Governing AI-Assisted Mammography Screening in Viet Nam: Radiologist Review, Recall Decisions and Interval Cancer Monitoring

Governing AI-Assisted Mammography Screening in Viet Nam: Radiologist Review, Recall Decisions and Interval Cancer Monitoring

IMPLEMENTATION INSIGHTS

Last reviewed: 28 October 2026

A woman attends a hospital for routine breast cancer screening. An artificial intelligence (AI) system analyzes her mammogram and highlights a small area of asymmetry. The radiologist reviews the images, compares them with previous examinations, and finds that the appearance has remained stable. There is no convincing evidence of a suspicious lesion.

Should the patient be recalled because AI has flagged the area? Would additional imaging provide useful information, or might it lead to unnecessary investigations and anxiety? What should happen when an algorithm produces a high suspicion score without identifying an abnormality that the radiologist can confidently interpret?

Another situation raises a different concern. Several months after receiving a negative screening mammogram, a woman develops a palpable breast lump. Subsequent diagnostic imaging and biopsy confirm breast cancer. Her earlier examination had been interpreted by a radiologist with AI assistance, and neither had identified a finding requiring further assessment.

The new diagnosis raises questions about the earlier screening examination. Was a subtle abnormality present but overlooked? Was the cancer not visible on mammography? Did breast density or image quality affect detection? How should the hospital review what happened without assuming that every cancer diagnosed after a negative mammogram represents a preventable error?

These situations illustrate why AI-assisted mammography screening requires more than an accurate image-analysis algorithm. AI can influence how radiologists interpret images, which patients are recalled, how additional investigations are arranged, and how screening performance is evaluated over time.

For healthcare organizations in Viet Nam, the challenge is not simply whether AI can identify more suspicious findings. It is whether the technology supports reliable interpretation, appropriate recall decisions, timely diagnosis, and meaningful evaluation of cancers that become apparent after screening.

Why screening performance cannot be judged by detection alone

Mammography is an established method of breast cancer screening, but its performance varies according to breast density, lesion characteristics, image quality, screening intervals, and the experience of interpreting radiologists.

Some cancers produce recognizable masses, suspicious calcifications, architectural distortion, or asymmetries that appear new or more conspicuous compared with previous examinations. Others may be difficult to distinguish from surrounding breast tissue, particularly when the breast is dense.

Screening also involves potential harms. Suspicious findings may lead to additional imaging or biopsy even when no cancer is ultimately identified. Some cancers detected through screening may never have caused symptoms or affected a person's health during their lifetime, creating the possibility of overdiagnosis and subsequent overtreatment.

The aim is therefore not to maximize cancer detection at any cost. A screening program must balance the potential benefits of earlier diagnosis against false-positive findings, unnecessary investigations, patient anxiety, and the resources required for follow-up.

AI may change that balance, but its effect depends on the technology and how it is incorporated into clinical practice.

An algorithm that highlights suspicious regions for radiologists serves a different purpose from one that prioritizes examinations for single or double reading. Software that estimates future breast cancer risk also differs from an application designed to identify abnormalities on the current mammogram.

These technologies may use similar image data while supporting different clinical decisions. Evidence demonstrating the performance of one application should not automatically be applied to another.

For hospitals considering AI-assisted mammography, the relevant question is whether the technology improves the overall screening pathway, rather than simply whether it achieves high accuracy in a development or validation dataset.

What the latest clinical evidence shows

The evidence for AI-assisted mammography screening has advanced considerably, particularly through large prospective studies examining cancer detection, radiologist workload, recall decisions, and interval cancers.

One of the most important studies is the Mammography Screening with Artificial Intelligence (MASAI) trial conducted in Sweden.

The randomized trial compared an AI-supported screening pathway with standard double reading by radiologists. In the AI-supported group, the system was used to direct examinations to single or double reading according to a predefined protocol and to highlight potentially suspicious findings.

An analysis published in The Lancet Digital Health in 2025 included 105,915 participants in the final evaluation [2].

