Thesis title: Smart Cross-Polarized Scattering: Optical technique and intelligent processing for bacteria recognition
Over the years, bacteria have been studied in various sectors, with both healthcare and industrial applications. This has led to the development of multiple chemical-physical techniques for their characterization, including PCR and mass spectrometry, which over time have become among the most used in microbiological laboratories. However, these techniques present several limitations, such as the need for chemical intermediaries, destructive analysis and the long times for characterization. As a result, research has been oriented towards the development of new, more efficient techniques.
In particular, optical techniques are attracting growing interest due to their speed, non-destructive nature and possibility of integration with machine learning. However, a problem common to all new techniques is the need to have a large database consisting of thousands of data points in order to reliably evaluate their potential. The construction of such databases represents an extremely high economic and temporal cost. To overcome this difficulty, the use of synthetic databases, obtained through numerical simulations, is being explored.
In this context, we introduce Cross Scattering Polarization, a cross-polarization tech-nique for the recognition of individual bacteria. Thanks to the use of numerical models and COMSOL Multiphysics software, we generated a total of 2880 images of four different bacteria: Bacillus subtilis, Bacillus globigii, Salmonella and Vibrio cholerae. The objective is twofold: to evaluate the recognition capacity of the technique and to analyze the effect of acquisition noise on the classification of models trained with a noise-free dataset.
For this study, we used two classification algorithms: a convolutional neural network, specifically VGG-11, and an algorithm based on k-nearest neighbors (KNN) combined with a principal component transform (PCA). In the case of the convolutional neural network, the results were excellent in the absence of noise; however, in the presence of noise it was necessary to apply pre-processing to the images to improve accuracy. On the contrary, in the case of the KNN + PCA algorithm, no pre-processing was necessary to obtain high accuracies.