Tuberculosis (TB) is an airborne infection and a common cause of deaths related to antimicrobial resistance. Under-reporting of the disease and inadequate care for controlling the spread of the disease remain a challenge, especially in resource-constrained settings. There is a significant lack of expertise in interpreting radiology images in countries with heavy TB burden adversely impacting screening efficacy. Image-analysis-based computer-aided diagnosis (CADx) tools have gained significance because they offer a promise to alleviate the human burden in screening in countries that lack adequate radiology resources. However, a majority of these tools use handcrafted features to capture relevant characteristics in underlying data requiring expertise in morphological and textural image analysis. On the other hand, convolutional neural networks (CNN), a class of deep learning (DL) models, are shown to deliver promising results on visual recognition tasks with end-to-end feature extraction and classification. Ensemble learning (EL) methods combine multiple models to offer promising predictions because they allow blending of intelligence from different learning algorithms. In this work, we compare four different strategies for improving disease detection accuracy. The results obtained with a stacked ensemble of models from these proposals are found to be promising compared to the state-of-the-art methods to detect findings consistent with TB.

Keywords: Tuberculosis, computer-aided diagnosis, convolutional neural networks, deep learning, classification, ensemble learning, stacking, stacked ensembles.