Computer-aided interpretation of chest radiography reveals the spectrum of tuberculosis in rural South Africa

Tuberculosis causes over 1 million deaths annually, challenging the WHO End TB Strategy goals. In high-burden, resource-limited settings, community-based screening can increase case finding but is constrained by the costs of molecular and culture diagnostics and limited availability of radiologists to interpret CXRs. Computer-aided detection (CAD) systems, such as CAD4TB, can rapidly score digital CXRs (0–100) for abnormalities suggestive of TB and may reduce diagnostic costs. Prior work has shown CAD4TB can identify abnormal lung fields and TB-related features and can serve as a triage tool, but optimal score thresholds likely vary by setting (TB prevalence, demographics, and available tests). HIV infection may alter radiologic manifestations of TB, potentially impacting CAD performance. Most evaluations of CAD4TB have been in clinical settings among symptomatic patients; data from population-based, symptom-agnostic screening are limited, and threshold selection for prospective active case-finding remains uncertain. This study prospectively applied CAD4TBv5 in a rural, HIV-endemic South African community during a mobile health screening program to maximize detection of active TB and characterize the spectrum of disease, comparing CAD performance with an expert radiologist and evaluating the newer CAD4TBv6 retrospectively.

The authors summarize WHO guidance supporting CXR as a screening tool and highlight the emerging role of CAD in systematic screening. Prior studies reported CAD4TB’s ability to detect TB-related abnormalities and reduce costs, but emphasized that triage thresholds require context-specific calibration. A retrospective study suggested lower thresholds in population screening versus clinical settings. Evidence on CAD performance among people living with HIV is mixed, as HIV-associated TB can present atypically on CXR. Previous CAD4TB evaluations have been mainly in symptomatic clinic populations, with limited prospective, population-based data guiding threshold selection for active case-finding.

Design and setting: Vukuzazi, a community-based, multi-disease screening program using mobile vans in rural uMkhanyakude district, KwaZulu-Natal, South Africa (May 25, 2018–May 24, 2019). Eligibility: residents aged ≥15 years; written informed consent (and guardian consent for minors). Data collection: structured questionnaires on TB symptoms (fever, weight loss, cough, night sweats), smoking, TB and HIV history; 4th-generation HIV Ag/Ab testing on venous blood. Imaging and CAD: Posterior–anterior digital CXRs (Canon CXDI-NE) acquired by health workers; DICOM images stored and scored locally offline by CAD4TBv5. CAD4TB methodology includes lung field segmentation and analysis of lung shape, symmetry, and costophrenic angles to produce a 0–100 abnormality score reflecting likelihood of active TB. Triage procedures: Per WHO survey guidelines, participants were referred for sputum testing if they reported any TB symptom or had a CAD4TB score at/above the triage threshold. Pilot phase: initial threshold 60 (based on literature) was tested on 1,090 participants; poor capture of radiologist-defined lung abnormalities prompted reduction to 25 to maximize sensitivity and in-field sputum collection. Radiologist reading: A senior radiologist, blinded to CAD scores and clinical data, centrally reviewed all CXRs, classifying lung fields as normal vs abnormal and indicating whether findings were diagnostic of active TB. For participants not triaged in the field but deemed to have abnormal lung fields by the radiologist, home visits were conducted to collect sputum. Microbiology: For triaged participants able to produce sputum on-site, specimens underwent Xpert MTB/RIF Ultra and liquid mycobacterial culture. Definitions: Definite TB = microbiological confirmation of M. tuberculosis by Xpert Ultra (including trace) and/or culture. Definite TB, trace excluded = as above but excluding cases with Xpert Ultra trace as the sole evidence. Probable TB = radiologist’s CXR diagnostic of active TB without microbiological confirmation. Analysis: Compared CAD4TBv5 and retrospectively computed CAD4TBv6 triage performance against the radiologist for detecting definite and probable TB; evaluated sensitivity, specificity, PPV, ROC AUC, PRAUC, number needed to test, and proportion triaged across thresholds; stratified analyses by HIV status. Logistics: Mobile radiography was solar powered; CAD processed CXRs offline within ~1 minute, enabling point-of-care triage.

