Deep learning for bacterial colony detection

During vaccine development, after vaccine samples are added in Petri dishes with blood agar, within several days from a single bacteria many bacterial colony forming units (CFUs) will grow. Two types of colonies, virulent (bvg+) and avirulent (bvg-), appear in the dishes. Counting the number of viable bacteria and computing the proportion of virulent bacterial colonies is an important step to assess the vaccine yield. To make the overall vaccine creation process easier, in recent years deep learning models have been implemented to automate the counting of the CFUs. However the hyperparameter optimization of the deep learning models are tedious. In order to deal with the configuration problem of the models, in this study to automate CFUs counting we implemented nnUNet which is a self-configuring framework for deep learning-based biomedical image segmentation. A dataset of 105 images of Petri dishes containing bacterial colonies and their associated masks were used for segmentation and detection of the bacterial colonies. The results were compared with the previous studies where some of the Automated Machine Learning (AutoML) methods were used for hyperparameter optimization. This study shows that nnUNet is a good alternative to other solutions in the previous studies for automating the bacterial colony counting and classification problem. While nnUNet shows similar performance with previous studies, it brings us the extra advantage of the self-adapting framework thus we save time and resources from hyperparameter optimization process. It is also once again shown that nnUNet can be a good starting framework for an image segmentation task. Even though nnUNet outputs state-of-the-art results over many datasets, it can be also used as a high-quality baseline for new segmentation tasks. Finally, starting a segmentation task with nnUNet can provide hyperparameter values that can be used for warm starting the AutoML methods like Bayesian hyperparameter optimization, genetic algorithms.