Classificação de bactérias usando imagens hiperespectrais e redes neurais convolucionais

Abstract

Deep learning according to Bezerra (2016) would be a sub-area for instruction in Machine Learning that analyzes and researches processes and methods to resemble human brain actions in functions such as speech recognition, or the process of body language and recognition visual, this has been sufficiently effective in pattern recognition and classification of hyperspectral images. With this study, the objective is to apply deep learning with convolutional neural networks in the recognition of hyperspectral images of bacteria in order to accelerate the sample identification process, reducing the cost of classification and increasing the effectiveness of the treatment. Beekmann et al. (2003) has bloodstream alterations (BSIs) as the tenth leading cause of death in the United States, and is also associated with a higher rate of hospitalization, where the “[...] hospital fatality rate ranged from 4.2% with ’bacteremia’ and 6.9% with E. coli septicemia at 22.2% with ’septicemia due to a Gram-negative, unspecified organism’ and 26.8% with ’inexact septicemia’ ” (BAINE; YU; SUMME, 2001). There was also a longer length of stay, where Pittet, Tarara e Wenzel (1994) cites during the study period the length of stay outside the hospital and in the SICU attributable to bloodstream infection of 24 and 8 days, respectively. The extra costs attributable to the infection averaged $40,000 dollars per survivor. Thus, according to Beekmann et al. (2003), the early identification of BSI enables an immediate optimization of antimicrobial therapy and the need for additional diagnostic studies, which can serve to reduce the length of hospital stay and these costs. The objective is to train a network capable of classifying bacteria by hyperspectral images, in order to reduce the cost of the patient’s stay in the hospital and efficiency in combating the microbe present in the body, since late identification is harmful to it. The method for the pre-processing of the images was through bibliographical research with the updating of improvement and updating of knowledge through a scientific investigation of works already published, together with a combination of empiricism and articles to compose the network’s architecture.