Governing AI-Assisted Pulmonary Tuberculosis Screening in Viet Nam: Chest X-Ray Triage, Diagnostic Confirmation and Safe Follow-Up

Governing AI-Assisted Pulmonary Tuberculosis Screening in Viet Nam: Chest X-Ray Triage, Diagnostic Confirmation and Safe Follow-Up

IMPLEMENTATION INSIGHTS

Last reviewed: 29 October 2026

A 58-year-old man with diabetes attends a community health screening program in Viet Nam. He has experienced an occasional cough but has not sought medical attention because the symptoms seem mild. During the screening visit, he undergoes a digital chest X-ray.

An artificial intelligence (AI) system analyzes the image and assigns an elevated abnormality score, indicating findings that may be consistent with pulmonary tuberculosis (TB). The result suggests that further assessment is appropriate, but it does not establish that the man has TB.

The screening team now faces several decisions. Should he be referred directly for molecular testing, or should a radiologist review the image first? If the molecular test is negative but clinical concern remains, what should happen next? Who will ensure that he receives the result and completes any recommended follow-up?

Another patient presents a different challenge. She has experienced a prolonged cough and weight loss and has recently been in close contact with someone diagnosed with pulmonary TB. However, her chest X-ray receives a low AI abnormality score.

The clinical team must decide whether further investigation is warranted despite the result. Relying on the algorithm alone could create false reassurance, particularly when symptoms and exposure history suggest that TB remains possible.

These situations illustrate why AI-assisted TB screening is not simply an image interpretation task.

Computer-aided detection (CAD) software can help identify chest radiographic abnormalities associated with pulmonary TB and may reduce the workload involved in screening large populations. Its usefulness, however, depends on how the results influence clinical decisions, which people receive diagnostic testing, and whether those requiring further care are successfully followed up.

For healthcare organizations in Viet Nam, the challenge is to connect AI-assisted screening with appropriate clinical assessment, diagnostic confirmation, result communication, and subsequent management. A screening program may identify many people with abnormal chest X-rays yet achieve limited benefit if those individuals do not complete diagnostic evaluation.

The objective is not simply to generate more abnormal AI scores. It is to identify people who need further investigation, avoid inappropriate reassurance, use diagnostic resources effectively, and prevent patients from being lost between screening and care.

Why pulmonary tuberculosis screening remains important

Tuberculosis is an infectious disease caused by bacteria belonging to the Mycobacterium tuberculosis complex. It most commonly affects the lungs, although other organs can also be involved.

Pulmonary TB is particularly important for public health because people with infectious disease may transmit the bacteria through the air.

The symptoms vary. Some people experience prolonged coughing, fever, night sweats, weight loss, or fatigue. Others have mild or nonspecific symptoms, and some may not recognize any symptoms when they are first assessed.

These differences can make early detection difficult. People may delay seeking medical attention because their symptoms appear minor, diagnostic services are not easily accessible, or they are concerned about the financial and social consequences of a TB diagnosis.

Systematic screening and active case finding can help identify people who might otherwise remain undiagnosed.

The World Health Organization (WHO) recommends systematic TB screening in selected populations according to epidemiological risk, clinical circumstances, and the screening setting [1].

Priority populations include household and close contacts of people with TB, people living with HIV, people exposed to silica dust, and people in certain congregate settings. Other populations may warrant screening depending on local TB prevalence, risk factors, and access to healthcare.

Chest radiography is an important screening method because it can reveal pulmonary abnormalities in people who may have few or no recognizable symptoms. However, the findings are not specific to TB. Previous infections, other respiratory diseases, chronic lung conditions, and noninfectious disorders can produce abnormalities that resemble those associated with tuberculosis.

A normal chest X-ray also does not exclude TB in every clinical situation.

Screening must therefore remain connected to clinical assessment and appropriate diagnostic testing. Introducing AI may improve access to chest X-ray interpretation, particularly in high-volume community screening programs, but it does not remove the limitations of radiography or replace the professional decisions required to determine whether further investigation is appropriate.

What AI-assisted chest X-ray screening actually does

CAD software uses computational methods, often including deep learning, to analyze digital chest X-ray images.

For TB screening, the system identifies image patterns associated with pulmonary abnormalities that may suggest tuberculosis. Depending on the product, it produces an examination-level abnormality score and may display highlighted regions, heat maps, or other image-related information.

The score supports a screening or triage decision. A program may select a threshold above which an examination is considered screen-positive, indicating that additional diagnostic evaluation should be considered.

However, an AI abnormality score is not necessarily a calibrated probability that the person has active TB.

Its meaning depends on the software, model version, scoring system, and intended population. For example, a score of 70 should not automatically be interpreted as a 70% probability of TB, nor should it be assumed to have the same meaning across different CAD products.

This distinction matters because the score may influence whether a person undergoes molecular testing.

A false-positive result may lead to unnecessary investigations, additional appointments, costs, and anxiety. A false-negative result may delay diagnosis if healthcare professionals interpret a low score as evidence that no further assessment is needed.

Healthcare organizations should understand what the system was developed and evaluated to detect, the population in which it can be used, and the decisions its output is intended to support.

An AI-generated screening result is not equivalent to bacteriological confirmation, a clinical diagnosis, or a treatment recommendation.

CAD can support the identification of people who may need further evaluation, but the diagnosis of TB requires additional evidence and appropriate professional assessment.

What WHO guidance says about CAD for TB screening

WHO has recommended CAD as an option for interpreting digital chest radiographs during TB screening in eligible populations since 2021 [1].

The recommendation is conditional and recognizes that CAD can support high-volume screening and help address limitations in access to trained chest X-ray readers.

In June 2025, WHO published an updated policy statement following an independent evaluation of commercially available CAD products [2].

Six products met WHO's performance standards for TB screening among people aged 15 years or older.

WHO did not extend this recommendation to children younger than 15 years because the available evidence did not support the same conclusion for that age group.

