Toward a lightweight model for pulmonary tuberculosis classification

As in every developing country, Madagascar is still fighting against the tuberculosis (TB), the second leading cause of death from a communicable infectious disease. Its incidence remains high, exacerbated by limited medical infrastructure and unequal distribution of health resources. Several researches were conducted to automatically diagnosis disease to greatly lowering the total expenses. Deep learning approaches demonstrated its high effectiveness in the domain using chest x-ray images (CXR) that are the most cost-effective and widely utilized imaging technology to classify TB. Some preprocessing was necessary for getting better features from the images. Using a more lightweight model inspired by SSD Lite X, this work achieves the highest accuracy (training, validation and testing), recall, F1-Score and specificity of 100% each, making this model the most effective one in the state-of-the-art. It could be deployed in a low-resource environments, enabling automated diagnosis in every region of Madagascar and other developing countries