Abstract
Background
Human reader variability and workforce constraints limit the quality and scale of chest X-ray (CXR) screening for tuberculosis (TB) in many high-burden settings. Although the WHO endorses computer-aided detection (CAD) software, most evaluations have focused on the use case where CAD software entirely replaces human readers. Evidence on alternate CAD software deployment models for TB screening and head-to-head comparisons remains limited.
Methods
A test library of 1466 CXR images was curated after an active case finding initiative in Ho Chi Minh City, Viet Nam. Images were processed by Lunit INSIGHT CXR CAD software and independently interpreted by 10 human readers across two rounds separated by a wash-out period. Five CXR interpretation methods were simulated: human-only reading, CAD-only reading, CAD-assisted human reading, independent double reading and sequential reading. Generalised estimating equations were used to estimate adjusted incidence rate ratios comparing the sensitivity and specificity of CAD-containing CXR interpretation methods with human-only reading, while operational performance metrics were summarised as percentage changes relative to human-only reading.
Results
Human-only reading achieved mean sensitivity 79.0% and specificity 84.8%. CAD-only reading significantly increased sensitivity (91.4%, adjusted incidence rate ratio (aIRR) 1.16, 95% CI 1.06 to 1.27, p=0.002) and significantly reduced specificity (81.9%, aIRR 0.96, 95% CI 0.94 to 0.98, p<0.001). Independent double reading achieved the highest sensitivity (95.0%, aIRR 1.20, 95% CI 1.12 to 1.30, p<0.001), but the lowest specificity (74.5%, aIRR 0.87, 95% CI 0.86 to 0.89, p<0.001) and increased diagnostic testing by +59.2% and the number needed to test by +32.3%. Sequential reading maintained sensitivity comparable to human-only reading (78.8%, aIRR 1.00, 95% CI 0.99 to 1.00, p=0.326), while achieving the highest specificity (90.3%, aIRR 1.07, 95% CI 1.06 to 1.07, p<0.001), reducing human workload by −67.5%, and lowering the number needed to test by −29.9%. CAD-assisted human reading produced non-significant sensitivity gains (82.9%, aIRR 1.05, 95% CI 1.00 to 1.11, p=0.057) and reduced specificity (82.2%, aIRR 0.97, 95% CI 0.96 to 0.98, p<0.001), with substantial inter-reader variability. Sensitivity gains were greatest among participants without prior TB, and specificity improvements were most pronounced among less experienced readers.
Conclusion
No single CAD software use case optimally balanced accuracy and operational efficiency. CAD-only and independent double reading maximised TB detection, while sequential reading improved efficiency and conserved resources. CAD software deployment should align with local workforce capacity, diagnostic infrastructure and programmatic priorities. Prospective evaluations of these use cases are warranted.
Keywords: Tuberculosis, Public Health, Digital Technology, Digital Health
WHAT IS ALREADY KNOWN ON THIS TOPIC
Human interpretation of chest X-rays (CXRs) for tuberculosis (TB) screening is variable and influenced by reader training, experience and workload constraints; computer-aided detection (CAD) software can achieve accuracy comparable to human readers.
Most CAD software evaluations in TB have focused on accuracy benchmarking or on use cases.
WHAT THIS STUDY ADDS
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
Introduction
Tuberculosis (TB) remains the leading infectious cause of death globally, yet over 20% of people with TB remain undiagnosed,1 while those who are diagnosed often experience prolonged delays,2 contributing to continued transmission and poor health outcomes. In recent years, efforts to strengthen health systems and expand access to more sensitive diagnostic tools have helped narrow the detection gap. Chest X-ray (CXR) screening has played an important role in this process. It has re-emerged as a rapid, non-invasive tool to assess large numbers of people and indicate them for molecular diagnostic testing.3
Although CXR screening for TB is now widely used, its performance can vary.4 Participant characteristics such as young age5 and comorbidities, including HIV6 and diabetes,7 can influence the radiographic presentation of TB and may compromise interpretation accuracy when these characteristics are unknown or atypical presentations are unfamiliar to the reader. Reader capacity is also a persistent programmatic constraint. In many high TB burden settings, access to qualified human readers is limited.8 As a result, in routine practice CXR images are interpreted by a broad range of healthcare providers, including radiologists, pulmonologists, TB clinicians, general practitioners and sometimes radiographers or technicians. Their performance can vary considerably and is known to be influenced by differences in training, clinical background, years of experience and workload.9
In response, computer-aided detection (CAD) software platforms have been developed to standardise CXR interpretation. A growing body of evidence indicates that CAD software can achieve accuracy comparable to or exceeding that of human readers.10,11 In light of this evidence, the WHO has endorsed the use of multiple CAD software platforms ‘in place of human readers’ for CXR interpretation during TB screening.4,12 However, entirely replacing human readers raises important ethical considerations. From the patient perspective, the absence of human oversight may raise concerns about legal or financial liability in settings without clear regulatory frameworks13 and how radiological findings are contextualised.14 It may also limit the detection and referral of incidental or non-TB abnormalities, although CAD models capable of supporting broader diagnostic assessments are currently being developed and piloted.15 Among healthcare workers, concerns may also arise about deskilling and job displacement.16
Despite these ethical concerns, most studies have primarily evaluated CAD software for TB screening as a replacement for human readers, focusing on threshold optimisation and benchmarking against human performance. This framing does not capture the full range of ways CAD software may be used in real-world settings. The WHO’s CAD software operational guidance explicitly acknowledges non-replacement use cases,17 and outside the field of TB, various models for integrating CAD software into clinical workflows have been well described.18–20 Although terminology is not standardised, these non-replacement use cases include: CAD-assisted reading, where software outputs are used in real time by a human during CXR review to support identification of TB-related abnormalities; independent double reading, where a human reader and CAD software review CXRs in parallel to detect abnormalities missed by either; and sequential reading, where CAD software identifies and triages normal CXRs away from human review to reduce workloads (figure 1).
