Optimizing Computer-Aided Detection for Tuberculosis (CAD4TB) Thresholds for Community TB Screening in Nigeria: Recalibrating Threshold with Local Data

Abstract

Background: Computer-aided detection (CAD) is a promising tool for tuberculosis (TB) screening, but its performance, particularly the optimal abnormality score thresholds, is highly context-specific. This study aimed to optimise the use of computer-aided detection (CAD4TB) software in tuberculosis (TB) screening by customising its threshold settings based on individual patient characteristics and reported symptoms. Method: The study was a retrospective observational study utilizing secondary data from community-based TB screening using portable digital X-ray (PDX) with CAD4TB AI. A total of 18,529 adults were screened across various settings, with 16,047 participants having complete results on both CAD4TB and molecular tests (GeneXpert). Most participants were from the general population (35.0%), slums (25.3%), and hard-to-reach areas (16.5%). The mean CAD4TB score was 43.13 (range: 0–99.3). CAD4TB AI findings were validated against molecular results to determine diagnostic accuracy, sensitivity, specificity; PPV, and NPV were calculated, and ROC curves were generated to assess AI score thresholds and case detection yield by age and setting. Findings: The mean age was 43.5 ± 18.4 years; 62.6% were males.  The study found a high TB prevalence of 10.9%. The performance of CAD exhibited significant variation among individuals with a prior history of tuberculosis. The optimal threshold for this subgroup was a score of 48 (sensitivity 91.0%, specificity 41.9%, AUC 0.76), compared to a score of 36 for those without prior TB (sensitivity 90.6%, specificity 39.0%, AUC 0.81). This confirms that CAD accuracy is reduced among people previously treated for TB, likely due to post-TB radiological sequelae. While no single threshold met the WHO Target Product Profile (≥90% sensitivity and ≥70% specificity) for the entire population, the study demonstrated CAD's good overall accuracy (AUC >0.80) and its value as a high-sensitivity screening tool. Conclusion: CAD is an effective tool for community-based TB screening in Nigeria. However, its implementation requires locally validated, context-specific thresholds rather than universal cut-offs. Programmes must adopt different thresholds for key subgroups, such as those with a previous history of TB, to balance case detection with cost-effectiveness. Continuous re-evaluation of thresholds with each software update is essential to maintain optimal performance and impact.