This distinction is important. Children may still require TB screening and diagnostic assessment, but the evidence supporting CAD use in adults and older adolescents should not automatically be generalized to younger children.

The WHO evaluation also does not mean that the six products perform identically.

A study published in ERJ Open Research in 2026 compared eight commercial CAD products using chest X-ray datasets representing community-based and healthcare facility-based TB screening [3].

Diagnostic performance varied considerably. The area under the receiver operating characteristic curve (AUROC) ranged from 0.72 to 0.84 in community screening and from 0.77 to 0.86 in facility-based screening.

At a fixed sensitivity of 90%, specificity ranged from approximately 26% to 56% in community screening and from 30% to 49% in facility-based screening.

Six products met the performance considerations supporting WHO's 2025 policy update, while two performed less favorably.

These findings demonstrate why product selection and threshold evaluation matter. Different systems can generate substantially different numbers of positive screening results, even when configured to achieve similar sensitivity.

Meeting WHO performance standards does not establish that every version of a product will perform equally well in all settings. It also does not independently confirm compliance with Vietnamese medical device or other applicable regulatory requirements.

WHO guidance supports the use of CAD for a defined TB screening purpose. It should not be interpreted as authorization for an algorithm to diagnose TB autonomously or provide a complete radiological interpretation of every abnormality visible on a chest X-ray.

What the October 2026 study in Ho Chi Minh City tells us

An important contribution to the evidence was published on October 9, 2026, in BMJ Public Health [4].

Codlin and colleagues examined how different CAD-supported chest X-ray reading strategies might influence TB screening accuracy, workload, and subsequent diagnostic testing.

The researchers developed a test library of 1,466 chest X-ray images from an active TB case-finding initiative in Ho Chi Minh City. Fifty-eight participants, representing approximately 4.0% of the library, were classified as having TB according to the study's composite reference standard.

The images were analyzed using Lunit INSIGHT CXR software, version 3.1.4.4, and independently interpreted by 10 human readers from five public and five private healthcare facilities. The readers had different professional backgrounds and levels of experience.

The investigators compared five simulated interpretation strategies: human-only reading, CAD-only reading, CAD-assisted human reading, independent double reading, and sequential reading.

The results showed important differences.

Human-only reading achieved mean sensitivity of 79.0% and specificity of 84.8%.

CAD-only interpretation achieved sensitivity of 91.4% and specificity of 81.9%, increasing sensitivity while reducing specificity compared with human-only reading.

CAD-assisted human reading achieved sensitivity of 82.9% and specificity of 82.2%. The improvement in sensitivity was not statistically significant at the conventional 5% level, and performance varied between readers.

Independent double reading achieved the highest sensitivity, at 95.0%, but specificity fell to 74.5%. Under this strategy, a positive finding from either the human reader or CAD could lead to further diagnostic testing.

Sequential reading produced a different balance. CAD first classified the images, and only those meeting a selected threshold proceeded to human interpretation.

This approach achieved sensitivity of 78.8% and specificity of 90.3%. It also reduced the simulated human reading workload by 67.5% compared with human-only reading.

The strategy achieving the highest sensitivity was therefore not the one providing the greatest reduction in workload or the highest specificity.

The researchers also examined potential consequences for diagnostic testing. Compared with human-only reading, independent double reading increased the number of people considered eligible for further testing by 59.2%, while sequential reading reduced that number by 30.1%.

These figures represent modeled eligibility for subsequent testing, not evidence that the corresponding numbers of molecular tests were actually completed in prospective clinical practice.

Sequential reading also reduced the number needed to test to identify one TB case, but it did not improve sensitivity compared with human-only reading.

These differences are highly relevant to screening programs where molecular testing capacity, staffing, and funding are limited.

However, the findings require careful interpretation.

The study was a retrospective comparative evaluation using simulated reading workflows. Participants were not prospectively assigned to receive one of the five strategies, and the analysis did not directly measure how the different approaches would affect patient care after implementation.

The CAD thresholds were also selected using the same dataset on which performance was evaluated. This may have made those thresholds particularly suited to the study population, limiting their generalizability.

Another important limitation concerns the reference standard.

During the original active case-finding program, decisions about who received molecular testing were influenced by a radiologist and a different CAD product, qXR. People whose chest X-rays were considered normal by both were not necessarily subjected to the same bacteriological testing process as those with abnormal findings.

Consequently, some people with TB may have been classified as not having the disease because their infection was not identified through the available reference process.

The study also included only 58 participants classified as having TB, limiting the precision of some comparisons.

These limitations do not invalidate the findings, but they affect how the results should be used.

The study provides valuable evidence from Viet Nam about the trade-offs between detection, workload, and diagnostic demand. It does not establish that any particular strategy will achieve the same performance if introduced prospectively into another healthcare organization.

Choosing a screening strategy is a clinical and operational decision

The Ho Chi Minh City study illustrates why healthcare organizations should decide how CAD will influence patient management before introducing the software.

A CAD-only strategy may be useful when the main challenge is limited access to trained readers or when a community screening program needs to evaluate large numbers of images.

Under appropriate conditions, CAD may substitute for human interpretation for the defined TB screening task. However, this approach requires confidence in the product's performance, selected threshold, image quality, and arrangements for subsequent assessment.

CAD-assisted human reading maintains professional interpretation while providing AI information to the reader. It may help identify subtle abnormalities, but the additional information can also influence decisions through automation bias or create uncertainty when the reader and algorithm disagree.

Independent double reading may be useful when the program gives particular priority to identifying as many potential TB cases as possible. The trade-off is greater demand for confirmatory testing.

Sequential reading may conserve human reading resources and reduce testing demand. However, people whose images fall below the CAD threshold may receive no human image review within that screening process.

This creates a different safety concern, particularly if a low score discourages further investigation in someone who has symptoms or significant risk factors.

The appropriate strategy depends on the purpose of the service.