Figure 1. Flow diagrams illustrating five simulated CXR interpretation methods for TB screening. CAD, computer-aided detection; CXR, chest X-ray; TB, tuberculosis.
Few studies have described or evaluated these non-replacement use cases for CAD software during TB screening,21–25 and even fewer have directly compared multiple use cases in head-to-head evaluations.26 This evidence gap is significant. For TB implementers, the key question is no longer whether CAD software works, but how it should be implemented. For governments, CAD software use cases that improve CXR interpretation accuracy while enabling more efficient use of resources may be particularly valuable in an era of rising healthcare costs and evolving donor funding.27 This study aims to compare the accuracy and operational performance of five CXR interpretation methods, to inform context-specific deployments of CAD software during TB screening.
Methods
Study participants
A test library of CXR images was developed using data from an active TB case finding (ACF) initiative in Ho Chi Minh City, Viet Nam. The ACF implementation and test library curation methods have been described elsewhere in detail.28 In brief, local health authorities mobilised at-risk community members to attend 133 ACF events, including people aged ≥55 years, household or close contacts of people with TB, people with a previous history of TB, people living with HIV and/or diabetes, people with TB-related symptoms and people living in socioeconomically disadvantaged areas. Each participant was screened using CXR, with images undergoing independent double reading by an on-site radiologist and qXR CAD software (Qure.ai, India), using the manufacturer-default abnormality score threshold of ≥0.50. Individuals with any abnormal result from either CXR interpretation were asked to provide a single sputum specimen for testing with the Xpert MTB/RIF Ultra (Xpert-Ultra) assay. The qXR CAD software was used only as an independent second reader during the source ACF activities and was not evaluated in the present study because its outputs contributed to determining which participants underwent diagnostic testing, and consequently, to the construction of the composite reference standard used for this study’s test library.
After the ACF events were fully implemented, a series of event-level and participant-level inclusion criteria were applied to the dataset to curate the test library. These criteria ensured that the library represented the demographics of the screened population and that all participants could be consistently assigned a composite reference standard for TB diagnosis. Individuals with abnormal CXR and positive Xpert-Ultra results were classified as having TB, while those with abnormal CXR and negative Xpert-Ultra results, as well as those with two normal CXR results (from both the radiologist and the qXR CAD software), were classified as not having TB. The final test library comprises 1466 adults, of whom 58 (4.0%) were diagnosed with TB.
CAD software processing and CXR reading
The test library’s de-identified CXR images were shared with Lunit (South Korea) for processing with their INSIGHT CXR CAD software (V.3.1.4.4); continuous abnormality scores and corresponding heat maps and segmentation overlays were returned. The CXR images were also interpreted by 10 human readers, from five public and five private health facilities in Ho Chi Minh City, across two rounds of reading. The professional backgrounds and experience levels of the 10 readers varied, but all regularly interpreted CXR images as part of their routine clinical duties. Approximately half had obtained a formal radiology certification within the past 15 years (Online supplemental table S1).
In the first round, the readers interpreted each CXR image using a four-category reporting scale: (1) TB-related abnormality in the lungs, (2) other abnormality in the lungs, (3) abnormality outside the lungs or (4) normal. They were blinded to participant clinical data and TB diagnoses and did not have access to INSIGHT CXR CAD software outputs. After a 6-week wash-out period, the study identifiers associated with each CXR image were replaced with new, randomly-generated identifiers. Then a second round of CXR reading was conducted using the same four-category reporting scale. The readers remained blinded to participant clinical data and TB diagnoses but were instructed to review the corresponding INSIGHT CXR CAD software outputs to support identification of TB-related abnormalities. Readers were allowed to assign a final CXR interpretation that differed from the CAD software outputs in either direction (eg, classifying a CXR as abnormal when the CAD software indicated it was normal or vice versa).
CXR interpretation scenarios
We compared five CXR interpretation methods for TB screening (figure 1). CXR screening outcomes were simulated by algorithmically re-classifying participants using the results from the first or second round of human reading and/or INSIGHT CXR CAD software outputs according to the following predefined decision rules. For each method, all participants with a positive CXR screen were assumed to undergo Xpert-Ultra testing, with results being determined according to the previously described composite reference standard for TB diagnosis.
Human-only reading: CXR screening outcomes were based on the first round of unaided reading. A positive screen was defined as a ‘TB-related abnormality in the lungs’.
CAD-only reading: CXR screening outcomes were based solely on the INSIGHT CXR CAD software outputs. A positive CXR screen was defined as an abnormality score ≥33. This CAD software threshold was determined using the study data set rather than a separate training or calibration data set by evaluating sensitivity across multiple possible score thresholds and selecting the highest to achieve the minimum Target Product Profile (TPP) criteria for a high-sensitivity (≥90%) TB screening test.29 Threshold selection was based solely on sensitivity; other performance metrics were not considered.
CAD-assisted human reading: CXR screening outcomes were based on the second round of reading, during which human readers had access to the INSIGHT CXR CAD outputs as a decision-support tool. A positive CXR screen was defined as a classification of ‘TB-related abnormality in the lungs’.