AI-supported screening detected 338 cancers among 53,043 participants, compared with 262 cancers among 52,872 participants receiving standard screening. The corresponding cancer detection rates were 6.4 and 5.0 per 1,000 screened participants.

The AI-supported pathway also reduced the number of screen readings by 44.2%. Recall and false-positive rates were not significantly higher than those observed with standard double reading, although the number of recalls was slightly greater in the AI-supported group.

These findings suggested that a carefully designed AI-supported reading strategy could increase cancer detection while reducing radiologist reading workload.

However, detecting more cancers during screening does not necessarily establish that fewer clinically important cancers will be diagnosed later. Nor does it prove that breast cancer mortality will decline.

Further evidence became available in January 2026, when the MASAI investigators published their interval cancer analysis in The Lancet [1].

The study reported interval cancer rates of 1.55 per 1,000 participants in the AI-supported group and 1.76 per 1,000 in the standard double-reading group.

The AI-supported strategy met the prespecified criterion for noninferiority in interval cancer rates. This means the evaluated approach satisfied the trial's predefined requirement that it should not be unacceptably worse than standard double reading, using a 20% relative noninferiority margin.

It did not establish statistically significant superiority in reducing interval cancer rates.

Screening sensitivity was higher with AI support, at 80.5%, compared with 73.8% in the control group. Specificity was 98.5% in both groups.

The investigators also observed fewer interval cancers with certain unfavorable characteristics in the AI-supported group, including fewer invasive cancers and fewer cancers classified as T2 or higher. These findings are encouraging, but they should not be interpreted as independently established reductions in advanced disease or breast cancer mortality.

Additional evidence comes from the German PRAIM study, published in Nature Medicine in 2025 [3].

This prospective, multicenter observational implementation study included 463,094 women undergoing organized mammography screening, of whom 260,739 received AI-supported screening.

The breast cancer detection rate was 6.7 per 1,000 examinations in the AI-supported group, compared with 5.7 per 1,000 in the standard double-reading group. Recall rates were 37.4 and 38.3 per 1,000 examinations, respectively.

The positive predictive value of recall was also higher in the AI-supported group, at 17.9%, compared with 14.9% without AI support.

PRAIM provides valuable evidence from routine screening practice. Nevertheless, it was not a randomized trial. Radiologists could choose whether to use AI, and differences between the groups may have influenced the observed results.

A further contribution appeared in March 2026, when researchers in Spain published the AITIC study in Nature Medicine [4].

This prospective, paired noninferiority trial included 31,301 women undergoing digital mammography or digital breast tomosynthesis. Each examination was assessed using two parallel reading strategies: conventional double reading and an AI-based approach in which examinations classified as low-risk were considered normal without additional radiologist reading, while higher-risk examinations received double reading with AI support.

The evaluated AI strategy reduced the calculated radiologist reading workload by 63.6% and increased the cancer detection rate from 6.3 to 7.3 per 1,000 participants.

However, the recall rate increased by 14.8% relative to standard double reading, and the AI strategy did not meet the study's noninferiority criterion for recall.

This finding illustrates an important trade-off. Reducing reading workload and increasing cancer detection do not necessarily reduce the number of patients requiring further investigations.

The paired study design also requires careful interpretation. Both reading strategies were evaluated for each participant rather than assigning separate patient groups to experience only one approach. Consequently, the study did not establish the interval cancer rate that would result from independently implementing each pathway over time.

Its findings should not be interpreted as general authorization to exclude low-risk mammograms from professional review. Such a workflow requires evidence supporting the specific AI system and clinical setting, together with consideration of applicable regulatory and patient-safety requirements.

Taken together, MASAI, PRAIM, and AITIC show that AI can improve selected screening performance measures under defined conditions. They also demonstrate why reading configuration, recall consequences, and longer-term outcomes must be evaluated separately.

The evidence does not establish that all mammography AI products are equally effective, that professional interpretation can safely be removed from every screening pathway, or that improved detection necessarily produces better long-term patient outcomes.