A community active case-finding program may place greater emphasis on sensitivity because its objective is to find people who would otherwise remain undiagnosed.

A facility with limited molecular testing capacity must consider how many additional investigations the strategy will generate and whether the laboratory can provide timely results.

A hospital using chest X-rays for broader clinical evaluation also needs to consider whether a TB-focused CAD system can coexist safely with the professional interpretation required to identify other abnormalities.

These decisions cannot be made solely by comparing sensitivity, specificity, or AUROC.

Healthcare organizations need to understand how the selected strategy affects the complete screening-to-care pathway, including who receives testing, who may be missed, and whether the necessary follow-up services are available.

CAD thresholds should reflect the intended population and service

Most TB CAD applications use an abnormality score and a selected threshold to classify screening results.

The threshold affects the balance between sensitivity and specificity.

A lower threshold generally increases the proportion of people classified as screen-positive. This may help identify more potential TB cases but can also increase confirmatory testing demand.

A higher threshold may reduce unnecessary investigations while increasing the possibility that some people with TB will receive a screen-negative result.

The appropriate balance depends on the population and the objectives of the program.

A high-prevalence screening setting may produce a different diagnostic yield from a lower-prevalence population. People with previous TB may have residual abnormalities that influence CAD scores even when active disease is not present.

Other factors, including age, HIV status, concurrent lung disease, image quality, and patterns of healthcare access, may also influence observed performance.

WHO-related technical resources recognize the importance of selecting and evaluating thresholds according to the intended setting.

Healthcare organizations should therefore avoid adopting a numerical threshold simply because it appears in a vendor presentation or international study.

The thresholds examined in the October 2026 Ho Chi Minh City study were specific to the evaluated software version, study population, and simulated workflows [4].

They should not become universal clinical cutoffs for Vietnamese healthcare services.

Local evaluation should examine how threshold selection affects the number of people referred for testing, the likely number of missed cases, laboratory workload, and the consequences of incomplete follow-up.

The threshold is ultimately part of a clinical service decision, not merely a technical software setting.

A negative CAD result should not override clinical suspicion

An important safeguard is the distinction between a low AI abnormality score and a low clinical likelihood of TB.

A person may have symptoms, exposure history, or medical conditions that justify diagnostic assessment even when CAD does not detect substantial radiographic abnormalities.

This is particularly relevant to people with persistent coughing, unexplained weight loss, close contact with someone who has pulmonary TB, immunosuppression, or other circumstances increasing clinical concern.

Chest radiography cannot identify every case of pulmonary TB. Some individuals may have subtle abnormalities or apparently normal imaging despite having the disease.

CAD inherits these limitations and introduces additional uncertainty related to model performance.

A low score should therefore not be interpreted as definitive evidence that TB is absent.

The clinical pathway should distinguish routine screening from diagnostic investigation of people who already have symptoms or other clinical indications.

When TB remains clinically plausible, further investigation may be appropriate regardless of the AI classification.

The same principle applies to children younger than 15 years. The absence of a WHO recommendation for CAD screening in this age group does not remove the need for appropriate pediatric TB assessment.

An algorithm developed to prioritize pulmonary TB screening should not become a gatekeeper that prevents clinically indicated diagnostic care.

Positive screening findings require diagnostic evaluation

An elevated CAD score indicates that the chest X-ray contains patterns potentially associated with pulmonary TB.

It does not establish that Mycobacterium tuberculosis is present.

Similar radiographic findings may result from previous TB, other infections, chronic pulmonary diseases, or noninfectious conditions.

Further diagnostic assessment is therefore necessary.

WHO's consolidated TB diagnosis guidelines, published in 2025, recommend low-complexity automated nucleic acid amplification tests (NAATs) on respiratory samples as initial diagnostic tests for adults and adolescents with symptoms of TB or positive pulmonary TB screening results, rather than smear microscopy or culture as the initial test [5].

These molecular tests can help identify TB bacteria and, depending on the assay, provide information about resistance to particular anti-TB medicines.

The specific test, specimen, and subsequent diagnostic pathway depend on the patient's circumstances, available technology, and applicable clinical guidance.

A positive molecular result can provide important bacteriological evidence, but interpretation may be more complex in people with previous TB treatment, low bacterial loads, trace results, or suspected drug resistance.

A negative result also does not automatically exclude TB when clinical or radiographic suspicion remains substantial.

Depending on the circumstances, clinicians may consider additional samples, further microbiological testing, imaging review, specialist assessment, or other appropriate investigations.

WHO's operational handbook published in August 2025 provides further guidance on diagnostic algorithms, test selection, and the interpretation of discordant findings [6].

Diagnostic options also continued to develop during 2026.

In March 2026, WHO announced additional recommendations concerning near-point-of-care molecular tests, tongue swab specimens for selected adults and adolescents unable to produce sputum, and sputum pooling strategies intended to improve testing efficiency [7].

These developments may help bring diagnostic services closer to communities and primary healthcare settings.

However, the recommendations have specific limitations. Respiratory samples remain preferred when available, and the suitability of tongue swabs depends on the test and patient population. Some near-point-of-care tests identify TB without determining rifampicin resistance, requiring additional testing where drug-resistance assessment is indicated.

The recommendations should not be interpreted as evidence that every new testing method is already routinely available or authorized in Viet Nam.

For hospitals and screening programs, the distinction remains clear: CAD supports the identification of people who may need diagnostic evaluation. It does not replace the process required to establish or exclude TB disease.

Bacteriological confirmation should be pursued where appropriate, while the possibility of clinically diagnosed TB without bacteriological confirmation must also be addressed according to applicable professional guidance.

Follow-up determines whether screening produces meaningful care

A screening program may identify a person who needs further assessment without ensuring that the assessment is completed.

Someone with a positive CAD result may leave a community screening event before providing a sputum sample. Another person may complete molecular testing but never receive the result. A third may receive a diagnosis but fail to attend the designated TB treatment service.