Independent double reading: CXR screening outcomes were based on the data from both the first round of unaided reading and INSIGHT CXR CAD software outputs. A positive CXR screen was defined as either a human classification of ‘TB-related abnormality in the lungs’ and/or an abnormality score ≥33 (same threshold as Method 2).
Sequential reading: CXR screening outcomes were determined using a workflow in which all images were initially processed with the INSIGHT CXR CAD software. Images with an abnormality score <20 were classified as normal and did not undergo human review; images with an abnormality score ≥20 were subsequently reviewed by a human, using results from the first round of unaided reading. A positive CXR screen required both an abnormality score ≥20 and a subsequent human classification of ‘TB-related abnormality in the lungs’. Unlike CAD-only and independent double reading, where the ≥33 abnormality score threshold is used to identify people in need of diagnostic testing, in sequential reading the CAD software acts as a pre-screen to reduce human reader workload. A separate CAD software threshold was identified for this use case, by examining sensitivity across possible score thresholds within the study dataset. The ≥20 abnormality score threshold corresponds to the highest rounded score to achieve a <3% loss in sensitivity. This loss in TB detection was deemed acceptable for this study, based on the non-inferiority margin defined in a recent TB diagnostic testing evaluation.30
Statistical analysis
Characteristics of the test library were summarised using descriptive statistics. Differences in TB detection rates were assessed using the χ2 or Mann-Whitney U test, as appropriate. Sensitivity and specificity for each CXR interpretation method were calculated for the 10 readers individually and then averaged across readers. Generalised estimating equations (GEEs) were used to estimate adjusted incidence rate ratios (aIRRs) for pairwise comparisons of sensitivity and specificity between the CAD-containing CXR interpretation methods and human-only reading, adjusting for participant characteristics (eg, sex, age group, symptom status, contact history, diabetes, HIV status and previous history of TB) and accounting for clustering of observations.31 Subgroup analyses examined differences in CXR interpretation method performance across selected participant (eg, sex, age group, symptom status and history of TB) and reader characteristics (eg, working in a public vs private health facility and level of experience (<5 years vs ≥5 years)) using GEEs. Pairwise comparisons were also estimated within each subgroup, relative to human-only reading. Operational performance metrics were calculated for each CXR interpretation method for the 10 readers individually and averaged across readers. These included the number of CXRs requiring human interpretation, the number of individuals eligible for diagnostic testing, the number of people diagnosed with TB and false-negative rate and the number needed to test (defined as the number of Xpert-Ultra tests performed divided by the number of individuals diagnosed with TB). All statistical analyses were performed using Stata V.17.0 (StataCorp). Statistical significance was defined as a two-sided p value <0.05.
Results
Demographics
In the study’s test library, TB detection rates were higher among males, those without health insurance and individuals reporting TB symptoms, a past history of TB and contact exposure (table 1). No significant differences in TB detection rates were observed by age or diabetes comorbidity, while people living with HIV had a significantly lower rate of TB. Inconsistencies between the TB detection patterns in this test library and the published literature likely reflect the participant mobilisation strategies used for ACF, rather than true differences in the epidemiology of TB in Viet Nam.
Table 1. Description of the test library’s participants.
| All participants | TB | No TB | P value | |
|---|---|---|---|---|
| All participants | 1466 | 58 (4.0%) | 1408 (96.0%) | N/A |
| Demographic factors | ||||
| Sex | ||||
| Male | 825 (56.3%) | 44 (5.3%) | 781 (94.7%) | 0.002 |
| Female | 641 (43.7%) | 14 (2.2%) | 627 (97.8%) | |
| Age, median (IQR) | 49 (38–61) | 49 (39–58) | 49 (38–61.5) | 0.875 |
| 18–54 years | 884 (60.3%) | 36 (4.1%) | 848 (95.9%) | 0.779 |
| ≥55 years | 582 (39.7%) | 22 (3.8%) | 560 (96.2%) | |
| Lacking health insurance | 141 (9.6%) | 11 (7.8%) | 130 (92.2%) | 0.014 |
| Presence of TB symptoms | ||||
| Cough (any duration) | 251 (17.1%) | 37 (14.7%) | 214 (85.3%) | <0.001 |
| Fever | 39 (2.7%) | 14 (35.9%) | 25 (64.1%) | <0.001 |
| Weight loss | 58 (4.0%) | 15 (25.9%) | 43 (74.1%) | <0.001 |
| Night sweats | 20 (1.4%) | 6 (30.0%) | 14 (70.0%) | <0.001 |
| ≥1 reported symptom from above | 283 (19.3%) | 41 (14.5%) | 242 (85.5%) | <0.001 |
| TB risk factors | ||||
| Contact of person with TB | 475 (32.4%) | 41 (8.6%) | 434 (91.4%) | <0.001 |
| Past history of TB | 251 (17.1%) | 27 (10.8%) | 224 (89.2%) | <0.001 |
| Diabetes | 71 (4.8%) | 3 (4.2%) | 68 (95.8%) | 0.905 |
| HIV | 364 (24.8%) | 8 (2.2%) | 356 (97.8%) | 0.047 |
TB, tuberculosis.