Radiologist review remains central to clinical decisions

The clinical value of AI depends partly on how its findings interact with radiologist interpretation.

An AI system may generate an examination-level suspicion score, identify specific image regions, or provide information that influences the reading workflow. These outputs should not be treated as interchangeable.

A high examination-level score indicates that the software has identified image patterns associated with malignancy according to its underlying model. It does not necessarily mean that a radiologist can identify a corresponding suspicious lesion.

Similarly, an AI-generated mark may highlight a genuine abnormality, normal anatomy, overlapping tissue, or an imaging artifact.

Radiologists need to assess these findings using the available mammographic views, breast composition, previous examinations, and relevant clinical information.

Comparison with earlier mammograms is particularly important when evaluating asymmetries that appear new or have changed over time, or when determining whether an appearance has remained stable.

The point at which AI information becomes available can also influence interpretation.

Some systems show AI findings before the radiologist begins reading. Others provide the information after an initial assessment or use AI within a second-reading or arbitration process.

Presenting AI findings early may help direct attention toward subtle abnormalities. However, it may also create anchoring or automation bias, particularly when the software assigns a high suspicion score to a finding that is not clinically convincing.

Conversely, a low AI score may create false reassurance if a radiologist gives it greater weight than a suspicious finding visible on the images.

Hospitals should therefore examine how AI affects professional interpretation rather than assume that combining a radiologist and an algorithm will always improve performance.

The objective is to support a dependable reading process in which AI contributes relevant information while clinically important findings remain subject to appropriate professional assessment.

Different reading configurations require different evidence

AI can be integrated into mammography screening in several ways.

In conventional double reading, two radiologists interpret the examination, with disagreements resolved through an established consensus or arbitration process.

AI may be added to this arrangement as detection support. Alternatively, it may be used to identify examinations suitable for single reading while directing others to two radiologists.

Other approaches investigate AI as an independent reader or as a component of automated triage, potentially allowing some examinations to be classified as negative without individual radiologist review.

These configurations should not be considered equivalent because they change the role of the technology and the opportunities for professional interpretation.

The European Commission Initiative on Breast Cancer (ECIBC) issued two relevant conditional recommendations in July 2023 [5].

Based on the evidence available at that time, the guideline development group suggested against replacing conventional double reading with single reading supported by AI in organized population-based screening programs.

It separately suggested using AI-supported double reading rather than double reading without AI support.

Both recommendations were based on evidence assessed as having very low certainty.

The subsequent MASAI and AITIC findings provide important additional information because they prospectively evaluated different AI-supported screening arrangements.

However, neither study demonstrates that every reduced-reading or partially autonomous configuration is appropriate for routine implementation.

MASAI used a defined protocol in which AI directed examinations to single or double reading, with radiologists remaining involved. AITIC evaluated a different approach that allowed low-risk examinations to be classified as normal within the AI strategy without further human reading.

These differences matter. A hospital cannot assume that evidence supporting one configuration also validates another.

International recommendations can inform decisions, but they do not automatically establish a mandatory reading model for every healthcare organization in Viet Nam.

Screening examinations and diagnostic breast imaging must remain distinct

A safe AI-assisted mammography service should distinguish screening from diagnostic evaluation.

Screening mammography is generally performed in people without breast symptoms according to an established screening approach.

Diagnostic breast imaging addresses a specific concern, such as a palpable lump, suspicious nipple discharge, skin changes, or an abnormality identified during a previous examination.

This distinction matters because the expected clinical decisions are different.

A woman attending routine screening may proceed through the established mammographic assessment pathway. By contrast, someone presenting with a new palpable breast lump may require clinical examination, targeted imaging, and further investigation even if a recent screening mammogram was reported as negative.

A low AI suspicion score should not be used to dismiss a concerning clinical finding.

Similarly, evidence supporting a model developed for screening mammography should not automatically be extended to diagnostic mammography or other breast imaging applications.

Healthcare organizations should understand whether the proposed software supports full-field digital mammography, digital breast tomosynthesis, synthesized mammographic images, or another specified modality.

