Our study was conducted in 3 counties of Yichang city, Hubei province (Yidu, Zigui, and Changyang counties), a region characterized by a high PTB burden, with an overall incidence of PTB of nearly 100 per 100,000 population between 2022 and 2024. This study involved patients aged above 15 years (including outpatients and inpatients) who underwent CXR examinations at health care facilities due to symptomatic presentation. We used a paired diagnostic yield study design where each chest radiograph underwent 2 independent assessments—from an AI-powered CAD system and from clinical radiologists whose interpretations were the original reports from the time of patient presentation, thereby reflecting real-world conditions. To address the issue of multiple CXRs potentially occurring for the same patient during the study period, we established a stratified selection protocol—for non-PTB cases, the initial CXR examination was used and, for PTB cases, the CXR temporally closest to the confirmation date was selected. To reduce temporal misclassification bias between CXR acquisition and TB confirmation, we set a 6-month window period as the validity criterion.

All CXR images were extracted in Digital Imaging and Communications in Medicine format from the Picture Archiving and Communication System-Radiology Information System of county- and township-level health care facilities in 3 study counties. The data collection period spanned from January 1, 2022, to December 31, 2024. The system also recorded patient demographics (gender and age) and the original radiology reports, including examination time, report date, the original radiological findings, and the diagnostic conclusions issued by the radiologists at the time of patient care. In this study, a suggestive PTB case was based on the content of the “diagnostic conclusions” field, which contained the radiologist’s ultimate clinical judgment on the CXR. If specific terms related to active PTB (such as “tuberculosis” or “pulmonary tuberculosis”) appeared in this field, the case was classified as “abnormalities suggestive of PTB.”

Data on PTB cases were retrieved from the Tuberculosis Information Management System (TBIMS). TBIMS is a population-based surveillance system that mandates reporting of all PTB cases, and the dataset used in this study spanned from January 1, 2022, to June 30, 2025, and included patient diagnosis type and confirmation date [16]. PTB cases in the TBIMS included clinically diagnosed PTB cases and bacteriologically confirmed PTB cases, in accordance with the diagnostic criteria and case classification guidelines issued by the Chinese national health authorities [17]. Clinically diagnosed cases were identified based on chest imaging, supplemented by clinical manifestations, immunology tests, and pathological examination or bronchoscopy. Bacteriologically confirmed cases were diagnosed based on a comprehensive assessment involving chest imaging, bacteriological tests (sputum smear microscopy for acid-fast bacilli or mycobacterial culture), molecular biological tests (such as Xpert MTB/RIF or Xpert Ultra), and lung tissue pathological examination.

For AI-based CXR interpretation, we used the JF CXR-1 software (version 2; JF Healthcare). It is an advanced AI system built on deep learning technology, primarily used for TB detection in CXRs with external validation support [18-20]. In 2022, the software was approved by China’s National Medical Products Administration as a class III medical device. The system generates probability scores ranging from 0 to 1, where higher values indicate a greater likelihood of active PTB. According to the manufacturer’s manual, scores exceeding 0.35 mandate additional diagnostic evaluation for TB. Consequently, this study used a threshold of >0.35 as the criterion for radiographic abnormalities suggestive of active PTB.

The following steps were performed on the software:

We evaluated diagnostic performance using 2 primary indicators—diagnostic yield among all those diagnosed (DYD) and positive predictive value (PPV). DYD is defined as the proportion of confirmed patients with TB in whom the diagnostic tool (CXR interpreted by CAD or radiologists) successfully identified the disease [21]. Unlike a conventional sensitivity indicator, the DYD is calculated within a pragmatic clinical context. Its denominator includes all confirmed patients who underwent CXR as part of routine care, thereby accounting for real-world variables such as the time interval between imaging and confirmatory diagnosis. As a result, the DYD reflects the actual effectiveness of a tool in detecting TB among known patients under typical clinical circumstances, rather than its maximum theoretical accuracy under idealized research conditions. PPV is defined as the proportion of patients confirmed to have TB among all results identified as positive by a diagnostic tool (CXR interpreted by CAD or radiologists). It is crucial to acknowledge that this calculated PPV, based on bacteriological confirmation alone, should be interpreted as a lower-bound estimate. The true PPV is likely higher, as some CAD-positive cases not confirmed by bacteriology might represent occult TB missed by conventional diagnostics.