Average performance relative to human-only reading
Human-only reading achieved a mean sensitivity of 79.0% and a mean specificity of 84.8% across the 10 readers (figure 2). In comparison, CAD-only reading had significantly higher mean sensitivity (aIRR 1.16, 95% CI 1.06 to 1.27, p=0.002) and significantly lower mean specificity (aIRR 0.96, 95% CI 0.94 to 0.98, p<0.001). CAD-assisted human reading resulted in a modest, non-significant increase in mean sensitivity (aIRR 1.05, 95% CI 1.00 to 1.11, p=0.057) and a significant reduction in mean specificity (aIRR 0.97, 95% CI 0.96 to 0.98, p<0.001). Independent double reading achieved the highest mean sensitivity (aIRR 1.20, 95% CI 1.12 to 1.30, p<0.001), but this gain was accompanied by the largest decline in mean specificity (aIRR 0.87, 95% CI 0.86 to 0.89, p<0.001). Sequential reading maintained a mean sensitivity comparable to human-only reading (aIRR 1.00, 95% CI 0.99 to 1.00, p=0.326), while achieving the highest mean specificity among all CXR interpretation methods (aIRR 1.07, 95% CI 1.06 to 1.07, p<0.001).
Figure 2. Pairwise comparisons of mean sensitivity and specificity for CAD-containing CXR interpretation methods relative to human-only reading. Mean sensitivity (left) and mean specificity (right) of five simulated chest X-ray interpretation methods for TB screening. Bars represent mean point estimates across all readers, with adjusted incidence rate ratios shown relative to human-only reading (reference). Pairwise comparisons are annotated with p values. Vertical lines indicate the Target Product Profile performance criteria for a high-sensitivity (≥90%) and high-specificity (≥80%) TB screening test. aIRR, adjusted incidence rate ratios; CAD, computer-aided detection; CXR, chest X-ray; TB, tuberculosis.
Individual reader variation
Individual human-only reader performance varied substantially, with sensitivities ranging from 55.2% to 91.4% and specificities from 50.9% to 95.7% (online supplemental table S2). This variability was eliminated with CAD-only reading, reflecting complete standardisation of CXR interpretation. CAD-assisted human reading increased variability in sensitivity (55.2%–98.3%) while reducing variability in specificity (59.7%–92.8%). Independent double reading narrowed sensitivities to a uniformly high level (93.1%–96.6%), while specificity variability was modestly reduced relative to human-only reading and remained greater than that observed with all other CAD-containing CXR interpretation methods (42.5%–81.4%). In contrast, sequential reading largely preserved reader-specific sensitivity variation (53.4%–91.4%), but produced the highest and most consistent specificity across human-involved CXR interpretation methods (83.2%–96.0%).
Subgroup analyses
Across all CXR interpretation methods, sensitivity did not differ significantly across participant or reader characteristics (online supplemental tables S3-1, S3-2, S3-3, S3-4 and S3-5), with the exception of higher sensitivity with independent double reading among older participants (aged ≥55 years). In contrast, specificity varied more widely across subgroups. Within-subgroup analyses revealed heterogeneous performance across the CXR interpretation methods, though clear trends emerged (online supplemental tables S4-1, S4-2, S4-3 and S4-4). Sensitivity gains relative to human-only reading were most pronounced among participants with no history of TB, while specificity gains were greatest among readers with less than 5 years of experience.
Operational metrics
Clear differences were observed in operational metrics across the five CXR interpretation methods (table 2; individual reader results are available in online supplemental tables S5-01, S5-02, S5-03, S5-04, S5-05, S5-06, S5-07, S5-08, S5-09 and S5-10). Three scenarios required human readers to interpret all CXR images (eg, human-only, CAD-assisted and independent double reading), whereas CAD-only reading eliminated the need for human interpretation entirely. Sequential reading represented an intermediate approach, reducing human reading workloads by −67.5%. Compared with human-only reading, eligibility for follow-on diagnostic testing increased by +14.7% with CAD-assisted human reading, by +18.2% with CAD-only reading and by +59.2% with independent double reading, but declined by −30.1% with sequential reading. Human-only reading missed 21.0% of people with TB. This proportion decreased to 17.1% (−18.9%) with CAD-assisted human reading, 8.6% (−59.0%) with CAD-only reading and 5.0% (−76.2%) with independent double reading, but remained essentially unchanged at 21.2% (+0.8%) with sequential reading. Human-only reading had a number needed to test of 5.7; this increased by +2.2% with CAD-only reading, by +9.3% with CAD-assisted human reading and by +32.3% with independent double reading, but declined by −29.9% with sequential reading.
Table 2. Operational performance metrics for CAD-containing CXR interpretation methods relative to human-only reading, sum of all readers.
| CXRs requiring human review n; Δ vs human-only |
Eligible for diagnostic testing n (%); Δ vs human-only |
TB detected n |
Missed TB (false negative rate) n (%); Δ vs human-only |
Number needed to test n; Δ vs human-only |
|
|---|---|---|---|---|---|
| Human-only reading | 14 660; Ref | 2605 (17.8%); Ref | 458 | 122 (21.0%); Ref | 5.7; Ref |
| CAD-only reading | 0; −100% | 3080 (21.0%); +18.2% | 530 | 50 (8.6%); −59.0% | 5.8; +2.2% |
| CAD-assisted human reading | 14 660; 0% | 2989 (20.4%); +14.7% | 481 | 99 (17.1%); −18.9% | 6.2; +9.3% |
| Independent double reading | 14 660; 0% | 4146 (28.3%); +59.2% | 551 | 29 (5.0%); −76.2% | 7.5; +32.3% |
| Sequential reading | 4760; −67.5% | 1822 (12.4%); −30.1% | 457 | 123 (21.2%); +0.8% | 4.0; −29.9% |
CAD, computer-aided detection; CXR, chest X-ray; n, number; Ref, reference; TB, tuberculosis; Δ vs human-only, relative percent change compared with human-only reading.
