Computer-aided X-ray boosts TB screening accuracy in Uganda

Tuberculosis remains a global health challenge, and finding the best way to screen people quickly and accurately is essential. Modern computer-aided detection (CAD) programs read chest X-rays and give a numeric score that reflects how suggestive the image is of TB. But until now, it has been hard to measure how well CAD works across the whole group of people who show up for screening, including those with low X-ray scores or no symptoms. Joowhan Sung and colleagues addressed that gap by offering portable chest X-ray screening with CAD (qXR v3) to people aged 15 and older in Uganda. Screening was open to everyone regardless of symptoms. The study used a low cut-off score of 0.1 (on a scale from 0 to 1) to decide who should give a sputum sample for confirmatory molecular testing with Xpert Ultra. By starting with broad, community-style screening and collecting confirmatory tests for people above the pre-set score, the team could estimate CAD performance in a real screening setting rather than only among people already suspected of having TB.

The study screened 52,835 people with CAD-scored chest X-rays. Of those, 8,949 (16.9%) had scores at or above 0.1 and were asked to provide sputum. The team obtained valid Xpert Ultra results for 7,219 participants; 382 (5.3%) of these were Xpert-positive, including 81 with trace results. To estimate how well qXR v3 would detect Xpert-positive TB across the whole screened population, the researchers made conservative assumptions about people with scores below 0.1 — for example, assuming 0.1% of that low-score group would have been Xpert-positive if tested. Under that assumption, qXR v3 had an estimated area under the curve (AUC) of 0.92 (95% confidence interval 0.90–0.94) for detecting Xpert-positive TB. The team also compared using one universal score threshold for everyone to using thresholds stratified by age and sex. At a fixed specificity of 96.1%, sensitivity was 75.0% with a universal threshold (≥0.65) versus 76.9% with age- and sex-stratified thresholds, a statistically significant improvement (p=0.046).

These results suggest CAD can be more accurate for TB screening than earlier estimates implied, because this study measured performance in a broad population that included people without symptoms or clear X-ray abnormalities. Using qXR v3 together with Xpert Ultra confirmatory testing, the researchers showed high overall discrimination (AUC 0.92) and demonstrated that small gains are possible by personalizing the score threshold by age and sex. In practical terms, programs that rely on a single universal cut-off may miss opportunities to improve detection or to use resources more efficiently. Stratified thresholds could allow screening programs to maintain high specificity while nudging up sensitivity in groups where that matters most. The study, supported by the National Institutes of Health, points toward a more tailored approach to chest X-ray screening that could better match testing to individual characteristics and the realities of community screening.