Modelos de detecção de tuberculose ativa via imagens de raio X utilizando curvas principais

Abstract

Tuberculosis (TB) remains one of the most lethal infectious diseases worldwide, particularly in low- and middle-income regions where access to efficient healthcare systems is still limited. When not diagnosed early, the disease can progress to severe stages, leading to death or causing significant long-term respiratory sequelae, which reinforces the need for continuous improvement of screening and diagnostic tools, especially given the increasing volume of examinations and the demand for methods that are fast, robust, and accessible. This dissertation proposes and evaluates a computational intelligence method for tuberculosis detection using chest Xray image analysis, whose main distinguishing feature is its parsimonious nature, emphasizing simplicity, a reduced number of parameters, and low computational complexity, aligning with contemporary Green AI principles that highlight the importance of energy-efficient systems with minimal environmental impact. The principal method integrates the Principal Curves (PC) technique, employed as a specialist mechanism to extract relevant geometric structures from images, with Principal Component Analysis (PCA), responsible for dimensionality reduction while preserving essential variability in the dataset(CPE+PCA), but was also tested a method using SVD for reduction of dimensions (CPE+SVD), one where it was only used principal curves (CPE-Solo), and finally one where was utilized a Resnet-50 for feature extraction together with principal cuves and Principal Component Analysis (CPE+Resnet+PCA). For evaluation, a widely referenced database was used, the TBX11K, containing three classes: tuberculosis patients (800 images), patients with other pulmonary diseases (3,800 images), and healthy individuals (3,800 images), enabling the simulation of a realistic clinical screening scenario. The results indicated an accuracy of 0.87, a precision of 0.69 for TB, an average processing time of 0.020 seconds per image, and a carbon footprint of 0.010 kg of CO2-eq and an energy expenditure of 0.065 kWh, values that are competitive when compared to more complex deep learning-based methods, particularly given the model’s low computational cost and high energy efficiency. These findings demonstrate that the proposed approach offers a promising alternative for applications in resource-constrained settings and reinforce the feasibility of developing assisted diagnostic solutions that are simultaneously accurate, efficient, and environmentally responsible, contributing to the advancement of artificial intelligence applications in healthcare. Keywords: Tuberculosis; Artificial intelligence; Main Curve; K-Segments, Image Analysis, Ethics, Green AI