Discussion
This comparative evaluation of simulated CXR interpretation for TB screening highlights clear differences in accuracy and operational performance across methods. Relative to human-only reading, CAD-only and independent double reading achieved the highest sensitivities, but at the cost of increased follow-on diagnostic testing and higher numbers needed to test. In contrast, sequential reading maintained the same relatively low sensitivity as human-only reading, while achieving a lower number needed to test and large reductions in human reader workload. CAD-assisted human reading produced more modest and inconsistent gains, reflecting reader-dependent variation in how INSIGHT CXR CAD software outputs were taken into account during CXR interpretation decision-making. The sensitivity gains with CAD-containing CXR interpretation were concentrated among individuals without a prior history of TB, while specificity improvements were greatest among less experienced readers.
These findings indicate that a CXR interpretation strategy should be selected to match programme priorities, rather than being applied as a single standard approach. CAD-only and independent double reading may be most appropriate for screening projects and programmes aiming to maximise TB detection and willing to accommodate higher volumes of diagnostic testing. These conditions are well suited to ACF initiatives and prevalence surveys, which are often supported by international donors, but may also be appropriate for screening programmes where the consequences of missed diagnoses are particularly severe, such as drug-resistant TB contact investigations. In contrast, sequential reading and CAD-assisted reading may be better suited to settings where laboratory capacity, workforce availability and budgets constrain the scale-up of screening, as reductions in testing volumes and reader workload may translate into lower programme costs, although this requires evaluation through formal economic analyses. These more resource-efficient approaches may be particularly relevant for domestically financed programmes where budgets are often scrutinised and politicised. Policymakers considering the implementation of CAD software for TB screening should therefore weigh the trade-offs between maximising TB detection and minimising downstream burdens during decision making.
Human reader performance in this study was highly variable, and on average, fell well below the TPP criteria for a high-sensitivity (≥90%) and high-specificity (≥80%) TB screening test.29 Only one of 10 readers met these performance standards. A key strength of this study is the inclusion of 10 readers from both public and private clinics across Ho Chi Minh City. Although CXR images were interpreted under study conditions, these 10 readers better reflect the level of service provision available across the city. Previous CAD software evaluations from this setting have used experienced radiologists as comparators, even though their services are limited to specialised facilities. These results highlight the need for improved quality and consistency in CXR interpretation, which may be achieved through the use of CAD software, particularly as CXR screening is already being implemented at scale in Vietnam.
CAD-only reading was the sole CXR interpretation method to meet TPP performance standards, adding to the growing evidence base demonstrating strong standalone performance of INSIGHT CXR CAD software for TB screening. However, implementation of CAD-only reading raises important ethical and operational implications. As CAD software is a relatively new technology in most high TB burden settings, the acceptability of CAD-only reading is incompletely understood among programme beneficiaries and healthcare providers, including whether this approach aligns with their values and preferences for healthcare delivery. Reliance on CAD software alone may also increase the risk of interrupted CXR screening services in the event of software, hardware or connectivity failures, particularly during community-based ACF. However, in settings where access to trained human readers is severely limited, the benefits of CAD-only reading may easily justify these trade-offs. In contrast, the three non-replacement CAD software use cases evaluated in this study may offer more acceptable CAD software deployment models in settings with fewer human resource constraints. These use cases combine the consistency and performance of the CAD software with the contextual judgement, accountability and service continuity provided by human readers.
Although independent double reading is a recognised quality assurance standard in diagnostic imaging, it is rarely feasible in routine TB screening due to workforce and cost limitations.32 CAD software may offer a scalable alternative for identifying CXR abnormalities missed by initial human reading. However, discordance between human and CAD software interpretations requires careful integration into clinical workflows. In this study, double reading with CAD software increased the number of people eligible for diagnostic testing by +59.2% and the number needed to test by +32.3% compared with human-only reading. These findings suggest that human readers in this study both over-read and under-read CXR images. Over-reading by humans which does not ultimately yield TB diagnoses is often accepted, despite being an inefficient use of resources. In contrast, CAD-driven referrals may be perceived as excessive because they challenge established norms of human judgement and could strain laboratory capacity when added to existing human referrals, even if they result in additional TB detection. Moreover, transparent and continuous performance feedback from CAD software may alter longstanding norms of professional autonomy, by making diagnostic discrepancies more visible and attributable at the individual level. Unlike mammography screening programmes, where double human reading is already standard practice and programmes are resourced to support two reads, TB screening typically relies on a single human reader. In mammography, CAD software has been used to replace the second human read,33 whereas introducing double reading into TB screening (whether through a second human reader or a CAD software) would represent an additional workflow and may therefore require additional resources. Therefore, future studies should evaluate the clinical, operational and economic implications of double reading with CAD software, including its impact on laboratory capacity and diagnostic yield, its acceptability among beneficiaries and health workers and its cost-effectiveness.
