CAD4TB is an innovative chest x-ray analysis software platform developed by Radboud University Nijmegen in the Netherlands. The CAD4TB platform uses deep-learning technology to analyse chest x-rays and detect signs of tuberculosis fully automatically. The output from the algorithm is a heatmap, indicating suspicious areas in the lungs and an overall score from 0 to 100 indicating likelihood of tuberculosis based on the analysis.
TB TRIAGE+
Participation in the TB TRIAGE+ project will enable research into additions and improvements to the CAD4TB platform as follows:
Detection of other significant radiological findings.
A chest x-ray provides information on many other conditions, which may be present apart from TB. In this project we will focus on detection of an enlarged heart (cardiomegaly), lung cancer, and emphysema. This will enrich the capabilities of the platform while providing information about the prevalence of these conditions in hard-to-reach African populations.
Investigation of CAD4TB performance in the presence of HIV
It is known that HIV affects the appearance of chest x-rays and that patterns of TB may be more difficult to identify in these subjects. Since TB TRIAGE+ will collect information on HIV status this will enable research into the performance of CAD4TB in HIV positive subjects
Modelling the effect of aging on the chest X-ray
The age of a subject affects the appearance of the chest X-ray in multiple ways. In particular, an older subject is more likely to have scarring from previous illness in their life-time. Such issues can affect the performance of CAD4TB. The research during TB TRIAGE+ will investigate ways to incorporate subject age as part of the modelling process.
Simplification of the procedure to determine TB score threshold
CAD4TB outputs a score indicating likelihood of TB, which is typically thresholded at a certain value to decide which cases should be treated as TB positive. Determination of this threshold to obtain a desired sensitivity is not straightforward and the ideal threshold varies between datasets for many reasons. Using the large dataset collected during the TB TRIAGE+ project will enable us to research modelling of the differences between datasets to simplify the process of threshold selection.
Figure 1: (a) The original x-ray image (b) The heatmap indicating regions of abnormality in the lungs per pixel. The final composite score for this subject is 91.7, indicating a very high likelihood of tuberculosis.
Figure 2. (a) Cardiomegaly (enlarged heart), width illustrated by red line. (b) A mass identified on chest X-ray, indicative of lung cancer. (c) Emphysema, indicated by flattened diaphragm, enlarged lungs, small vertical heart shape and hyperlucent lung fields.
Figure 2. (a) Cardiomegaly (enlarged heart), width illustrated by red line. (b) A mass identified on chest X-ray, indicative of lung cancer. (c) Emphysema, indicated by flattened diaphragm, enlarged lungs, small vertical heart shape and hyperlucent lung fields.
Figure 1. Sample output from CAD4TB (a) The original radiograph, (b) The radiograph with abnormality heatmap overlay.