An algorithm validated on one image type may perform differently when applied to images obtained using other acquisition or processing methods.

The intended clinical purpose, eligible patient population, and supported imaging conditions should therefore be established before the software influences patient management.

Recall decisions should follow clinical assessment

One of the most consequential decisions in mammography screening is whether a patient should return for further assessment.

A recall may lead to additional mammographic views, comparison with earlier images, targeted ultrasound, or other clinically indicated investigations.

Recalling more patients may identify additional cancers, but it also exposes more people to anxiety, inconvenience, costs, and potentially unnecessary procedures.

Reducing recalls can lessen these burdens, but only if clinically important abnormalities are not overlooked.

AI can influence this balance by identifying findings that radiologists might otherwise miss or by providing additional information about the likelihood of malignancy.

However, the final clinical decision should follow an appropriately governed interpretation process. An AI score should not be automatically converted into a recall order or a recommendation for biopsy without the assessment required by the applicable clinical workflow.

The American College of Radiology's Breast Imaging Reporting and Data System (BI-RADS) provides standardized terminology, assessment categories, and management recommendations for breast imaging [6].

The BI-RADS v2025 Manual includes updated guidance supporting structured reporting, medical audit, and outcomes monitoring.

In screening mammography, BI-RADS 0 generally indicates that the assessment is incomplete and additional imaging evaluation or comparison with prior examinations is required.

A completed assessment of BI-RADS 1 or 2 generally supports routine screening, subject to the patient's clinical circumstances.

BI-RADS 3 describes a probably benign finding for which short-interval imaging follow-up is generally appropriate. It should ordinarily be assigned after an appropriate diagnostic assessment rather than used to avoid completing an initially incomplete screening examination.

BI-RADS 4 and 5 indicate suspicious or highly suspicious findings for which tissue diagnosis is generally recommended.

These categories represent professional imaging assessments. They are not direct translations of algorithmic suspicion scores.

A high AI score does not automatically establish BI-RADS 4 or 5. Likewise, a low AI score does not justify a negative assessment when the radiologist identifies a suspicious abnormality.

Healthcare organizations should ensure that AI-generated classifications are not entered into the medical record as final clinical assessments without the professional review required by the intended workflow.

Managing disagreements between AI and radiologists

Disagreements between AI and radiologists are likely to occur in routine screening.

A radiologist may identify suspicious architectural distortion even when the AI score is low. In another examination, AI may highlight an area that the radiologist considers benign or stable after comparison with previous images.

These situations require clinical interpretation rather than a rule that automatically favors either the algorithm or the radiologist.

When a radiologist identifies an abnormality warranting additional assessment, a low AI score should not prevent appropriate investigation.

When AI highlights a region that appears benign, the radiologist may review the relevant views, assess image quality, compare previous examinations, and decide whether additional imaging is justified.

A high AI score without a convincing imaging correlate presents a more difficult problem. It may warrant closer review, but it does not independently establish an indication for biopsy or repeated investigations.

The clinical team should consider whether the system has identified an interpretable finding, whether the examination meets the software's technical requirements, and whether the proposed response is supported by its intended use and available evidence.

Where multiple readers are involved, the hospital should establish how clinically significant disagreements are resolved. Depending on the screening model, this may involve a second radiologist, consensus review, or another appropriate arbitration process.

Measuring the frequency of agreement between AI and radiologists can provide useful information, but agreement alone is not a sufficient measure of safety.

Both the radiologist and AI may miss an important abnormality. Conversely, disagreement may reflect a radiologist correctly rejecting a false-positive AI finding.

A more meaningful evaluation examines how disagreements are resolved and whether the resulting decisions are consistent with subsequent diagnostic findings.

Where feasible and appropriate, preserving the original AI output, radiologist assessment, and final decision can support later quality review.

Such records can help distinguish algorithm-related concerns from interpretation errors, technical problems, or weaknesses in the clinical workflow.