Sequential reading maintained comparable sensitivity to human-only reading while reducing interpretation workload by more than two thirds, making it well suited for high-volume screening settings where experienced readers are available but capacity is insufficient to meet workload guidelines.34 Because only CAD-flagged CXRs undergo human review, the sensitivity of this approach is determined largely by human reader expertise, limiting its ability to realise the full sensitivity gains achievable with a CAD software. The one-third reduction in eligibility for diagnostic testing observed in this study likely reflects the INSIGHT CXR CAD software correctly classifying normal CXR images that were over-read during human-only reading; this effect would be expected to diminish when using more experienced readers. Taken together these findings suggest that the performance of sequential reading will be shaped by the level and distribution of human reader expertise within the available workforce. However, interpretation of these findings requires caution since readers did not prospectively interpret a restricted subset of CAD-flagged CXR images in this study. Awareness that a CXR has been classified as potentially abnormal by a CAD software may increase the likelihood of over-reading or improve detection of subtle abnormalities. The net effect of these behavioural responses cannot be inferred from the study’s simulated design. Therefore, prospective implementation studies comparing human-only reading and real-time sequential reading with CAD software during TB screening are needed to more accurately estimate this approach’s relative diagnostic performance and operational impact.
CAD-assisted human reading produced limited performance improvements, with substantial variation across readers. Some readers achieved clear gains in sensitivity, while others showed little or no improvement, despite having received the same training and access to detailed heat maps and segmentation overlays that highlighted areas of radiographic interest. The reasons for inconsistent engagement with the CAD software outputs were outside of the scope of this study but may be related to several factors. CAD software is still a relatively new technology in TB screening, and some clinicians may view it with scepticism, particularly if it is perceived as a threat to their professional roles. In addition, this was the first time the study’s readers had used the INSIGHT CXR CAD software in a clinical decision making context, and the training period may have been insufficient for them to integrate the software outputs into their established interpretation workflows. With longer exposure, ongoing performance feedback and a clearer framing of the CAD software as a supportive tool rather than a replacement for clinical judgement, the readers may engage with the software outputs more effectively. These findings highlight the importance of adequate training and user trust in shaping the adoption of this technology and approach, and warrant further prospective assessments.
This study provides a robust comparison of CXR interpretation methods for TB screening, but the generalisability of the findings may be limited by its setting and design. The test library data were collected in a single geography and included only adult participants. The library also had a modest sample size of just 58 people with TB, limiting the insights that could be drawn from subgroup analyses. In addition, the test library was curated from an ACF initiative, and thus likely includes a higher proportion of people with low-grade CXR abnormalities and paucibacillary TB disease compared with facility-based TB screening. Although the use of a composite reference standard can be considered a methodological strength of this study, some individuals with TB may have been excluded from TB testing during ACF due to very low-grade CXR abnormalities missed by both the on-site radiologist and qXR CAD software. Additional people with TB may have been detected during ACF if a lower CAD software abnormality threshold had been used to indicate testing, if more sensitive diagnostic approaches had been employed (eg, double Xpert-Ultra testing or culture)35 and/or if additional follow-up screening had been conducted for participants with negative Xpert-Ultra and CXR abnormal results.36–38 Importantly, the CXR interpretation methods evaluated in this study were retrospectively simulated using a test library rather than prospectively implemented as screening workflows. Consequently, the analysis may not fully capture changes in reader behaviour, workflow efficiency or other operational factors that could influence performance when these approaches are implemented under real-world conditions.
INSIGHT CXR CAD software thresholds were selected to achieve predefined sensitivity targets using the same study dataset on which performance was evaluated. Because threshold identification and evaluation were conducted with a single test library rather than separate training and evaluation test libraries, the selected thresholds may be partially tailored to the characteristics of this dataset and may not be generalisable, particularly in populations with different demographics and TB prevalence. However, this analysis compares the relative performance of different CXR interpretation methods in addition to their standalone performance. Therefore, this limitation is less likely to affect the comparative conclusions across CXR interpretation methods. This study evaluated only a single CAD product and software version; therefore, the findings are specific to INSIGHT CXR CAD software and may not reflect the performance of other CAD software products when deployed across these different use cases. These analyses also focused on only accuracy and performance metrics, and did not evaluate downstream outcomes, such as diagnostic delays, treatment initiation, patient outcomes or cost-effectiveness, all of which are likely to be critical determinants of CAD software uptake.
Conclusion
In this comparative evaluation of five CXR interpretation methods for TB screening, we found that no single approach optimally balanced accuracy and operational efficiency across all metrics, but that different CAD software use cases could be more optimally deployed to meet specific public health objectives rather than a single deployment model being universally applied. Programmes seeking to maximise TB detection may favour CAD-only or independent double reading, whereas those facing workforce or laboratory constraints may benefit more from sequential workflows. Policy decisions should therefore consider local epidemiology, reader expertise and availability, diagnostic capacity and the trade-offs between sensitivity, specificity, workload and potential cost. Prospective implementation studies on non-replacement use cases are needed to validate these simulated findings under real-world conditions and to assess downstream impacts on patient outcomes and programme effectiveness, as well as acceptability among beneficiaries and providers.
Supplementary material
Acknowledgements
We are grateful to the Viet Nam National Tuberculosis Control Programme, Pham Ngoc Thach Hospital and the District TB Units in Districts 6, 8, Binh Chanh and Thu Duc for their collaboration and support in implementing the ACF events. We also extend our sincere appreciation to the more than 25 000 community members who participated in those ACF events. In addition, we thank Lunit for providing access to their INSIGHT CXR CAD software to process the study’s test library at no cost.
Footnotes
Funding: USAID supported the ACF events which were the source data for this evaluation’s test library (agreement number: 72044020FA00001). Additional funding for CXR reading and analysis was provided by the Stop TB Partnership’s TB REACH initiative (agreement number: STBP/TBREACH/GSA/W9-9110). These funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.
Provenance and peer review: Not commissioned; externally peer reviewed.
