Avaliação da radiografia de tórax com algoritmos de interpretação automatizada para rastreamento de tuberculose em massa em prisões

Abstract

Introduction: Tuberculosis remains a serious global public health problem and it is the leading isolated cause of death among infectious diseases. The incidence of tuberculosis in the population deprived of liberty is 10 times higher than in the general population, early diagnosis is a control strategy in these high-incidence environments. The use of x-ray combined with artificial intelligence can accelerate the detection of active tuberculosis cases, during mass screening in prisons. In this perspective, this study aimed to compare four automated x-ray interpretation methods for mass tuberculosis screening in prisons. Methods: A cross-sectional study was carried out in three prisons in the state of Mato Grosso do Sul, from which the data were collected from 2017 to 2019. Demographic data, prison history, lifestyle, health history and symptoms suggestive of tuberculosis were collected through a questionnaire. Then, the xray was performed, and the images were electronically transferred for automated analysis by CAD4-TB, versions 5 and 6, and to the LUNIT INSIGHT CXR2 system, versions Tb and MCA. For both CAD4-TB versions, with score range from 0 to 100, scores ≥ 60 was considered altered. While for LUNIT system, the score ranges from 0 to 1, being altered, the value ≥ 0.88. After the x-ray, a sputum sample was collected to be processed in the Xpert MTB/RIF. All information was recorded in REDcap® and sensitivity, specificity and area under the AUCcurve were calculated, and the results was compared using the confidence interval. Results: 7081 x-ray images were analyzed and observed an AUC of 0.92 was observed for LUNIT-TB, 0.88 for CAD4TB v5 and v6. It was identified that 33.4% of individuals with altered radiographs in LUNIT-TB were diagnosed with tuberculosis and specificity of 94.8%. In addition, in LUNIT-TB it performed more powerfully to differentiate individuals with low and high scores when compared to the CAD4TB versions. Conclusion: The results suggest that the use of chest radiography supported by automated interpretation can be an efficient strategy for tuberculosis mass screening in high-risk environments.