A recall is only useful when the patient completes assessment

A screening program does not end when a radiologist recommends further examination.

The patient must receive the result, understand what is recommended, and have a practical opportunity to complete the next step.

A clinically justified recall may still fail to produce timely diagnosis if the patient cannot be contacted, does not receive an appointment, encounters financial barriers, or is uncertain about the significance of the finding.

These challenges become more complicated when screening and diagnostic services are provided by different healthcare organizations.

A woman may undergo screening at one hospital, receive further imaging at another, and ultimately have a biopsy or treatment elsewhere. Unless information is reliably communicated, the original screening service may never learn whether the recommended assessment was completed.

Hospitals should therefore evaluate more than the number of recall recommendations generated.

They need to understand whether patients complete additional imaging, whether biopsy is performed when indicated, and whether unresolved findings reach an appropriate diagnostic conclusion or documented disposition.

An incomplete assessment should not disappear from the follow-up process simply because no cancer diagnosis has yet been recorded.

The World Health Organization's Global Breast Cancer Initiative emphasizes the relationship between early detection, timely diagnosis, and comprehensive management [8,19].

The WHO training manual for breast cancer program managers, published in August 2026, further addresses patient navigation, multidisciplinary care, monitoring, and sustainable service delivery.

For AI-assisted screening, identifying a suspicious finding contributes to better care only when the patient can receive appropriate further assessment and management.

Why interval cancer monitoring matters

Some of the most important outcomes of mammography screening become apparent only after the original examination.

An interval breast cancer is generally a primary breast cancer diagnosed after a negative screening examination and before the next scheduled screening round, within the program's defined follow-up period.

The exact definition and observation window should be specified because screening intervals vary between programs.

Interval cancers are important because they may represent different situations. Some cancers were not visible on the original mammogram. Others were present but not recognized. Some may have developed or progressed after the examination.

These situations should not automatically be treated as equivalent to missed diagnoses.

Dense breast tissue can obscure a lesion that is not distinguishable from surrounding anatomy. A tumor may also develop rapidly after an apparently normal examination.

In other cases, an abnormality may become recognizable only when the earlier images are reviewed with knowledge of the subsequent cancer diagnosis.

This distinction is particularly important when evaluating AI.

A model may retrospectively assign a high suspicion score to an earlier mammogram after cancer has been diagnosed. However, this does not establish that using the model prospectively would have led to an appropriate recall or prevented an interval cancer.

The clinical significance of a retrospective finding depends on whether the information would reasonably have changed the original decision.

The 2026 MASAI results are valuable because interval cancer rates were evaluated prospectively within a randomized trial [1].

An additional review published in the Journal of Breast Imaging in August 2026 examined the potential role of AI in interval cancer detection and prediction, including the difficulties of comparing definitions, evaluating clinical workflows, and managing high AI scores without visible mammographic abnormalities [7].

The review reinforced an important principle: retrospective detection performance does not automatically demonstrate prospective clinical benefit.

For healthcare organizations, interval cancer monitoring should provide a longer-term view of screening performance rather than serve only as a retrospective test of whether AI could recognize abnormalities after the diagnosis was already known.

Reviewing interval cancers without hindsight bias

A structured interval cancer review should distinguish different reasons why cancer was not identified during the previous screening examination.

Some cancers may be classified as true interval cancers because the earlier mammogram showed no recognizable abnormality and the tumor became detectable later.

Other cancers may be mammographically occult, remaining difficult or impossible to demonstrate even when the subsequent diagnosis is known.

Some cases contain minimal signs that can be recognized retrospectively but would not necessarily have justified recall during the original examination.

Others may involve abnormalities that were sufficiently visible that further assessment would reasonably have been expected.

Technical limitations may also contribute, including incomplete tissue coverage, positioning problems, or inadequate image quality.

These distinctions matter because the appropriate improvements differ.

An abnormality that was clearly visible but not recognized may justify examining radiologist interpretation, training, and the interaction between AI and professional review.

A technically inadequate examination may indicate a need to improve mammography acquisition and quality assurance.

