Contribution of artificial intelligence to the imaging diagnosis of pediatric pulmonary tuberculosis

Abstract

Pediatric tuberculosis (TB) remains a diagnostic challenge in Brazil and worldwide. The Brazilian Ministry of Health recommends a clinical scoring system (S-MoH) for children and adolescents with suspected TB. Interpretation of radiographs within this scoring system may require specialist input. AI-based systems, such as CAD4TB (Delft Imaging Systems B.V.), approved by the WHO for adults, are not yet recommended for standalone use in children under 15 years of age. A retrospective study was conducted at a pediatric institute from January 31, 2017, to January 29, 2025, including 179 patients aged 0–14 years with pulmonary TB or other diseases. CAD4TBv7.1 analyzed chest radiographs using two cutoff points established by Youden's index: 53.48 for analyses against the S-MoH score and 53.89 for analyses against microbiological confirmation. Results were compared with both microbiological confirmation and S-MoH score. Among the 179 participants, 61 (34.1%) had TB, 25 of which were microbiologically confirmed. CAD4TBv7.1 showed an area under the ROC curve (AUROC) of 0.71, with a sensitivity of 52% and a specificity of 86.3% compared with microbiological diagnosis. Against S-MoH, AUROC was 0.59, with a sensitivity of 34.43% and a specificity of 86.44%. CAD4TBv7.1 demonstrated low sensitivity and high specificity, particularly regarding its overall discriminative capacity. Thus, CAD4TBv7.1 emerges as a promising complementary screening tool for pediatric TB. Although its standalone use is not yet recommended, it may complement S-MoH in settings lacking radiologists. Investments in AI must be accompanied by consistent pediatric validation and strategies that combine technological innovation with traditional and cost-effective clinical approach.