Sensitivity analysis was performed to assess the robustness of CAD and radiologist performance in identifying PTB cases using CXR under different scenarios: (1) time window validation—assessment of different PTB confirmation intervals (3 mo and 9 mo vs the primary 6 mo window), (2) image selection validation—impact of image selection strategy (using the last CXR examination vs the initial CXR examination for patients without TB), and (3) case definition expansion—inclusion of clinically diagnosed PTB cases, to re-evaluate the performance of CAD and radiologists.

The study was approved by the Institutional Review Board of the Chinese Academy of Medical Sciences and Peking Union Medical College (CAMS&PUMC-IEC-2025‐04). The study retrospectively analyzed deidentified CXR images and PTB data obtained from routine clinical practice. A waiver of informed consent was received from the institutional review board. To protect privacy and confidentiality, all patient data were anonymized at the source, and the analysis was conducted exclusively on aggregated or coded datasets.

From January 2022 to December 2024, a total of 132,159 posteroanterior chest radiographs were collected from 7 county-level hospitals and 32 township health centers across 3 counties (Yidu, Zigui, and Changyang) in Yichang city, China. All images were accompanied by complete radiological reports. By linking PTB case data from the TBIMS, the final study population comprised 93,319 patients (non-PTB cases: n=93,046, 99.7%; bacteriologically confirmed PTB cases: n=273, 0.3%). The study outline is presented in Figure 1.

A total of 93,319 patients (Figure 1) were included in this study. Characteristics of these patients are presented in Table 1. The median age of patients with PTB was higher than that of patients without PTB (69 years vs 60 years; P<.001), and male gender preponderance (70.7% vs 50.1%; P<.001) was observed. A higher proportion of patients with PTB came from township-level health care facilities compared to patients without PTB (88.3% vs 61.2%; P<.001). The proportion of patients with PTB from outpatient clinics was comparable with that of patients without PTB (29.7% vs 28.4%; P=.70).

Among 93,319 participants, 13,515 cases were identified as suspected PTB by CAD, with a positive rate of 14.5% (13,515/93,319), which significantly exceeded the positive rate of suspected PTB (679/93,319, 0.7%) identified by radiologists. Among the 273 patients with confirmed PTB, CAD achieved DYD of 83.9% (229/273), which was substantially higher than that of radiologists (70/273, 25.6%). CAD captured all 70 PTB cases detected by radiologists, demonstrating a low risk of missed detection. However, the PPV of CAD was only 1.7% (229/13,515), which was much lower than the radiologists’ PPV of 10.3% (70/679; Figure 2).

Strategy 3 required radiologists to interpret all 93,319 CXR images. It achieved a PPV of 1.7% (229/13,515), reflecting a marginal decrease compared to strategy 2 (PPV: 229/13,556, 1.7%). The DYD of strategy 3 was 83.9% (229/273), consistent with strategy 2. Strategy 4 similarly necessitated full interpretation of all CXRs. Its PPV increased to 11% (70/638), slightly exceeding strategy 1 (PPV: 70/679, 10.3%). The DYD of strategy 4 remained 25.6% (70/273), consistent with strategy 1. Strategy 5 enabled a radiologist workload reduction of 85.5% (1‐13,515/93,319) while maintaining a minimum system DYD of 25.6% (ie, finding all cases the radiologists originally found) and a potential DYD of 83.9% (if radiologists trusted the AI on all 229 positive cases; Table 2).

Among patients with abnormal CAD, each patient’s CAD score was classified into 4 subgroups by quartiles. The distribution of CAD scores was found to be linearly correlated with the proportion of suspected patients with PTB identified by radiologists and confirmed patients with PTB (P for trend <.001). Most confirmed patients (189/229, 82.5%) had CAD scores exceeding the 75% quantiles (Q75) threshold, which accounted for 34.2% (4618/13,515) of all CAD-positive patients (Table 3).

