Abstract
BACKGROUND: In resource-limited areas, severe shortages of radiologists contribute to high rates of missed pulmonary tuberculosis (PTB) cases when relying solely on conventional chest X-ray (CXR). Although artificial intelligence-powered computer-aided detection (CAD) has proven effective in PTB diagnosis, its real-world performance remains underexplored. OBJECTIVE: This study aimed to evaluate the real-world diagnostic yield of CAD technology as a triage tool for detecting PTB in primary health care facilities in high-burden areas. METHODS: We conducted a retrospective paired-design diagnostic yield study using CXR images collected from 7 county- and 32 township-level health care facilities in Yichang city between 2022 and 2024 year. All images were retrospectively reprocessed with CAD software (JF CXR-1), and the original reports interpreted by radiologists at the time of patient admission were extracted. CAD and radiologist performances were compared using 2 primary evaluation indicators-diagnostic yield among diagnosed cases (DYD) and positive predictive value (PPV). Subgroup analysis (by region, age, sex, health care facility tier, and patient category) and sensitivity analysis were conducted to assess the robustness of the results. RESULTS: Among 93,319 enrolled study patients, including 273 (0.3%) bacteriologically confirmed PTB cases, CAD demonstrated a substantially higher DYD (229/273, 83.9%) than radiologists (70/273, 25.6%), although the PPV was much lower (1.70% vs 10.31%). This high-sensitivity performance achieved an 85.5% (79,804/93,319) reduction (only 13,515 instead of 93,319 CXRs) in radiologist workload via selective review of CAD-positive images, without missing any radiologist-identified PTB cases. Furthermore, probability scores greater than 0.75 were a key threshold for identifying high-risk patients with PTB, and these patients were prioritized for radiologist review. Subgroup analysis further revealed that CAD outperformed radiologists in identifying PTB cases across all scenarios, despite some heterogeneity. CAD performance was significantly better in township-level medical facilities (DYD: 86.7%; PPV: 2%) than in county-level hospitals (DYD: 62.5%; PPV: 0.6%). CONCLUSIONS: CAD technology is valuable for detecting PTB in primary health care facilities. Combined with a tiered artificial intelligence prescreening with selective human review strategy, this approach effectively alleviates the workload of radiologists in resource-constrained regions, offering a scalable solution for tuberculosis prevention and control.