DENSITY MAP BASED BACTERIAL COLONY COUNTING USING DILATED CONVOLUTIONAL NEURAL NETWORK

Abstract

The counting of colonies of bacteria is a basic procedure in the field of microbiology and it is necessary to be able to estimate the level of concentration of microorganisms as well as to determine the results of experiments. The traditional approaches of counting are time-consuming, laborious, and human error prone, especially when high density colonies are present, overlapping or irregular in distribution. These drawbacks motivate this paper to suggest an automated framework of bacterial colony counting that relies on CSRNet, a convolutional neural network which trains on density map regression. The given method uses a VGG-16-based front-end to perform an effective feature extraction, and then dilated convolutional layers, which produce high-resolution density maps of the spatial distribution of bacterial colonies. To determine the total count of colonies, the predicted density map is summed up, which enables the model to be effective in dealing with partial, clustered and closely packed colonies without the need to explicitly identify or segment the objects. Ground truth density maps are created by convolving annotated locations of colonies with a Gaussian kernel, such that an adequate amount of supervision is available in training. The proposed method proves to be effective as it is shown to reach the level of counting accuracy of 97.36%  and low error rates (MAE = 4.15 and RMSE = 4.82) when evaluated experimentally using a dataset of bacterial colony images. The model additionally achieves high precision (96.92%), recalls (97.14%), and F1-score (97.03%) meaning that it agrees with manual counts and is also consistent in performance at different colony densities and imaging settings. A comparative analysis has revealed that CSRNet-based strategy offers competitive performance but with lower computational complexity as compared to detection-based models like Faster R-CNN and Efficient Models. Such findings prove that CSRNet-based density estimation is a strong, precise, effective method to automated counting of bacterial colony, and has a high probability of practical application to microbiology as well as biomedical fields.