A cancer that was genuinely not visible on the earlier examination raises different questions about screening limitations and whether another imaging approach would have been appropriate for the patient's risk profile.

Reviewers must also consider hindsight bias.

Once the location and diagnosis of a cancer are known, subtle findings on earlier images can appear more obvious than they did during the original examination.

A structured review can reduce this bias by separating an initial assessment conducted without knowledge of the later diagnosis from a subsequent review informed by the diagnostic findings.

European breast screening quality assurance literature describes approaches involving independent review, consensus, and categories such as true interval cancers, minimal signs, false negatives, and occult cancers.

The classification method should be defined consistently before it is used to compare outcomes.

Interval cancer review can support professional learning and service improvement. It should not automatically be interpreted as evidence of individual negligence or legal liability.

Building reliable interval cancer monitoring

Meaningful interval cancer monitoring requires more than storing mammography images.

A screening service needs to identify which examinations were considered negative, when the next screening was expected, and whether breast cancer was diagnosed during the intervening period.

This can be challenging when patients receive subsequent care at different facilities.

A hospital that records only cancers diagnosed within its own organization may underestimate interval cancer rates if patients are diagnosed or treated elsewhere.

Similarly, comparing two screening programs without examining differences in follow-up completeness can produce misleading conclusions.

One program may appear to have fewer interval cancers simply because it has less complete information about what happened after screening.

Healthcare organizations should therefore distinguish an absence of documented interval cancers from reliable evidence that no interval cancers were identified during the defined follow-up period.

Where legally permissible and operationally feasible, information may be obtained through hospital records, pathology systems, authorized referral arrangements, or suitable cancer registry linkages.

The data required for evaluation should be determined in advance. Relevant information may include the original screening date, final radiologist assessment, AI system and software version, recall decision, subsequent diagnostic findings, pathology confirmation, and date of cancer diagnosis.

The monitoring period and denominator should also be consistent.

An interval cancer rate calculated per 1,000 negative screening examinations is not automatically interchangeable with a rate calculated per 1,000 screened participants.

Differences in screening intervals, eligibility criteria, observation periods, and completeness of cancer identification can affect comparisons.

It is also important to distinguish a cancer diagnosed after a genuinely negative screening examination from one diagnosed following a positive screening result whose diagnostic workup was delayed or incomplete.

Both may reveal problems in the screening service, but they should not automatically be combined into the same outcome category.

For Vietnamese healthcare organizations, establishing reliable longitudinal follow-up information may be more difficult than generating AI scores from mammography images. Nevertheless, without dependable outcome information, it becomes difficult to determine whether the technology is contributing to safer screening over time.

Breast density and patient differences deserve attention

Breast density affects mammography because dense fibroglandular tissue can obscure lesions and reduce the sensitivity of the examination.

Dense breast tissue is also associated with increased breast cancer risk, although density alone does not determine the appropriate screening or diagnostic pathway.

AI performance may differ across breast density categories, age groups, imaging modalities, and patient populations.

A model developed using mammograms from an organized European screening program may not perform identically when introduced into a Vietnamese hospital with different patient characteristics, equipment, or clinical workflows.

Local evaluation should therefore examine relevant differences rather than relying only on overall accuracy.

An algorithm may perform well across the full evaluation population while showing weaker performance in a subgroup that represents an important part of the hospital's patients.

Breast implants, postsurgical changes, variations in mammographic acquisition, and image-processing techniques may also influence whether a particular model can be used reliably.

It is equally important to distinguish AI detection scores from individualized breast cancer risk assessment.

A score indicating suspicion on the current mammogram is not necessarily a validated estimate of a woman's likelihood of developing breast cancer over the next five or ten years.

When supplemental ultrasound, magnetic resonance imaging, or another examination is clinically appropriate, the decision should reflect the patient's circumstances and applicable professional guidance rather than an unsupported interpretation of an AI score.

The Vietnamese screening context

The introduction of mammography AI in Viet Nam should be considered within existing breast cancer screening and diagnostic services.