These are different points at which the screening-to-care pathway can fail.

A reliable program needs to establish who explains the screening result, arranges diagnostic testing, reviews laboratory findings, contacts the patient, and confirms that the necessary next steps have occurred.

This becomes particularly important when screening and diagnostic services operate at different locations.

A person may undergo chest X-ray screening through a mobile community program, provide a specimen at a local health facility, and receive a laboratory result processed elsewhere.

Unless information is communicated reliably, the original screening team may not know whether testing was completed or whether the person subsequently received care.

Healthcare organizations should distinguish a referral being issued from a referral being completed.

They should also distinguish specimen collection from completion of testing, and the availability of a result from its communication and appropriate clinical interpretation.

For people diagnosed with TB, timely linkage to treatment services and applicable public health reporting arrangements is important.

The aim is not simply to increase screening volume. It is to improve the proportion of people who move through the necessary steps toward diagnosis and appropriate management.

Vietnamese guidance provides a foundation for active case finding

Viet Nam already has professional guidance addressing active TB case finding and diagnostic follow-up.

The Ministry of Health issued Decision No. 162/QĐ-BYT on January 19, 2024, promulgating guidance on TB diagnosis, treatment, and prevention [9].

The document addresses clinical and diagnostic considerations for pulmonary TB and other forms of tuberculosis, including situations in which bacteriological confirmation has not been obtained.

Another important document is Decision No. 2627/QĐ-BYT, issued on September 5, 2024 [10].

This decision promulgated guidance on active and intensified detection of TB, TB infection, and selected respiratory diseases in communities and healthcare facilities.

The guidance describes screening approaches combining clinical assessment, chest radiography, and molecular diagnostic testing, including the approach commonly referred to as the 2X strategy.

It also explicitly recognizes that AI software may be used to support the identification of chest X-ray abnormalities suspicious for TB.

This is particularly relevant because it places AI within an existing Vietnamese professional screening pathway rather than treating it as a separate diagnostic system.

However, this reference should not be interpreted as authorization of every commercial CAD product or as permission for AI to independently establish a TB diagnosis.

The same guidance describes the need to continue evaluation when molecular testing is negative but clinical or imaging findings remain concerning.

It also identifies circumstances in which molecular testing may be appropriate despite a normal chest X-ray, depending on symptoms, exposure history, and other risk factors.

These provisions reinforce the importance of clinical assessment beyond the imaging result.

The practical opportunity for healthcare organizations is to introduce CAD within established screening and diagnostic arrangements while preserving the relevant professional responsibilities.

Tuberculosis screening and infection prevention must remain connected

People attending a screening program may have infectious pulmonary TB before their diagnosis has been confirmed.

A TB screening service should therefore operate alongside appropriate infection prevention and control arrangements.

Depending on the setting, these may include symptom assessment at entry, respiratory hygiene, appropriate patient flow, environmental controls, and measures to protect healthcare workers.

The Ministry of Health issued Decision No. 579/QĐ-BYT on March 12, 2024, promulgating guidance on preventing TB transmission in healthcare facilities and communities [11].

The guidance addresses responsibilities and practical measures relevant to people with suspected or confirmed TB.

Introducing CAD does not replace these measures.

A high AI score does not independently establish that someone is infectious, while a low score should not override appropriate infection prevention precautions when symptoms or other clinical information raise concern.

Screening, diagnostic confirmation, infection prevention, and treatment are connected activities, but they should not be treated as interchangeable.

The screening pathway should allow clinically important information to trigger appropriate action without waiting for an algorithm to establish something it cannot independently confirm.

Non-TB abnormalities on chest X-rays also require attention

A chest X-ray obtained during TB screening may show findings unrelated to tuberculosis.

These could include abnormalities suggesting other pulmonary infections, chronic lung disease, cardiac conditions, or possible malignancy.

Their clinical significance depends on the imaging appearance, the patient's history, and the clinical context.

A CAD system developed for TB screening should not automatically be considered capable of identifying every clinically important chest X-ray abnormality.

This concern becomes particularly relevant when programs use CAD-only interpretation or sequential reading to reduce human workload.

An image classified as low-risk for TB may still contain an abnormality requiring separate clinical assessment.

Healthcare organizations should therefore define whether the examination serves only a specified TB screening purpose or is also expected to provide a broader diagnostic assessment.

Where broader chest imaging interpretation is required, an appropriate professional reading or another suitable pathway should remain available.

The organization should also determine how non-TB abnormalities are communicated and followed up.

Without these arrangements, a program could improve TB-focused screening efficiency while creating unintended delays in identifying other clinically significant conditions.

The purpose and limitations of CAD should be reflected accurately in patient communication, clinical documentation, and operational procedures.

Image quality and patient identification affect reliability

The performance of chest X-ray CAD depends on the images supplied to the software.

Positioning, patient rotation, inspiration, movement, exposure, artifacts, equipment characteristics, and image-processing methods can affect quality.

A system developed or evaluated using particular digital radiography images may not perform identically when applied to images obtained with different equipment or acquisition conditions.

Healthcare organizations should understand the image formats and technical conditions supported by the proposed software.

They should also determine how invalid, incomplete, or unsupported examinations are handled. A technically unsuitable image should not silently generate a reassuring screening result.

Depending on the circumstances, an appropriate response may involve professional review, repeat imaging when justified, or another diagnostic approach.

Correct patient identification is equally important.

A technically accurate AI result can contribute to harm if it is associated with the wrong person. This risk is particularly relevant during high-volume community screening, mobile radiography operations, or transfers between imaging, laboratory, and clinical information systems.

The patient identity, original image, examination date, AI result, and subsequent diagnostic findings should remain reliably linked.

Duplicate records, incorrect identifiers, and delayed synchronization may interfere with referrals or result communication.

A positive CAD result assigned to the wrong patient could cause unnecessary investigations while leaving the person who needs assessment without appropriate follow-up.

