Abstract:

Objective To evaluate the effect of artificial intelligence (AI) automatic film reading technology in tuberculosis screening of HIV/AIDS patients, and to provide scientific basis and reference for the formulation and improvement of active tuberculosis detection strategies in high risk population. Methods A retrospective study was conducted in 633 HIV/AIDS patients who were actively screened for tuberculosis using digital X-ray photography (digital radiography, DR) in 2019. Of them, 47 (7.4%) were pathogen-positive tuberculosis patients. Data of chest X-ray (“chest film”) of the patients were read using the AI automatic film reading technology from 3 different enterprises, and 3 senior experts were also invited to read the film. The results were analyzed, and the diagnostic effectiveness of manual reading and AI automatic reading technology was compared by reference to the etiological diagnosis of pulmonary tuberculosis. Results Of the 633 subjects, 198 (31.3%, 198/633) were diagnosed as suspected active tuberculosis by manual reading. The numbers of suspected cases detected by the three reviewers were 139 (22.0%, 139/633), 100 (15.8%, 100/633), and 90 (14.2%, 90/633), respectively. Of the 47 patients complicated with pathogen-positive tuberculosis, 19 (40.4%, 19/47), 29 (61.7%, 29/47) and 21(44.7%, 21/47) were diagnosed by the 3 experts, respectively, and a total of 14 cases (29.8%, 14/47) were misdiagnosed. By AI automatic film reading technology, 434 cases (68.6%) were diagnosed as suspected active tuberculosis. Using the technologies of 3 enterprises, the numbers of suspected cases detected by were 260 (41.1%, 260/633), 299 (47.2%, 299/633) and 247 (39.0%, 247/633), respectively. Of the 47 patients complicated with pathogen-positive tuberculosis, 32 (68.1%, 32/47), 30 (63.8%, 30/47), and 33 (70.2%, 33/47) were diagnosed, respectively. A total of 5 cases (10.6%) were misdiagnosed. Based on pathogen-positive diagnosis, the sensitivity of HIV/AIDS complicated with pathogen-positive tuberculosis was 70.2% (33/47), the specificity was 71.8% (421/586), and the consistency rate was 71.7% (454/633) by manual film-reading method; while by AI automatic film reading technology, the sensitivity was 89.4% (42/47), specificity was 33.1% (194/586), and the consistency rate was 37.3% (236/633). Conclusion AI automatic film reading technology is more sensitive than manual reading methods in diagnosis of pathogen-positive tuberculosis in HIV/AIDS population, and is suitable for active screening of tuberculosis.

Key words: HIV, Acquired immunodeficiency syndrome, Tuberculosis,pulmonary, Artificial intelligence, Diagnostic techniques and procedures