Mammography may be provided through preventive health examinations, outpatient services, specialized breast clinics, or dedicated cancer screening programs.

These settings differ in their patient populations, equipment, staffing, access to previous images, and capacity to provide diagnostic follow-up.

Some hospitals have experienced breast radiologists, dedicated ultrasound services, biopsy capability, and coordinated pathology reporting. Others may offer mammography screening but refer patients to another facility for further assessment.

These differences can influence both the benefits and the risks of introducing AI.

Viet Nam's Ministry of Health issued Decision No. 3898/QĐ-BYT on December 18, 2025, promulgating professional guidance on breast cancer screening and early detection [9].

The guidance applies to healthcare facilities nationwide and addresses breast cancer risk assessment, screening methods, screening schedules, management of abnormal findings, diagnostic imaging, and subsequent evaluation.

For women at average risk aged 40 to 70, the document describes annual screening involving clinical breast examination and mammography, with supplemental ultrasound for women with dense breast tissue.

It also provides different considerations for women classified as having increased breast cancer risk.

These recommendations should be understood within the scope of the Vietnamese professional guidance and interpreted according to the patient's circumstances. They should not be presented as universally identical to screening recommendations in other countries.

On August 6, 2026, the Ministry of Health issued Decision No. 2507/QĐ-BYT, promulgating technical guidance on breast cancer screening and early detection through clinical breast examination [10].

This document focuses on the clinical breast examination process and related assessment, counseling, and referral activities. It is distinct from the broader breast cancer screening guidance issued under Decision No. 3898/QĐ-BYT.

Both documents are relevant to the organization of breast screening services in Viet Nam. However, neither should be interpreted as independently validating a particular AI mammography product or authorizing an algorithm to replace radiologist interpretation.

A healthcare organization considering AI must assess the specific technology, proposed clinical workflow, and applicable regulatory requirements separately.

Vietnamese legal and professional responsibilities

Introducing AI into mammography screening requires consideration of the legal framework governing medical examination and treatment, artificial intelligence, medical devices, and personal data protection.

These responsibilities are related, but compliance in one area does not automatically establish compliance in another.

Evidence that an AI system performed well in an international study, or that it has received authorization in another country, does not independently establish that it may be used for every proposed clinical purpose in Viet Nam.

The relevant assessment depends on the actual software, its intended medical purpose, deployment arrangements, data-processing activities, and the decisions its outputs may influence.

Medical examination, treatment, and professional accountability

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 organizations, professional practice, patient rights, and clinical services [11].

Healthcare organizations considering AI-assisted mammography should ensure that the proposed use is consistent with their authorized services, applicable technical requirements, and the professional scope of the practitioners involved.

An AI-generated image classification does not independently establish a breast cancer diagnosis.

Clinical interpretation, assessment of suspicious findings, decisions about further investigations, and subsequent management remain subject to applicable professional responsibilities.

The pathway may involve radiologists, treating physicians, breast specialists, pathologists, and other appropriately authorized healthcare professionals.

The introduction of AI does not independently expand professional scope or transfer clinical authority to a software provider.

International references such as BI-RADS and European breast screening guidance provide useful professional standards and evidence-based recommendations. However, they should not automatically be described as legally binding Vietnamese clinical protocols.

Relevant Ministry of Health requirements and institution-specific professional obligations should be assessed according to their applicability.

Artificial intelligence regulation

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [12].

Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions concerning AI activities, including risk classification and management [13].

Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, promulgates a list of high-risk AI systems [14].

The legal classification of a mammography AI application depends on its actual functions, intended purpose, deployment arrangements, and applicable legal criteria.

The healthcare-specific entries expressly identified in the high-risk list concern certain AI-assisted surgical and robotic systems. The list should not automatically be interpreted as assigning every mammography AI application to the same category.

A system that highlights suspicious findings for radiologist review may also present different operational and regulatory considerations from software intended to make autonomous screening decisions or determine which examinations receive reduced professional review.

