Automatic diagnostics of tuberculosis using convolutional neural networks analysis of MODS digital images

Abstract

Tuberculosis is an infectious disease that causes ill health and death in millions of people each year worldwide. Timely diagnosis and treatment is key to full patient recovery. The Microscopic Observed Drug Susceptibility (MODS) is a test to diagnose TB infection and drug susceptibility directly from a sputum sample in 7-10 days with a low cost and high sensitivity and specificity, based on the visual recognition of specific growth cording patterns of M. Tuberculosis in a broth culture. Despite its advantages, MODS is still limited in remote, low resource settings, because it requires permanent and trained technical staff for the image-based diagnostics. Hence, it is important to develop alternative solutions, based on reliable automated analysis and interpretation of MODS cultures. In this study, we trained and evaluated a convolutional neural network (CNN) for automatic interpretation of MODS cultures digital images. The CNN was trained on a dataset of 12,510 MODS positive and negative images obtained from three different laboratories, where it achieved 96.63 +/- 0.35% accuracy, and a sensitivity and specificity ranging from 91% to 99%, when validated across held-out laboratory datasets. The model's learned features resemble visual cues used by expert diagnosticians to interpret MODS cultures, suggesting that our model may have the ability to generalize and scale. It performed robustly when validated across held-out laboratory datasets and can be improved upon with data from new laboratories. This CNN can assist laboratory personnel, in low resource settings, and is a step towards facilitating automated diagnostics access to critical areas in developing countries.