Machine Learning-Aided Screening of Tuberculosis in Chest X-Rays.
Abstract
As a highly infectious and contagious disease, Tuberculosis(TB) still remains a major threat to
global health especially in low and middle income countries such as Uganda and ranks highest
in the causes of death from a single infectious agent. A very important step in the TB diagnosis
and treatment is screening and early detection of the disease, with chest x-ray ( CXR) being
the mostly widely used tool. However, there are challenges of limited number of radiologists
and inherent human error during CXR interpretation. In an effort to solve this problem, there
has been a significant rise in the interest to use artificial based intelligence-based TB screening
tools.
In this work, we explored the use of deep learning techniques in the automatic detection
of TB, not only distinguishing TB positive from health but also pneumonia and COVID-19
that usually manifest in a similar manner. A dataset of 700 TB images and 720 images for
each of the other classes was accrued from two publicly available datasets. Transfer learning
was performed on five models. Three of them based on the Convolutional Neural Network
(CNN) architectures; the EfficientNet-B4, ResNet50 and Xception models. The other two
were Vision transformers; the ViT B-16 and ViT B-32.
The EfficientNet-B4 posted the best overall performance with a test accuracy of about
97.9%. For ease of use, the best performing model was deployed in a web application to allow
the user upload the image and get results in a matter of seconds.
The achieved results show that the trained model has the capability to serve as a clinical
decision support to radiologists during TB screening.