Geoffrey Raposo: Active tuberculosis detection from frontal chest X-ray images

Tuberculosis (TB) is one of the leading causes of death from a single infectious agent in the world. In many high-burden regions, which often lack specialized healthcare professionals, Chest X-Ray (CXR) exams continue to be of vital importance in the diagnosis and follow-up of the various presentations of the disease. In this context, automated systems to support diagnosis from CXR images constitute a fundamental cog as the World Health Organization (WHO) confirmed in early 2021 that they can be used in place of human readers for the interpretation of digital CXRs.

In this study, we investigate the benefits of automatic Pulmonary Tuberculosis (PTB) detection methods based on radiological signs found on CXR. Contrary to direct scoring from images, implemented in most related work, indirect detection offers natural interpretability of automated reasoning. We identify generalization difficulties for direct detection models trained exclusively on the modest amount of publicly available CXR images from PTB patients. We subsequently show that a model, pre-trained on tens of thousands of CXR images using automatically annotated radiological signs, offers a more adequate base for development. By relaying radiological signs through a simple linear classifier, one is able to obtain state-of-the-art results on three publicly available datasets (test AUC on Montgomery County-MC: 0.97, Shenzhen-CH: 0.90, and Indian-IN: 0.93). We further discuss limitations imposed by the limited number of PTB-specific radiological signs available on public datasets, and evaluate possible performance gains that could be obtained if more were available (test AUC MC: 0.98, CH: 0.98, IN: 0.93).

We then analyze the relative importance of each of the radiological signs for PTB prediction using two distinct methods and conclude that more than a specific sign, it is their combination that allows a reliable detection of the disease.

Finally, we propose a visual overview of the radiological signs predictions over radiographs using grad-CAMs and highlight the importance of annotating PTB datasets to study the reliability of these visualizations.