Data quality and traceability are therefore clinical safety concerns, not merely administrative requirements.

Local evaluation should assess the complete screening pathway

International studies provide useful evidence, but they do not guarantee identical performance in Vietnamese healthcare services.

Local TB prevalence, patient characteristics, previous TB history, HIV status, imaging equipment, and the mix of symptomatic and asymptomatic individuals can influence screening performance.

The CAD threshold and reading strategy also affect which people proceed to molecular testing.

Healthcare organizations should evaluate the technology in relation to the population and workflow in which it will be used.

Relevant measures include sensitivity, specificity, the proportion of examinations classified as screen-positive, diagnostic testing demand, and the number of potential cases missed.

Positive predictive value is also important, but it depends on disease prevalence and should not be assumed to remain constant across populations.

Evaluation should examine relevant patient groups where the available data are sufficient.

For example, residual pulmonary abnormalities in people with previous TB may influence the frequency of positive CAD results. Performance may also differ among older adults, people with mild symptoms, or those with other lung conditions.

Another important consideration is how TB cases are identified in the evaluation dataset.

If only people with positive screening results receive confirmatory testing, some cases among screen-negative participants may remain undetected.

This creates verification bias and can make apparent screening sensitivity more favorable than it would be under a more complete reference process.

The October 2026 Ho Chi Minh City study illustrates why this limitation matters [4].

Healthcare organizations should therefore understand not only how a model was evaluated but also who received confirmatory testing and how negative cases were classified.

An additional international reference became available on October 20, 2026, when WHO published Guidance on Evidence Generation on New Tuberculosis Diagnostics [8].

The document addresses how evidence should be generated for new TB diagnostic technologies, specimen types, and testing strategies. It emphasizes that diagnostic accuracy alone is not sufficient to establish their broader value.

Evidence should also consider feasibility, acceptability, equity, costs, and outcomes that matter to patients and health systems.

This guidance is not a product-specific endorsement of CAD or a mandatory Vietnamese evaluation protocol. Nevertheless, its emphasis on the intended place of a diagnostic technology within the clinical pathway provides a useful perspective for evaluating AI-assisted screening services.

A system that performs well in a carefully selected imaging dataset may still produce limited improvement if diagnostic testing is inaccessible, referrals are not completed, or the screening strategy does not suit the population.

Where appropriate and legally permissible, supervised local evaluation can help determine whether the technology supports safe and useful decisions before wider deployment.

The objective is not to reproduce a vendor's performance figure. It is to understand whether the system adds value within the organization's actual screening service.

Measuring whether AI improves tuberculosis screening

A healthcare organization should define how it will evaluate AI-assisted TB screening before introducing the software.

Sensitivity and specificity matter, but neither is sufficient to determine whether a program improves patient care.

The organization should understand how many eligible people are screened, how many receive positive results, and how many complete the recommended diagnostic evaluation.

The proportion of screen-positive individuals who complete molecular testing is particularly relevant.

A screening program may achieve high sensitivity yet provide limited benefit if many people identified for further testing do not complete the examination.

Time to diagnostic assessment is another important measure.

Delays between chest X-ray screening, specimen collection, laboratory processing, result communication, and clinical review can undermine the benefit of early identification.

For people with confirmed or clinically diagnosed TB, evaluation should consider whether they are appropriately linked to treatment services and whether the necessary next steps occur.

The number of bacteriologically confirmed cases detected is informative, but it must be interpreted alongside the population screened, diagnostic strategy, and completeness of follow-up.

The organization should also examine the consequences of false-positive findings.

Increasing molecular testing without a corresponding improvement in meaningful case detection may create costs and pressure on laboratory capacity.

Conversely, reducing testing demand may improve operational efficiency while increasing the risk of missed TB.

Both outcomes should be considered rather than assuming that a higher detection rate or lower workload independently demonstrates success.

The costs of screening also extend beyond the AI software fee. They may include X-ray equipment, image acquisition, software integration, connectivity, professional review, specimen collection, molecular testing, patient communication, referral arrangements, and follow-up.

Workload reductions during image interpretation may therefore be offset by additional demand elsewhere in the pathway.

Longer-term outcomes such as treatment initiation, diagnostic delays, and subsequent case detection are clinically important, but attributing improvements to AI requires appropriate evaluation methods.

Routine monitoring should not be presented as proof that AI independently reduces TB transmission, mortality, or other long-term adverse outcomes.

These measures are proposed organizational evaluation considerations, not a mandatory Vietnamese AI TB screening scorecard or an independently validated Digital Medicine Vietnam assessment instrument.

Monitoring must continue after deployment

An AI system may perform acceptably during initial evaluation but behave differently as clinical conditions change.

Software updates, changes in X-ray equipment, shifts in the target population, and modifications to screening procedures may influence outputs.

Changes in TB prevalence or diagnostic testing practices can also affect the proportion of positive screening results and confirmed cases.

Healthcare organizations should therefore examine both the algorithm and the clinical service in which it operates.

An unexpected shift in CAD scores may warrant review, particularly after a software or equipment update.

However, an increase in positive screening results does not necessarily mean that the algorithm has become more sensitive. It could reflect changes in the population, image quality, threshold, or underlying disease prevalence.

Similarly, a reduction in confirmed TB cases may reflect incomplete follow-up rather than improved screening specificity.

Monitoring should examine whether the selected interpretation strategy remains appropriate and whether patients continue to receive the necessary next steps.

Clinicians and screening staff should have a practical means of reporting unexpected AI results, technically unsuitable images, missed abnormalities, and failures in communication or referral.

The response may involve reviewing thresholds, changing the workflow, providing additional training, or restricting particular uses.

If the system becomes unavailable or unreliable, an appropriate fallback should remain accessible.

The purpose of monitoring is to maintain a dependable screening-to-care pathway rather than assume that the technology will remain suitable indefinitely after initial deployment.

