Técnicas de inteligência computacional aplicadas ao diagnóstico de tuberculose

Abstract

Tuberculosis (TB) remains one of the deadliest infectious diseases globally, necessitating rapid and accurate diagnostic methods. Sputum smear microscopy, a microscopic technique to identify TB bacilli, is prevalent due to its simplicity and affordable cost. However, it is a laborious, entirely manual approach that requires intensive training of the examiner. In this context, this work investigates the application of Deep Learning (DL) techniques to assist the examiner and enhance the accuracy of sputum smear microscopy. For this, an extensive evaluation of various convolutional neural network models and object detection methods was conducted on diverse datasets, aiming to understand their performance and applicability in detecting and counting TB bacilli. In addition, aiming to improve the models performance, several strategies are proposed and evaluated, including creating a new public dataset, data augmentation techniques, image partitioning strategies, and implementing a voting system to facilitate the annotation of images for training. Data augmentation proved to be essential for increasing the models robustness, while image partitioning improved performance and allowed for handling variations in image size, maintaining standardized model parameters. The proposed voting system has the potential to become a valuable tool for creating customized training sets, vital for the generalization of models in different contexts. A Computer-Aided Diagnosis software prototype was also developed, integrating the assessed techniques and models, aiming to assist the examiner during the counting of bacilli in TB diagnosis. This prototype represents a significant advancement for the practical application of DL techniques in sputum smear microscopy, with the potential to accelerate and make the diagnostic process more precise. The experiments conducted provided insights into the possibilities and limitations of using DL in TB bacilli detection, also evaluating the potential for autonomous diagnosis. This research substantially contributes to the field, offering a detailed analysis of the use of DL in TB diagnosis via sputum smear microscopy, and methods that can be tested in real environments. In summary, this thesis provides an in-depth overview of the implementation of DL in TB bacilli detection, highlighting its potentialities, challenges, and establishing a solid foundation for future research and practical applications.