Deep learning for bacterial colony counting

Details

Supervisors

Lee, John Aldo ; Smyth, Paul

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées

Abstract

The counting of bacterial colony on Petri dishes is an important step in the elaboration of a vaccine. The part of virulent and avirulent colonies and their number allows the laboratory workers to check the efficiency of the vaccine. For the time being, the counting task is done manually or semi-automatically. The goal of this thesis is to use deep learning in order to automate this task. The first step of the work was to create a dataset. With a labeling tool, we constituted a dataset of 105 images of Petri dishes containing more than 2500 bacterial colonies. After, to address the problem, we used image segmentation techniques. We implemented three families of network structure: the UNet, the UNet with ResNet encoder and the UNet++. For the learning part, we introduced two losses: the cross-entropy loss and the dice loss. After all, we used the hybrid loss that is the linear combination of this last two. The learning was performed by three optimizers: SGD, Adagrad and Adam. Finally, we implemented three methods of data augmentation: the rotation of images, the addition of noise in the network and the mix-up technique. By using a genetic algorithm, we found hyper-parameters of the best model. This model reached precision and recall around 99% for colony counting.