• Population screened: 10,320 enrolled; 9,914 underwent CXR and CAD4TB triage (406 excluded due to pregnancy or inability to access van). Median age 39 (IQR 24–59); 29.8% HIV-positive; 10.7% reported any TB symptom. - TB yield: 99/9,914 (1.0%) had microbiologically confirmed (definite) TB; 75 (0.76%) when excluding Xpert Ultra trace-only positives. Additional 172 (1.7%) had probable (radiologic) TB. - Symptoms: 79.8% (78/99) of definite TB were asymptomatic; only 20 (20.2%) reported any TB symptom. - Radiology among definite TB: 80/99 (80.8%) had any lung field abnormality; 30/99 (30.3%) had CXR diagnostic of TB; 19/99 (19.2%) were read as normal by the radiologist. - CAD score distributions: Median CAD4TBv5 score: definite TB 63 (IQR 44–77.5); definite TB, trace excluded 66 (51.5–84); probable TB 87 (73–95), significantly higher than definite TB (p=5.3×10^-11); no evidence of TB 27 (22–40). CAD4TBv6 showed similar trends. - HIV status: CAD performance and scores did not differ significantly by HIV status across TB groups (p>0.05). - Pilot thresholding: At threshold 60, CAD4TBv5 captured only 27.3% (54/198) of radiologist-defined abnormal CXRs; triage threshold reduced to 25 for main phase to maximize sensitivity (84.8%, 95% CI 79.1–89.5) with specificity 65.7% (62.5–68.8) for detecting radiologist-defined abnormality. - Triage performance for definite TB: Radiologist sensitivity 80.8% (71.7–88.0), specificity 66.9% (65.6–68.2), PPV 4.7% (3.8–5.8). Comparable CAD thresholds: v5 at scores 39–40 achieved similar sensitivity (39: 82.8% [73.9–89.7]; 40: 79.8% [70.5–87.2]) but lower specificity (55.4% and 57.4%); PPV 3.6–3.7%. v6 at scores 47–48 achieved similar sensitivity (47: 82.8% [73.9–89.7]; 48: 76.8% [67.2–84.7]) with specificity comparable to radiologist (47: 62.6%; 48: 68.0%); PPV 4.3–4.6%. - ROC AUC: For definite TB vs no TB: v5 AUC 0.78 (0.73–0.83), v6 AUC 0.79 (0.73–0.84). For definite TB, trace excluded: v5 AUC 0.82 (0.77–0.87), v6 AUC 0.84 (0.79–0.89); not significantly different from definite TB detection. For probable TB: v5 and v6 AUC 0.96 (0.95–0.98), significantly higher than for definite TB (p<0.001). PRAUC lowest for definite TB (0.09) and definite TB trace excluded (0.08), highest for probable TB (v5: 0.42; v6: 0.44). - Proportion triaged and NNT: Using radiologist-defined abnormality for field triage would require sputum testing in 20.2% (2,002/9,914); NNT=25 for definite TB. At matched sensitivity thresholds: v5 would test 27.0–28.4% (NNT=34); v6 would test 20.1–23.7% (NNT=26–29), comparable to radiologist. - Low threshold impact: Using v5 threshold 25 in-field led to sputum collection for 59.6% (5,906/9,914), and identified 19/99 (19.2%) definite TB cases with radiologist-normal CXRs; 15 of these were asymptomatic and would have been missed by symptom or radiologist triage. Four definite TB cases had v5 scores <25 and were triaged due to symptoms. - Spectrum of disease: Definite TB included cases with highly abnormal CXRs (e.g., score 95, radiologist diagnostic) and radiologically subtle/normal cases (e.g., score 39, radiologist normal), underscoring subclinical presentations in community screening.

This study demonstrates that CAD4TB can triage CXRs for sputum testing in a community-based, symptom-agnostic screening program with sensitivity comparable to an expert radiologist. While CAD4TBv5 required testing a larger proportion of participants at matched sensitivity, CAD4TBv6 achieved similar sensitivity and specificity to the radiologist and comparable efficiency (NNT). The findings highlight that most microbiologically confirmed TB detected via community screening is asymptomatic and often radiologically subtle, supporting WHO’s move toward symptom-agnostic CXR-based screening. The wide range of CAD scores and radiologic appearances among definite TB cases suggests detection of earlier, potentially subclinical disease in community settings and indicates that CAD algorithms, typically trained on clinically diagnosed TB, may need refinement to better capture subtle patterns. Threshold selection for CAD-based triage is setting- and goal-dependent; piloting is crucial to balance case-finding against laboratory capacity and costs. HIV status did not significantly affect CAD performance, possibly reflecting effective ART coverage and immune reconstitution in this population.

In a rural, HIV-endemic community, CAD4TBv5 and v6 enabled efficient triage of digital CXRs for TB screening, with CAD4TBv6 matching radiologist performance in sensitivity and specificity. A low triage threshold maximized case-finding and uncovered asymptomatic, radiologically subtle TB that would be missed by symptom or radiologist-only strategies. CAD-based triage can feasibly replace human readers in TB prevalence surveys, provided thresholds are tailored via setting-specific piloting. Future work should focus on refining CAD to detect subclinical TB, improving transparency of training data, and optimizing threshold selection frameworks balancing yield and resource use.

• Only one radiologist performed blinded CXR readings, potentially biasing assessment and pilot threshold selection (radiologist instructed to overcall subtle abnormalities). - Microbiological confirmation relied on a single spot sputum sample, possibly underestimating definite TB. Inclusion of Xpert Ultra trace-only positives may have overestimated definite TB; sensitivity analyses excluding trace-only cases showed similar results. - No sputum was collected from: asymptomatic participants with CAD4TBv5 score <25 (n=3,772), those unable to produce sputum in camp (n=1,112), and those not reached for follow-up (n=54); true TB status of these groups is unknown. - High local HIV prevalence may limit generalizability to other settings. - Thresholds and performance may vary with CAD version; requires piloting for each context.