Patient consent for publication: Not applicable.
Data availability free text: Data are available from the corresponding author upon reasonable request.
Patient and public involvement: Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Ethics approval: The implementation of the ACF initiative (July 2022 to March 2023) which was the source data for this study’s test library was approved by the Ho Chi Minh City People’s Committee (Decision No. 2681/QĐ-UBND, 29 July 2020) in accordance with Vietnamese Ministry of Health’s national TB guidelines (Decision No. 1314/QĐ-BYT, 24 March 2020). The protocol for this de-identified, retrospective CXR interpretation evaluation was subsequently reviewed and approved by the Pham Ngoc Thach Hospital Institutional Review Board (Decision No. 1688/HĐĐĐ-PNT, 6 September 2023). All study procedures adhered to the ethical principles outlined in the Declaration of Helsinki and complied with local regulatory requirements.
Data availability statement
Data are available upon reasonable request.
References
- 1.Geneva, Switzerland: World Health Organization; 2025. Global tuberculosis report 2025. [Google Scholar]
- 2.Teo AKJ, Singh SR, Prem K, et al. Duration and determinants of delayed tuberculosis diagnosis and treatment in high-burden countries: a mixed-methods systematic review and meta-analysis. Respir Res. 2021;22:251. doi: 10.1186/s12931-021-01841-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Miller C, Lonnroth K, Sotgiu G, et al. The long and winding road of chest radiography for tuberculosis detection. Eur Respir J. 2017;49:1700364. doi: 10.1183/13993003.00364-2017. [DOI] [PubMed] [Google Scholar]
- 4.WHO Consolidated Guidelines on Tuberculosis. Module 2: Screening – Systematic Screening for Tuberculosis Disease. Geneva, Switzerland: World Health Organization; 2021. [PubMed] [Google Scholar]
- 5.Pillay T, Andronikou S, Zar HJ. Chest imaging in paediatric pulmonary TB. Paediatr Respir Rev. 2020;36:65–72. doi: 10.1016/j.prrv.2020.10.002. [DOI] [PubMed] [Google Scholar]
- 6.Frey V, Phi Van VD, Fehr JS, et al. Prospective evaluation of radiographic manifestations of tuberculosis in relationship with CD4 count in patients with HIV/AIDS. Medicine (Baltimore) 2023;102:e32917. doi: 10.1097/MD.0000000000032917. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Zafar MI, Chen L-L, Xiaofeng Y, et al. Impact of Diabetes Mellitus on Radiological Presentation of Pulmonary Tuberculosis in Otherwise Non-Immunocompromised Patients: A Systematic Review. Curr Med Imaging Rev. 2019;15:543–54. doi: 10.2174/1573405614666180806124416. [DOI] [PubMed] [Google Scholar]
- 8.Hricak H, Abdel-Wahab M, Atun R, et al. Medical imaging and nuclear medicine: a Lancet Oncology Commission. Lancet Oncol. 2021;22:e136–72. doi: 10.1016/S1470-2045(20)30751-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Brady AP. Error and discrepancy in radiology: inevitable or avoidable? Insights Imaging. 2017;8:171–82. doi: 10.1007/s13244-016-0534-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Worodria W, Castro R, Kik SV, et al. An independent, multi-country head-to-head accuracy comparison of automated chest x-ray algorithms for the triage of pulmonary tuberculosis. medRxiv. 2024 doi: 10.1101/2024.06.19.24309061. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Qin ZZ, Van der Walt M, Moyo S, et al. Comparing the accuracy of computer-aided detection (CAD) software and radiologists from multiple countries for tuberculosis detection in chest X-Rays. Sci Rep . 2025;15:22540. doi: 10.1038/s41598-025-06164-w. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Geneva, Switzerland: World Health Organization; 2025. Use of computer-aided detection software for tuberculosis screening: WHO policy statement. [Google Scholar]
- 13.Hwang EJ, Goo JM, Park CM. AI Applications for Thoracic Imaging: Considerations for Best Practice. Radiology. 2025;314:e240650. doi: 10.1148/radiol.240650. [DOI] [PubMed] [Google Scholar]
- 14.Udupa S, Sharma O, Sharma S, et al. Patient Perceptions of the Use of Artificial Intelligence in Radiology: A Scoping Review. Acad Radiol. 2026;33:774–93. doi: 10.1016/j.acra.2026.01.001. [DOI] [PubMed] [Google Scholar]
- 15.Sierra JFT, Herrera JAC, Sutachan MAT, et al. Innovations in artificial intelligence to minimize diagnostic error - a comparison with human interpretation of chest radiographs in the clinical context: a scoping review. Health Technol. 2025;15:1021–36. doi: 10.1007/s12553-025-00999-z. [DOI] [Google Scholar]
- 16.Akudjedu TN, Torre S, Khine R, et al. Knowledge, perceptions, and expectations of Artificial intelligence in radiography practice: A global radiography workforce survey. J Med Imaging Radiat Sci. 2023;54:104–16. doi: 10.1016/j.jmir.2022.11.016. [DOI] [PubMed] [Google Scholar]
- 17.WHO Operational Handbook on Tuberculosis. Module 3: Diagnosis - Rapid Diagnostics for Tuberculosis Detection, 2021 Update. Geneva, Switzerland: World Health Organization; 2021. [Google Scholar]
- 18.Tang A, Tam R, Cadrin-Chênevert A, et al. Canadian Association of Radiologists White Paper on Artificial Intelligence in Radiology. Can Assoc Radiol J. 2018;69:120–35. doi: 10.1016/j.carj.2018.02.002. [DOI] [PubMed] [Google Scholar]