The Vietnamese health-system context

AI-assisted pulmonary TB screening is particularly relevant to Viet Nam because of the country's continuing TB burden and efforts to strengthen early detection.

At the launch of Viet Nam's National TB Strategic Plan 2026–2030 on May 7, 2026, WHO noted that Viet Nam remained among the 30 countries with the highest TB burden worldwide [12].

WHO also recognized progress in case notification and treatment outcomes while highlighting continuing gaps in detection, service delivery, financing, and access to care.

The national strategy emphasizes active case finding, stronger TB services within primary healthcare, expanded public–private collaboration, prevention, and sustainable financing.

These priorities provide an important context for considering AI-assisted screening.

However, the value of CAD will differ across healthcare settings.

A specialized pulmonary hospital may have experienced radiologists and established molecular diagnostic services. A local healthcare facility may have access to chest radiography but limited rapid molecular testing capacity.

A mobile community screening program may reach people who rarely attend healthcare facilities but struggle to arrange subsequent investigations.

A private clinic may identify someone with possible TB but need to refer the person to an external TB service for diagnostic confirmation and treatment.

In each setting, the main problem may be different.

Where trained chest X-ray readers are difficult to access, CAD may support more consistent screening.

Where laboratory capacity is limited, the screening strategy must consider how many additional tests it will generate.

Where patients frequently fail to complete diagnostic assessment, better navigation, communication, and referral arrangements may provide greater immediate benefit than generating more positive screening results.

Decision No. 2627/QĐ-BYT offers a relevant professional foundation because it addresses active case finding in both community and healthcare settings [10].

Circular No. 34/2026/TT-BYT, issued and effective September 8, 2026, also provides guidance on selected activities under the National Target Programme on Healthcare, Population and Development for 2026–2035, with the first phase covering 2026–2030 [13].

The circular includes activities related to active TB screening and detection under applicable Ministry of Health professional guidance.

These policy developments support the strengthening of TB services, but they should not be described as independently requiring or authorizing the deployment of a particular AI product.

The practical priority is to align technology with the clinical pathway, available diagnostic resources, and needs of the population being served.

Vietnamese legal and professional responsibilities

Introducing AI-assisted chest X-ray screening requires consideration of the Vietnamese legal framework governing medical examination and treatment, disease prevention, artificial intelligence, medical devices, personal data protection, and radiation safety.

These responsibilities are related but distinct.

Evidence that a CAD product meets WHO performance standards or has received authorization in another country does not automatically establish that it may be used for every proposed clinical purpose in Viet Nam.

The applicable requirements depend on the specific software, its intended purpose, the way it is supplied and deployed, and the clinical decisions its outputs may influence.

Medical examination, treatment, and disease prevention

Viet Nam's Law on Medical Examination and Treatment No. 15/2023/QH15, effective January 1, 2024, as amended, provides an important legal foundation for healthcare services, professional practice, and patient rights [16].

A further development is the Law on Disease Prevention No. 114/2025/QH15, adopted on December 10, 2025, and effective July 1, 2026 [14].

This law establishes a broader legal framework for disease prevention, including the prevention and control of infectious diseases.

Decree No. 165/2026/NĐ-CP, effective July 1, 2026, provides implementing provisions for parts of the Law on Disease Prevention [15].

These documents are relevant to healthcare organizations conducting TB screening and other public health activities. However, they should be interpreted according to their actual provisions rather than treated as blanket authorization for any AI screening product.

Healthcare organizations considering AI-assisted TB screening should ensure that its use is consistent with their authorized services, the professional scope of participating healthcare practitioners, and applicable technical requirements.

The Ministry of Health's Decisions No. 162/QĐ-BYT and 2627/QĐ-BYT provide relevant professional guidance for TB diagnosis, active case finding, and clinical follow-up [9,10].

An AI-generated abnormality score does not independently establish a TB diagnosis or expand a practitioner's legally permitted professional scope.

Decisions about diagnostic testing, interpretation of results, diagnosis, and treatment remain subject to appropriate professional assessment.

WHO recommendations provide valuable evidence-based references, but they should not automatically be treated as legally binding clinical protocols that supersede applicable Vietnamese requirements.

Artificial intelligence and medical device regulation

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [17].

Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions concerning AI activities, including risk classification and management [18].

Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, promulgates a list of high-risk AI systems [19].

The legal classification of a TB CAD system depends on its actual functions, intended purpose, deployment arrangements, and applicable criteria.

A system designed to identify chest X-rays warranting further TB testing may have a different regulatory profile from software intended to make autonomous diagnostic or treatment decisions.

However, the absence of a particular CAD application from a named high-risk list does not independently establish that every such system is legally low-risk or exempt from other obligations.

Formal AI risk classification must also be distinguished from clinical safety risk.

A system can contribute to patient harm if its outputs lead to missed TB, unnecessary investigations, delayed diagnosis, or inappropriate reassurance.

CAD software intended for medical screening or diagnosis may also fall within Vietnamese medical device regulation, depending on its intended functions and applicable definitions.

Decree No. 98/2021/NĐ-CP on medical device management, as amended, including by Decrees No. 07/2023/NĐ-CP and No. 04/2025/NĐ-CP, forms part of the relevant framework [20].

Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [21].

Before routine deployment, healthcare organizations should determine the regulatory status of the particular CAD software and any relevant integrated imaging functions.

The regulatory status of a digital X-ray machine does not automatically establish that separately connected AI software may be used for every proposed clinical purpose.

Likewise, meeting WHO performance standards or holding foreign regulatory authorization does not independently demonstrate compliance with all applicable Vietnamese requirements.

The assessment should reflect the actual product, software version, intended medical purpose, and the role its outputs are expected to play in clinical decisions.

Personal data protection and confidentiality

Chest X-ray images and AI-generated results may contain or be linked to identifiable patient information.

These records can include patient identifiers, examination dates, symptoms, medical history, diagnostic results, and referral information.

Viet Nam's Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [22,23].