CAD consistently demonstrated superior DYD compared to radiologists across all evaluated demographic and health care system classification variables (Figure 3). However, CAD showed significant variation in performance with respect to health care facility tier, performing poorer in the county-level group than in the township-level group (P<.001; Table S1 in Multimedia Appendix 1). Multivariate logistic regression also demonstrated that advanced age (≥80 y) groups were less likely to achieve high PPV (OR 0.21, 95% CI 0.10‐0.44; P<.001). No significant differences were observed between the sexes and patient categories (outpatient or inpatient).

During temporal window validation, shortening the PTB confirmation window to 3 months slightly increased CAD’s DYD to 86.6% (increased to 86.6% from 83.9%; Table S2 in Multimedia Appendix 1), whereas extending it to 9 months reduced DYD to 81.8% (decreased to 81.8% from 83.9%; Table S3 in Multimedia Appendix 1), indicating a minor influence of temporal factors on performance. Image selection validation (using final CXRs vs initial CXRs for non-PTB cases) revealed only subtle differences with unchanged core conclusions (Table S4 in Multimedia Appendix 1), confirming the reliability of the original strategy. When clinically diagnosed PTB cases (n=71) were included (Table S5 in Multimedia Appendix 1), the DYD of CAD decreased by only 2.8% (decreased to 81.1% from 83.9%), remaining significantly higher than that of radiologists (81.10% vs 25.58%). The DYD of radiologists decreased slightly (decreased to 25.6% from 25.6%). The consistent trends in both groups confirm that the relative advantage of CAD was not substantially affected.

This retrospective study evaluated the real-world performance of CAD in PTB detection using CXR interpretation. Among 93,319 patients analyzed from health care facilities, 273 (0.3%) were bacteriologically confirmed as PTB cases within 3 years. The results demonstrated that CAD outperformed radiologists in identifying PTB cases, detecting an additional 159 cases. Compared to radiologist interpretation alone, the strategy of CAD screening followed by radiologist verification not only enhanced PTB case identification but also reduced the workload for radiologists. A CAD score exceeding 0.75 was a key threshold for identifying high-risk patients with PTB, demonstrating high consistency with radiologists’ judgments and supported by statistical trends confirming significantly higher PTB risk with elevated scores. Notably, CAD exhibited superior diagnostic performance in township medical institutions compared to county-level hospitals, effectively overcoming inherent resource constraints.

This study observed that CAD demonstrated substantially higher DYD than radiologists in routine clinical practice. This comparison is inherently asymmetric, contrasting a specialized AI algorithm operating under optimal conditions against the daily performance of radiologists working in high-volume, multitasking environments [22]. Therefore, our findings are not intended to demonstrate that CAD is superior to radiologists in a direct competition. Instead, we aim to reveal and address a critical systemic flaw—the DYD of the current passive case-finding system in real-world clinical settings is alarmingly low (70/273, 25.6%). The significantly higher DYD (229/273, 83.9%) of CAD underscores its potential as a system-level triage tool, effectively bridging the sensitivity gap caused by resource constraints and radiologist fatigue.

The PPV of the CAD system is the most conservative estimate, as the CAD is almost certainly identifying true PTB cases that the system missed. However, excessive reliance on CAD-based diagnoses may potentially contribute to a higher volume of unnecessary follow-up procedures, leading to an inefficient allocation of medical resources. In resource-constrained primary care settings, CAD systems can serve as a triage tool for rapid PTB diagnosis [23]. By adopting a strategy of computer-aided screening followed by selective radiologist review, this approach optimizes the TB triage workflow. Such a stepwise process can effectively alleviate the workload of manual film reading while maintaining high sensitivity and reducing false-positive results [24]. Consistent with previous studies, CAD scores not only function as a diagnostic threshold but also possess predictive value for PTB risk based on continuous score variations [25]. CAD interpretation outcomes demonstrate high consistency with radiologists’ CXR readings, providing an objective reference for clinical decision-making [26-28].

