Special ReviewArtificial Intelligence, Radiology, and Tuberculosis: A Review
Section snippets
INTRODUCTION
Tuberculosis (TB) is an epidemic in many parts of the world, being responsible for 1.6 million deaths in 2017; in the same year, 10 million people developed the disease (1). TB is the leading cause of death from infectious disease worldwide and disproportionately affects developing regions, such as Africa and South-East Asia (1). Frequently, countries that suffer from a high TB prevalence are also resource-poor, lacking in medical staff and equipment.
Perhaps the most pressing issue relating to
CHEST RADIOGRAPHY AND TB
The radiographic presentation of pulmonary TB is varied, making it a challenging diagnosis. Firstly, it is important to distinguish between active and latent TB. Active TB is characterized by the presence of consolidation and cavitary lesions in the lungs, and has a high risk of infectious spread (2). By contrast, latent TB is characterized by stable fibronodular changes, such as scarring and nodular opacification, and has a low risk of infectious spread (although the disease can be reactivated
COMPUTER-AIDED DIAGNOSIS
In its early development, TB detection using AI was studied with conventional CAD (8). In CAD systems, detection is performed using a manually created preset feature model and the computer does not “learn” given a greater caseload. As a result, it has proved difficult to incorporate the varied presentation of TB into a single CAD system. CAD systems developed from 1996 to 2013 achieved accuracies ranging from 42% to 100%, usually for specific features such as the presence of cavitations (8).
DEEP LEARNING
Recently, AI deep learning networks have been developed for TB detection. The first of these was developed by Hwang et al. in 2016 (18). The authors used a model called AlexNet, a pretrained deep learning network that had previously achieved success in the ImageNet Large Scale Visual Recognition Competition—an image classification challenge for nonmedical images. AlexNet had already been pretrained for image recognition—this capability merely needed to be adjusted for use in medical imaging, as
TOWARD REFINEMENT IN DEEP LEARNING
Most recently, models for TB detection are becoming more refined and streamlined. Previously described deep learning models, such as AlexNet and GoogLeNet, are pretrained on millions of images and are capable of distinguishing between thousands of classes of images even before they are used in the context of radiology. As a consequence, they require substantial computer memory and hardware requirements to function effectively, even when working on the narrow task of detecting TB on a chest
CONCLUSION
The use of AI to detect TB in chest radiographs has progressed significantly in the past 3 decades. The field began with the development of CAD programs, which showed promising results for limited narrow applications. However, these tools were limited by their small dataset size and their reliance on preset feature models, a fundamental weakness of the technology itself. In the last 5 years, deep learning has made rapid progress in medical imaging, with TB detection emerging as an area of
REFERENCES (23)
- et al.
CD4 T lymphocyte count and the radiographic presentation of pulmonary tuberculosis
Chest
(1995) - et al.
Miliary tuberculosis
Chest
(1996) Global Tuberculosis Report 2018
(2018)- et al.
Pulmonary tuberculosis: role of radiology in diagnosis and management
RadioGraphics
(2017) - et al.
Radiographic predictors of subsequent reactivation of tuberculosis
Int J Tuberc Lung Dis
(2007) - et al.
Chest radiographs and their reliability in the diagnosis of tuberculosis
JNMA J Nepal Med Assoc
(2005) - et al.
Value of chest X-ray in TB diagnosis in HIV-infected children living in resource-limited countries: the ANRS 12229-PAANTHER 01 study
Int J Tuberc Lung Dis
(2018) - et al.
Automatic screening for tuberculosis in chest radiographs: a survey
Quant Imaging Med Surg
(2013) - et al.
Novel coarse-to-fine dual scale technique for tuberculosis cavity detection in chest radiographs
EURASIP J Image Video Process
(2013) - et al.
Localization algorithm and implementation for focal of pulmonary tuberculosis chest image
Detecting tuberculosis in radiographs using combined lung masks
Cited by (38)
A novel tuberculosis diagnosis approach using feed-forward neural networks and binary pattern of phase congruency
2024, Intelligent Systems with ApplicationsDeep learning and radiomics of longitudinal CT scans for early prediction of tuberculosis treatment outcomes
2023, European Journal of RadiologyUpdate on the diagnosis and treatment of tuberculosis
2023, Anales de PediatriaMachine learning approaches in diagnosing tuberculosis through biomarkers - A systematic review
2023, Progress in Biophysics and Molecular Biology