Healthcare organizations should assess the applicable legal grounds and purposes for processing patient information, consent requirements and relevant exceptions, access controls, information security, retention arrangements, and responsibilities of technology providers.

Particular attention is required when chest X-ray images are transferred to external CAD platforms or processed through cloud services.

The organization should understand where patient information is stored and processed, who can access it, and whether the data may be used beyond the intended screening purpose.

Where personal data are transferred outside Viet Nam, applicable cross-border transfer requirements and relevant procedural obligations or exceptions should be assessed separately.

Secondary use of chest X-ray images for algorithm development, model retraining, commercial product improvement, or research requires separate consideration of the relevant legal and institutional requirements.

Removing patient names or identification numbers does not necessarily establish that images and associated metadata are fully anonymous.

TB-related information also requires careful handling because inappropriate disclosure may create privacy concerns or social stigma.

Communication with patients, referral partners, and public health services should therefore reflect lawful processing purposes and appropriate confidentiality safeguards.

Where the technology is evaluated through clinical research, applicable research ethics requirements and participant protections should also be considered.

Radiation safety

Chest X-ray screening involves exposure to ionizing radiation, although the radiation dose from an individual examination is generally low.

The justification for performing the examination and the safe operation of radiographic equipment remain important.

Viet Nam's Law on Atomic Energy No. 94/2025/QH15 and Decree No. 332/2025/NĐ-CP took effect on January 1, 2026 [24,25].

Healthcare organizations should assess the radiation safety, licensing, and technical requirements applicable to their actual imaging activities.

Introducing CAD does not independently justify obtaining X-rays from people who do not meet the relevant screening or clinical criteria.

The software also does not replace professional and technical responsibilities associated with image acquisition, radiation protection, and equipment quality assurance.

Mobile screening activities require particular attention to the applicable arrangements for healthcare services, imaging equipment, professional responsibilities, and radiation safety.

These obligations should be assessed according to the actual service model rather than assumed to be satisfied because AI software is available.

A practical direction for healthcare organizations in Viet Nam

Responsible implementation should begin with a clear understanding of the problem the healthcare organization intends to solve.

A screening service may wish to reach more people in high-risk communities, reduce delays in chest X-ray interpretation, or improve the proportion of people who complete diagnostic testing.

These objectives are related but require different operational responses.

Before selecting CAD software, the organization should examine its existing screening pathway. It needs to understand which people are eligible, how X-rays are acquired and interpreted, where molecular testing is available, and what happens after positive, negative, or inconclusive results.

Clinical leaders, radiology professionals, TB clinicians, laboratory representatives, and relevant public health personnel should participate in defining the intended use.

The organization should review the evidence supporting the proposed product, its target population, imaging requirements, regulatory status, and limitations.

Local evaluation should consider the actual X-ray equipment, patient characteristics, threshold strategy, diagnostic capacity, and arrangements for professional assessment.

The service must also preserve a pathway for patients with significant symptoms or risk factors despite a low CAD score.

Where a positive screening result requires molecular testing, responsibilities should be clear for arranging specimen collection, communicating findings, and confirming that the necessary next steps occur.

Similar attention is needed for negative molecular results with continuing clinical suspicion, inconclusive findings, and patients who do not complete referrals.

AI may help determine which images warrant additional attention, but it cannot compensate for inadequate specimen collection, delayed laboratory results, fragmented communication, or incomplete follow-up.

Where improvements to existing processes would address the main problem more effectively, those improvements should not be postponed simply to introduce new technology.

The objective is to build a dependable screening-to-care pathway in which AI contributes useful information without replacing professional and public health responsibilities.

From AI screening to reliable tuberculosis care

AI-assisted chest X-ray screening offers opportunities to support earlier identification of people who may have pulmonary tuberculosis.

WHO guidance recognizes the potential role of CAD in eligible populations, while the October 2026 study in Ho Chi Minh City provides locally relevant evidence about the trade-offs between screening accuracy and operational efficiency.

Other recent evaluations demonstrate that performance varies across products and screening environments. WHO's October 2026 guidance on evidence generation further reinforces the importance of looking beyond test accuracy to feasibility, equity, costs, and outcomes that matter to patients.

These developments are encouraging, but they do not establish that every CAD system or implementation strategy will produce the same benefits.

A model that increases sensitivity may generate greater demand for molecular testing. A strategy that reduces reading workload may not improve TB detection. A program that identifies more people with suspicious X-rays may still provide limited benefit if those individuals do not complete diagnostic evaluation.

For healthcare organizations in Viet Nam, the value of AI-assisted screening depends on how reliably the technology is connected to clinical assessment, diagnostic confirmation, infection prevention, referral arrangements, and follow-up.

A positive AI result should lead to appropriate consideration of further investigation, not an automatic TB diagnosis.

A negative AI result should not prevent diagnostic assessment when symptoms, exposure history, or other clinical findings remain concerning.

Diagnostic testing and subsequent management must follow applicable professional guidance, with clear responsibilities for communicating results and ensuring appropriate care.

The success of AI-assisted TB screening should ultimately be judged by the reliability of the complete screening-to-care pathway, rather than the number of chest X-rays analyzed or abnormality scores generated.

AI-assisted pulmonary tuberculosis screening creates clinical and public health value when it helps identify people who need further assessment, supports appropriate diagnostic confirmation, and improves the likelihood that patients receive timely and continuous care, not when an algorithmic screening result is treated as a diagnosis or a substitute for clinical judgment.

Professional Education Notice

This article is provided by Digital Medicine Vietnam (DMVN) for professional education, evidence interpretation, and knowledge exchange. It does not constitute individualized tuberculosis screening, diagnosis, treatment, infection prevention, public health, medical device, legal, regulatory, or patient-safety advice.

The opening clinical scenarios are illustrative and do not represent documented patient cases.

The discussion focuses primarily on AI-based computer-aided detection systems used to support pulmonary TB screening through digital chest radiography, together with subsequent diagnostic evaluation, referral, and follow-up.