We found that the diagnostic yield of CAD demonstrates significant variation across different levels of health care institutions, performing better in township-level facilities than in county-level hospitals. This performance disparity may be explained by several plausible factors related to the distinct patient populations and diagnostic pathways within the local health care ecosystem. We speculate that the reasons could be as follows: township-level facilities, serving as the primary contact point for a large volume of rural patients, may typically encounter cases at more advanced stages. Patients in these settings might tend to seek medical attention only when symptoms become severe and unmistakable, often presenting with more pronounced and characteristic radiographic abnormalities of PTB. These classic manifestations are likely to align well with the pattern recognition capabilities of current CAD systems. In contrast, county-level hospitals function as regional referral centers, potentially receiving a more complex case mix. This may include patients with early-stage or subtle disease, diagnostically challenging cases referred from township facilities, and a higher prevalence of nontuberculous pulmonary pathologies. Such complex and atypical presentations might approach the current detection limits of CAD algorithms. This finding underscores the importance of contextualizing CAD performance within the specific clinical environment and patient population it serves, and the proposed explanations warrant further investigation.

Previous research has centered on 2 PTB screening models—active screening in high-risk groups (eg, high-burden populations [19], cross-border migrants [29], and enclosed environments [30]) and passive screening in specialized facilities (eg, tertiary hospitals [11] and TB diagnostic centers [31]). This study pioneers PTB triage integration within China’s hierarchical health care system, offering novel evidence for AI deployment in primary care. While Kagujje et al [32] also studied primary settings, their focus was on prior TB history as an AI diagnostic confounder. The observed DYD of JF CXR-1 (229/273, 83.9%) aligns with high sensitivities reported previously (eg, Qin et al [20] reported a sensitivity of 86.4%). Compared with the sensitivity shown in clinical validation (eg, 78.9% in a study by Yang et al [12]), there is still a certain difference in performance in primary health care. Our proposed strategy of artificial intelligence prescreening with selective human review demonstrated its capability by reducing radiologist workload by 85.52%, aligning with the findings of Munjal et al [33]. On the basis of the TBIMS data, Xin et al [25] confirmed the effectiveness of CAD in individuals aged 65 years and above. This study further supplements evidence for the effectiveness of CAD in other application scenarios.

The 6-month interval between CXR acquisition and TB confirmation was defined primarily based on frontline physicians’ clinical experience, which indicates that the majority of diagnosable TB cases progress to a confirmable stage within this period. This window aligns with common diagnostic timelines in real-world practice and is methodologically consistent with previous studies (eg, Xin et al [25]). However, this exclusion may introduce selection bias, as the excluded cases with longer diagnostic delays may systematically represent patients with more subtle or early-stage disease that is inherently more challenging to detect. Consequently, the removal of these potentially harder-to-diagnose cases could lead to an overestimation of the DYD for both the CAD system and radiologists. Reassuringly, sensitivity analyses using both 3-month and 9-month windows consistently demonstrated the superior performance of the CAD system over radiologists, with only minor fluctuations in DYD values. This indicates that our primary conclusion remains robust despite the potential influence of this bias.

For CAD-positive cases not clinically diagnosed with PTB, the lack of subsequent verification makes it impossible to differentiate between true CAD false positives and occult TB cases missed by conventional diagnostics. The reported PPV of 1.7% is almost certainly an underestimate. The CAD system is likely identifying true active PTB cases that were missed by the current passive case-finding system, which relies on symptomatic presentation and subsequent bacteriological testing. Therefore, our findings not only demonstrate the high DYD of CAD but also point to a substantial number of potential PTB cases that merit further clinical investigation. This underscores the important role of CAD as a sensitive triage tool capable of uncovering the hidden burden of PTB in the community. Future studies with long-term clinical follow-up or additional diagnostic tests for CAD-positive individuals are needed to ascertain the true PPV.

This study was also limited to a unimodal radiographic input and did not integrate clinical symptom information. Since symptom profiles can influence both radiologists’ interpretations and the output of CAD systems, it would be valuable to evaluate whether diagnostic performance varies between symptomatic and asymptomatic subgroups. Although such a subgroup analysis fell outside the scope of this work, future investigations that incorporate multimodal data could help elucidate the relationship between symptomatic presentation and CAD performance, thereby contributing to the development of more clinically adaptive screening pathways.

CAD dramatically improves the detection of PTB among symptomatic patients presenting to primary health care facilities, while substantially reducing radiologist workload through a selective review strategy focused on CAD-positive images with reference to the probability score. Meanwhile, CAD demonstrates significantly superior performance in township medical facilities compared to county medical facilities, providing validated evidence for implementing tiered diagnosis and treatment models in resource-limited regions.