Tuberculosis (TB) infects 10 million people each year and disproportionately affects people in low-to-middle income countries. Early diagnosis can be difficult for several reasons, including that symptoms can be similar to other respiratory illnesses, the disease may present differently in patients who are also HIV-positive and access to and cost of medical care in some countries may be a challenge.
Sanjase
The World Health Organization (WHO) has endorsed the utilization of computer-aided diagnosis (CAD) for tuberculosis screening as a way to increase efficiency where there is scarcity of experienced personnel, according to Nsala Sanjase, MBchB, medical officer at the Center for Infectious Disease Research in Zambia.
Moving forward, the focus must be on developing tests that are high-quality enough to diagnose TB.
Dr. Sanjase's team worked to validate a tuberculosis screening AI algorithm developed by Google with support from the Bill and Melinda Gates Foundation, results of which were presented in a Thursday session.
"Computer-aided diagnosis has the potential to aid in the interpretation of chest X-rays in regions where expert radiologists are scarce," Dr. Sanjase said. "This has the potential to enhance screening, thereby reducing diagnostic test expenses."
AI System Was Comparable To Radiologists' Review
The study recruited adults at three clinical sites in Zambia who had symptoms suggestive of TB, were close contacts of TB patients or were newly diagnosed with HIV. Of the 1,807 patients with a definitive TB status, 641 were HIV positive and 190 were TB positive.
With a goal of using AI to interpret chest radiographs (CXR), this project prospectively evaluated a CXR AI model that included both an AI model that detected TB and one that detected CXR abnormalities.
CXRs were obtained from 1,923 patients with abnormal clinical status, of which 555 had CXR abnormalities. The TB AI's AUC was 88%, and the nine radiologists' sensitivities were comparable to or slightly below the AI at matching specificity.
The CXR abnormality AI's AUC was 97% with sensitivity and specificity respectively at 97% and 79%.
In addition, both the TB and CXR abnormality AIs demonstrated high performance in an HIV-enriched population.
While the AI system has promise, there is still work to be done in this area, according to Dr. Sanjase. Since this clinical study ended, Google has developed a second version of the AI with improved performance, but it has yet to be validated. There is also a need to develop a similar system to be used in pediatric patients.
"I am deeply committed to eradicating tuberculosis, a disease that has persisted for over a century. The significant toll it takes and the ongoing hardships it inflicts, particularly in resource-limited areas, are undeniable," Dr. Sanjase said. "To bring an end to this epidemic, we must break the chain of transmission. This requires identifying all individuals who are infected and promptly initiating treatment."
Access the presentation, "Prospective Multi-Site Validation of AI to Detect Tuberculosis and Abnormalities on Screening Chest Radiographs in an HIV Enriched Population in Zambia," (R1-SSCH09-4) on demand at Meeting.RSNA.org.
