Detecção automática de bacilos em baciloscopia de campo claro usando aprendizado profundo e técnica de imagem mosaico

Abstract

Tuberculosis (Tb) is one of the top 10 causes of death worldwide. The diagnosis and treatment of Tb in its early states are fundamental to reduce the rate of people affected by the disease, since the transmission of the Kock bacillus, the agent that causes Tb, is done through the respiratory route. To assist specialists in the diagnosis of this disease, many studies have been published to automatic detection of Kock's bacillus in bright field smear images, an exam frequently used to diagnose the disease. In this work, a bacillus detection method using convolutional neural networks (RNC) is presented to perform a segmentation task associated with a technique do build the images of the database, called image mosaic names. The methodology consists in the implementation of convolutional neural networks to perform a segmentation of objects of interest, in the case, the bacilli, in a mosaic image, followed by counting of segmented bacilli. Three RNC architectures, three optimization methods and four methods to evaluate a generalization of each architecture were evaluated. In total, 36 simulations were performed. Evaluating the simulation performances, we verified that networks with few layers, with a higher noise incidence, that is, some pixels are wrongly classified as bacilli. This is because few layers impair the network learning and the classes differentiation. An architecture with more layers, using the ADAM optimization method and the dropout generalization method shows the best results when compared to other models. This model reached values for metrics precision, precision, sensitivity, specificity and F1 score above 99%. These metrics were used for model’s evaluation.