after 40 000 iterations of the CNN, the prediction accuracy for each species was 93.9%. This shows, how well the species-specific value of the colonies annotated in the validation set matched the same value in the predicted set. When checking the CNN on the test set, the agreement between the number of colonies predicted and observed was 100% in the case 5 species, and above 95% in the case of 4 other species. In total, 10 species showed over 90% agreement, but 5 microbes had a result below 50% in the test. There are some species where the algorithm was very precise, but the lack of precision is not clear even for species with low agreement, as there were cases where the variance between different test images of some species was 66.3%. Based on the CNN weight database, a user-friendly software with a graphical user interface called CFU-detector was developed, which allows the user to detect and estimate the number of bacterial colonies in images taken of Petri dishes.

Viable cell counting is an essential task in microbiology and there is an international demand for its automation. The results of our research show that the CNN offers a potential solution to this problem, as is able to reproduce manual colony counting almost perfectly. In the future, this will not only reduce the manual labour requirements of the task, but also contribute to more accurate, faster and cheaper quantification.



List of lectures