| Abstract [eng] |
In this master‘s thesis, instance segmentation of bacterial image using the region-based convolutional neural network Mask R-CNN is performed and the distribution of bacteria in the images is evaluated. Data was obtained from a study of the antibacterial activity of silver and gold nanoparticle-coated titanium dioxide thin films against Veillonella parvula and Neisseria sicca species associated with oral disease. Summarized metrics for the best model: average precision at IoU = 0.5 is 0.873, mean average precision at different IoU values is 0.551, and F1 coefficient with particle accuracy is 0,898. The characteristics of bacteria in the images were calculated and it was observed that there are moderate / strong linear relationships between the model metrics and the circularity, solidity, roundness, aspect ratio and extent. The mean or median distance between the centroids to the nearest neighbor was used to estimate the distribution of bacteria in the images, whether the bacteria particles were random distributed, clustered, or self-avoiding. It has been found out that in most of the images the bacteria are randomly distributed, in a smaller part they are self-avoiding and form clusters in the smallest part. When comparing the mean values of the metrics in the images of different bacterial distributions, it was found that the mean of the metrics of the randomly distributed particles is statistically significantly higher than the mean of the metrics of the self-avoiding particles and the particles forming the clusters. Furthermore, the means of the self-avoiding particle metrics are in almost all cases statistically significantly higher than the means of the same metrics of the clustered particles. The study developed a methodological process for the evaluation of bacterial distribution in antimicrobial surfaces, which could significantly contribute to further studies of antibacterial activity against Veillonella parvula and Neisseria sicca, as it would allow the evaluation of each colony-forming bacterium separately. In summary, the weak point of the methodology used is the detection and instance segmentation of the bacteria that make up the cluster. However, it is also a difficult task to mark bacteria manually, as it is not always possible to pinpoint the limits or even the number of bacteria involved. To fully apply the developed methodology in practice (especially in the identification of cluster-forming bacteria), the model for the instance segmentation task should first be improved. |
