Deep learning could help detect TB earlier in low-income countries

Tuberculosis is a contagious disease caused by bacteria Mycobacterium tuberculosis. It affects the lungs, but can also affect other parts of the body such as the brain or the kidneys. To detect it, the WHO recommends chest X-rays. However, the low-income countries where this disease is most frequent lack the experts to interpret it, researchers have developed a system of deep learning to detect active pulmonary tuberculosis on chest X-rays and compared its performance with that of radiologists. The study ” Deep Learning Detection of Active Pulmonary Tuberculosis on Chest X-Ray Matches Clinical Performance of Radiologists” was published in the journal Radiology.

Radiology is a journal published by the Radiological Society of North America, RSNA, an association of radiologists, radiation oncologists and medical physicists.

Tuberculosis kills more than a million people worldwide each year, and the COVID-19 pandemic has exacerbated the problem. One in four people in the world is infected with the bacillus Mycobacterium tuberculosisand 5% to 10% of these people will develop active tuberculosis (TB) during their lifetime. Nearly 90% of active infections occur in approximately 30 countries, most of which have limited resources to combat this public health problem.

Using AI to detect lung disease

Studies based on deep learning to screen for chest x-rays, followed by a confirmatory nucleic acid amplification test (NAAT) have been shown to be more cost-effective than using NAATs alone.

WHO evaluated 3 computer-aided detection systems and determined that their diagnostic accuracy and performance was similar to that of human readers. Faced with the shortage of experienced readers, the WHO now recommends computer-assisted detection for screening and triage in people aged 15 or over.

The study

For this retrospective study, the researchers developed a system of deep learning (DLS) to interpret chest radiographs for imaging features of active pulmonary tuberculosis. They tested it on datasets from China, India, the United States, Zambia and South Africa and evaluated it under two conditions:

  • When it had a single predefined operating point across all datasets;
  • When customized based on radiologist performance in each region.

Since diagnostic performance can be influenced by disease prevalence, they compared DLS to two different groups of radiologists: one based in a region where tuberculosis is endemic (India) and the other in a region not endemic (United States).

Next, they estimated the cost savings of using this DLS as a triage solution for NAATs in screening settings. This study aimed to model real-world deployment scenarios and assess generalizability in four areas with high TB ​​rates and limited resources.

The DSL was trained on 165,754 images from 22,284 people. For the 1st set of tests, the data came from China, India, USA and Zambia, the test set consisted of 1236 images, of which 212 were identified as positive for tuberculosis based on microbiological tests or NAAT. These were scored binary by 9 radiologists from India and five from the United States.

Study results

The DSL achieved higher sensitivity than the Indian 9 radiologists assay, at 88% vs. 75%, with specificity not lower than 79% vs. 84%. The performance of DLS has also been shown to be excellent in two commonly used case-control datasets in China and the United States.

The second set of tests consisted of subjects from a gold mining population in South Africa, a group with a high prevalence of tuberculosis, but also lung problems such as silicosis, emphysema and chronic obstructive pulmonary disease. DLS gave results comparable to those of radiologists, as it did for HIV patients, but in this series the performance was found to be much poorer.

Study co-author Sahar Kazemzadeh, software engineer at Google Health, says:

“What’s particularly promising about this study is that we looked at a range of different datasets that reflected the breadth of TB presentation, different equipment, and different clinical workflows. ».

The AI ​​system met the thresholds set by the WHO in 2014 for any TB test in most datasets, according to Bram van Ginneken, professor of medical image analysis at Radboud University Medical Center. from Nijmegen, the Netherlands, co-author of the study. It could be very useful for low-income countries because the modeling could lower screening costs by 40 to 80%.

For Sahar Kazemzadeh:

“Filling the shortage of experts is where AI comes in. We can train computers to recognize TB from x-rays so that in these low-resource settings, a patient’s x-ray can be interpreted in seconds. »

Sources of the article: Deep Learning Detection of Active Pulmonary Tuberculosis on Chest X-Ray Matches Clinical Performance of Radiologists”

Radiology, DOI:10.1148/radiol.212213

Sahar Kazemzadeh, Jin Yu, Shahar Jamshy, Rory Pilgrim, BEng, Zaid Nabulsi, Christina Chen, Neeral Beladia, Charles Lau, Scott Mayer McKinney, Thad Hughes, Atilla P. Kiraly, Sreenivasa Raju Kalidindi, Muyoyeta World, Jameson Malemela, Ting Shih , Greg S. Corrado, Lily Peng, MD, Katherine Chou, Po-Hsuan Cameron Chen, Yun Liu, Krish Eswaran, Daniel Tse, Shravya Shetty, Shruthi Prabhakara.

Google Health, 1600 Amphitheater Pkwy, Mountain View, CA 94043 (SK, JY, SJ, RP, ZN, CC, NB, SMM, TH, APK, GSC, LP, KC, PHC, YL, KE, DT, SS, SP ); Advanced Clinic, Deerfield, Illinois (CL); Apollo Radiology International, Hyderabad, India (SRK); Tuberculosis Department, Center for Infectious Disease Research in Zambia, Lusaka, Zambia (MM); Sibanye Stillwater, Weltevreden Park, Roodepoort, South Africa (JM); and Clickmedix, Gaithersburg, Maryland (TS).

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *