Machine-learning Approach to Microbial Colony Localisation

Due to the massive expansion of the mass spectrometry, increased demands for precision and constant price growth of the human labour the optimisation of the microbial samples preparation comes into question. This paper deals with designing and implementing an image processing pipeline that takes an input in the form of a Petri dish image with cultivated colonies and outputs the position of possible sampling points. In total 547 samples were collected. The first block of the pipeline consists of a trained customised ENet model which predicts a binary mask. Architectures U-Net, UNet++ and ENet were examined, where ENet was found to perform with the highest Dice coefficient (0.979).