It does not establish the safety, effectiveness, regulatory status, or clinical suitability of any particular commercial, open-source, or internally developed AI system.

Different CAD products may have different intended purposes, eligible populations, imaging requirements, scoring methods, software versions, and clinical outputs. Evidence supporting one product or workflow should not automatically be generalized to another.

The October 2026 Ho Chi Minh City study provides evidence concerning simulated screening strategies and their potential effects on diagnostic accuracy and operational workload. It does not independently establish that implementing a particular CAD strategy reduces TB transmission, mortality, or other long-term adverse outcomes.

WHO guidance on CAD for TB screening applies to defined populations and intended uses. Meeting WHO performance standards should not be interpreted as automatic Vietnamese regulatory authorization or as evidence that CAD can independently confirm TB disease.

The referenced international recommendations and Vietnamese legal and professional documents must be interpreted according to their actual scope, applicable requirements, and effective dates.

The organizational approaches discussed, including threshold selection, diagnostic confirmation, patient follow-up, local evaluation, and continuing monitoring, are proposed considerations for responsible healthcare AI implementation and quality improvement.

They are not mandatory Vietnamese AI tuberculosis screening protocols, individual legal compliance determinations, or independently validated DMVN assessment instruments.

Clinical and public health decisions require appropriate professional assessment and must reflect the patient's circumstances and applicable guidance.

References

1. World Health Organization. WHO Consolidated Guidelines on Tuberculosis. Module 2: Screening – Systematic Screening for Tuberculosis Disease. Geneva: World Health Organization; 2021. ISBN: 978-92-4-002267-6.

2. World Health Organization. Use of Computer-Aided Detection Software for Tuberculosis Screening: WHO Policy Statement. Geneva: World Health Organization; 2025. ISBN: 978-92-4-011037-3.

3. Kohli M, Mace A, Miller C, et al. Evaluation of CAD Products in Community and Facility TB Screening: 2025 Policy Update. ERJ Open Research. Published online 2026. DOI: 10.1183/23120541.00215-2026.

4. Codlin AJ, Dao TP, Vo LNQ, et al. Accuracy and Operational Performance of Computer-Aided Detection Software Use Cases for TB Screening: A Retrospective Comparative Evaluation in Ho Chi Minh City, Viet Nam. BMJ Public Health. 2026;4(4). Published October 9, 2026. DOI: 10.1136/bmjph-2026-005566.

5. World Health Organization. WHO Consolidated Guidelines on Tuberculosis. Module 3: Diagnosis. Geneva: World Health Organization; 2025. ISBN: 978-92-4-010798-4.

6. World Health Organization. WHO Operational Handbook on Tuberculosis. Module 3: Diagnosis. Geneva: World Health Organization; 2025. ISBN: 978-92-4-011099-1.

7. World Health Organization. WHO Recommends Near Point-of-Care Tests, Tongue Swabs, and Sputum Pooling for TB Diagnosis. Departmental update. March 9, 2026.

8. World Health Organization. Guidance on Evidence Generation on New Tuberculosis Diagnostics. Geneva: World Health Organization; October 20, 2026. ISBN: 978-92-4-012379-3.

9. Ministry of Health of Viet Nam. Decision No. 162/QĐ-BYT Promulgating Professional Guidance on Tuberculosis Diagnosis, Treatment and Prevention. Issued January 19, 2024.

10. Ministry of Health of Viet Nam. Decision No. 2627/QĐ-BYT Promulgating Guidance on Active and Intensified Detection of Tuberculosis, Tuberculosis Infection and Selected Respiratory Diseases in Communities and Healthcare Facilities. Issued September 5, 2024.

11. Ministry of Health of Viet Nam. Decision No. 579/QĐ-BYT Promulgating Guidance on Tuberculosis Infection Prevention and Control in Healthcare Facilities and Communities. Issued March 12, 2024.

12. World Health Organization. Speech of Dr Angela Pratt at the 2025 TB Annual Review and Launch of Viet Nam's National TB Strategic Plan 2026–2030. Ha Noi, Viet Nam; May 7, 2026.

13. Ministry of Health of Viet Nam. Circular No. 34/2026/TT-BYT Providing Guidance on Selected Activities Under the National Target Programme on Healthcare, Population and Development for 2026–2035, Phase I: 2026–2030. Issued September 8, 2026; effective September 8, 2026.

14. National Assembly of Viet Nam. Law No. 114/2025/QH15 on Disease Prevention. Adopted December 10, 2025; effective July 1, 2026.

15. Government of Viet Nam. Decree No. 165/2026/NĐ-CP Detailing and Guiding the Implementation of Certain Provisions of the Law on Disease Prevention. Issued May 15, 2026; effective July 1, 2026.

16. National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment. Adopted January 9, 2023; effective January 1, 2024, as amended.

17. National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted December 10, 2025; effective March 1, 2026.

18. Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Artificial Intelligence. Issued April 30, 2026; effective May 1, 2026.

19. Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued June 30, 2026; effective August 15, 2026.

20. Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Medical Device Management, as amended, including by Decrees No. 07/2023/NĐ-CP and No. 04/2025/NĐ-CP.

21. Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Risk Determination and Management Measures for Products Classified as Medical Devices and Amendments to Circular No. 05/2022/TT-BYT. Issued June 30, 2026; effective July 1, 2026.

22. National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted June 26, 2025; effective January 1, 2026.

23. Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Personal Data Protection. Issued December 31, 2025; effective January 1, 2026.

24. National Assembly of Viet Nam. Law No. 94/2025/QH15 on Atomic Energy. Adopted June 27, 2025; effective January 1, 2026.

25. Government of Viet Nam. Decree No. 332/2025/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Atomic Energy Regarding Radiation Safety, Nuclear Safety and Security, Licensing, Inspection, Emergency Response and Related Matters. Issued December 18, 2025; effective January 1, 2026.