- 19.Hwang EJ, Park CM. Clinical Implementation of Deep Learning in Thoracic Radiology: Potential Applications and Challenges. Korean J Radiol. 2020;21:511–25. doi: 10.3348/kjr.2019.0821. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Creswell J, Vo LNQ, Qin ZZ, et al. Early user perspectives on using computer-aided detection software for interpreting chest X-ray images to enhance access and quality of care for persons with tuberculosis. BMC Glob Public Health . 2023;1:30. doi: 10.1186/s44263-023-00033-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Vijayan S, Jondhale V, Pande T, et al. Implementing a chest X-ray artificial intelligence tool to enhance tuberculosis screening in India: Lessons learned. PLOS Digit Health . 2023;2:e0000404. doi: 10.1371/journal.pdig.0000404. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Innes AL, Martinez A, Gao X, et al. Computer-Aided Detection for Chest Radiography to Improve the Quality of Tuberculosis Diagnosis in Vietnam’s District Health Facilities: An Implementation Study. Trop Med Infect Dis . 2023;8:488. doi: 10.3390/tropicalmed8110488. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Innes AL, Lebrun V, Hoang GL, et al. An Effective Health System Approach to End TB: Implementing the Double X Strategy in Vietnam. Glob Health Sci Pract. 2024;12:e2400024. doi: 10.9745/GHSP-D-24-00024. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Innes AL, Martinez A, Hoang GL, et al. Active case finding to detect symptomatic and subclinical pulmonary tuberculosis disease: implementation of computer-aided detection for chest radiography in Viet Nam. Western Pac Surveill Response J. 2024;15:1–12. doi: 10.5365/wpsar.2024.15.4.1118. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Munjal P, Mahrooqi AA, Rajan R, et al. Population-scale cross-sectional observational study for AI-powered TB screening on one million CXRs. NPJ Digit Med . 2025;8:418. doi: 10.1038/s41746-025-01832-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Park J, Hwang EJ, Lee JH, et al. Identification of Active Pulmonary Tuberculosis Among Patients With Positive Interferon-Gamma Release Assay Results: Value of a Deep Learning-based Computer-aided Detection System in Different Scenarios of Implementation. J Thorac Imaging. 2023;38:145–53. doi: 10.1097/RTI.0000000000000691. [DOI] [PubMed] [Google Scholar]
- 27.Apeagyei AE, Bisignano C, Elliott H, et al. Tracking development assistance for health, 1990-2030: historical trends, recent cuts, and outlook. Lancet. 2025;406:337–48. doi: 10.1016/S0140-6736(25)01240-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Codlin AJ, Dao TP, Nguyen BH, et al. Bridging the analog divide: a comparison of printed X-ray films and digital images when using computer-aided detection software for tuberculosis screening. BMC Glob Public Health . 2026;4:6. doi: 10.1186/s44263-025-00237-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Geneva, Switzerland: World Health Organization; 2025. Target product profiles for tuberculosis screening tests. [Google Scholar]
- 30.Moe CA, Luswata RK, Barrameda AJ, et al. Diagnostic yield of tongue swab- compared to sputum-based molecular testing for tuberculosis in four high-burden countries. Infectious Diseases (except HIV/AIDS) 2025 doi: 10.1101/2025.07.03.25330836. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31.Genders TSS, Spronk S, Stijnen T, et al. Methods for calculating sensitivity and specificity of clustered data: a tutorial. Radiology. 2012;265:910–6. doi: 10.1148/radiol.12120509. [DOI] [PubMed] [Google Scholar]
- 32.Geijer H, Geijer M. Added value of double reading in diagnostic radiology,a systematic review. Insights Imaging. 2018;9:287–301. doi: 10.1007/s13244-018-0599-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 33.Dembrower K, Crippa A, Colón E, et al. Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study. Lancet Digit Health. 2023;5:e703–11. doi: 10.1016/S2589-7500(23)00153-X. [DOI] [PubMed] [Google Scholar]
- 34.Brady AP, Loewe C, Brkljacic B, et al. Guidelines and recommendations for radiologist staffing, education and training. Insights Imaging. 2025;16:57. doi: 10.1186/s13244-025-01926-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Floyd S, Klinkenberg E, de Haas P, et al. Optimising Xpert-Ultra and culture testing to reliably measure tuberculosis prevalence in the community: findings from surveys in Zambia and South Africa. BMJ Open. 2022;12:e058195. doi: 10.1136/bmjopen-2021-058195. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Macpherson L, Kik SV, Quartagno M, et al. Diagnostic Accuracy of Chest X-ray Computer-Aided Detection Software for Detection of Prevalent and Incident Tuberculosis in Household Contacts. Clin Infect Dis. 2025;80:626–36. doi: 10.1093/cid/ciae528. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 37.Xin H, Wang W, Wang X, et al. The performance of computer-aided detection for chest radiography in tuberculosis screening: a population-based retrospective cohort study. Emerg Microbes Infect. 2025;14:2470998. doi: 10.1080/22221751.2025.2470998. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Ruperez M, Busang J, Mureithi L, et al. Clinical outcomes of participants of a TB prevalence survey with an abnormal chest X-ray but no evidence of TB disease after a median follow-up of 9 months in Zambia and South Africa. PLOS Glob Public Health. 2025;5:e0003787. doi: 10.1371/journal.pgph.0003787. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data are available upon reasonable request.