Automated Colony Detection: Evaluating U-Net Models with DenseNet Backbone

The manual analysis of bacterial colonies in microbiology is time-consuming and prone to errors. Automated solutions can support scientific staff and enable more efficient processes. The suitability of U-Net models for automatic colony detection was evaluated for fluorescence-based applications. To investigate the impact of different DenseNet backbones on segmentation quality, U-Net models were trained with these backbones, resulting in high segmentation accuracy with average IoU values of 0.601. A particular advantage of this method is the pixel-precise segmentation, which allows for further analyses such as the size and distance of the colonies to each other. The developed models operated efficiently with a mask prediction time of just 0.097 seconds and required no preprocessing of the images, even in the presence of artifacts. The implementation is based on freely available libraries, making it more accessible. Overall, U-Net models demonstrate high potential for automated colony detection and can significantly enhance the efficiency of microbiological processes.