Computer-aided detection for radiological disease severity classification on chest radiograph in children with intra-thoracic tuberculosis.

Implementation of the World Health Organization's (WHO) recommended shorter 4-month treatment regimen for non-severe tuberculosis (TB) in children requires classification of disease severity on chest X-ray (CXR). Access to specialists for CXR interpretation is limited. We explored the use of computer-aided detection of CXR ("CAD") to automate CXR classification of radiological disease severity. To do this, we combined three CXR datasets from children with confirmed and clinically diagnosed TB across the disease spectrum. CXRs were independently classified as radiologically severe or non-severe by two expert human readers. Definition of radiological disease severity aligned with WHO guidelines. CAD scores were generated by CAD4TB v7.0 and qXR v3.0 software. Neither software product was specifically trained with paediatric CXRs or for disease severity classification. We compared CAD scores between CXRs classified by human readers as non-severe versus CXRs classified by human readers as severe. CXRs from 526 children were included in this analysis: median age was 2.1 years (inter-quartile range 1-4.2 years); 57% of the children had microbiologically confirmed TB. We found that median CAD scores were significantly lower for CXRs classified as non-severe versus severe by human readers; the difference was greatest in children >5 years. The area under the receiver operating curve was 0.82 and 0.78 for qXR, and 0.79 and 0.76 for CAD4TB, against the reference of 'severe' as classified by each individual human reader respectively. These results demonstrate that CAD is a promising tool for TB disease severity stratification and has the potential to support access to shorter TB treatment regimens for children. Investment in paediatric CAD training and development to optimize solutions for children beyond the TB screening and diagnosis use-case is warranted.