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

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

An AI system may have significant patient-safety consequences if its outputs contribute to missed cancers, unnecessary recalls, delayed diagnosis, or incorrect communication of screening findings.

Healthcare organizations should therefore consider both regulatory compliance and the potential consequences of deploying the technology in clinical care.

Medical device regulation

AI software intended to analyze mammography images for medical purposes may fall within Vietnamese medical device regulation, depending on its functions, intended use, 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 regulatory framework [15].

Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [16].

Before routine clinical deployment, healthcare organizations should determine the applicable regulatory status of the particular AI software, mammography equipment, and any relevant integrated functions.

The regulatory status of a mammography machine does not automatically establish that separate AI software connected to it is permitted for every proposed clinical purpose.

Similarly, foreign regulatory authorization or published validation results do not independently demonstrate compliance with all Vietnamese requirements.

The assessment should reflect the actual product, software version, intended medical purpose, and decisions it is expected to support.

Personal data protection and secondary data use

Mammography images and associated screening records may contain or be linked to identifiable health information.

Patient identifiers, examination dates, imaging metadata, previous screening results, diagnostic findings, and pathology reports may all be relevant to AI-supported interpretation and outcome monitoring.

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 [17,18].

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

Particular attention is needed when mammography images are transferred to external AI platforms or processed through cloud infrastructure.

The hospital should understand where information is stored and processed, who can access it, and whether the technology provider may use the data for purposes beyond the agreed clinical service.

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

Secondary use of mammography images for algorithm development, commercial model training, or product improvement requires separate assessment of the relevant legal and institutional requirements.

Removing patient names or identification numbers does not necessarily establish that the images and associated records are fully anonymous. Metadata and linked clinical information may still permit re-identification.

Longitudinal interval cancer monitoring introduces additional considerations because it may involve linking screening examinations with later diagnoses and pathology results obtained from different healthcare organizations.

Such data linkage requires an appropriate legal and institutional basis.

Where AI is evaluated through clinical research, relevant research ethics requirements and participant protections should also be considered.

These matters should be addressed during implementation planning rather than after patient information has already been transferred or reused.

Measuring whether AI improves the screening service

A healthcare organization should establish how it intends to evaluate AI-assisted mammography before introducing the technology into routine practice.

The initial evaluation should determine whether the system operates reliably with the hospital's imaging equipment, supported formats, and intended patient population.

However, technical performance alone does not establish clinical effectiveness.

The hospital needs to understand whether AI changes radiologist interpretation, recall decisions, and diagnostic assessment in ways that are appropriate and beneficial.

Cancer detection rate is an important measure, but it should be considered alongside recall rate, false-positive rate, positive predictive value of recall, and the characteristics of cancers detected.

An increase in detected cancers accompanied by a substantial increase in unnecessary investigations presents a different benefit–harm balance from an increase achieved without additional false positives.

The type of cancer detected also matters.

Additional detection of ductal carcinoma in situ or other early lesions may contribute to earlier diagnosis, but a higher number of detected lesions does not automatically establish that overdiagnosis or overtreatment has been avoided.

Where reliable information is available, the organization should consider the distribution of invasive and in situ cancers, tumor size, stage, and relevant pathological characteristics.

Recall completion deserves equal attention.

A program may generate appropriate recommendations without improving care if patients do not receive or complete additional imaging or biopsy.

Hospitals should therefore examine the time from an abnormal screening assessment to further investigation and diagnostic conclusion.

Operational consequences also require evaluation.

AI may reduce the number of screening readings under certain protocols while increasing demand for diagnostic imaging, consultations, quality review, or technical support.

A reduction in workload at one stage does not necessarily mean that the total service burden has declined.

Longer-term monitoring should include interval cancers using a consistent definition, observation period, and reliable methods for identifying subsequent diagnoses.

Differences across relevant patient groups, breast density categories, imaging modalities, equipment types, and AI software versions should also be considered.

An algorithm that performs acceptably during initial evaluation may behave differently following changes in equipment, image processing, clinical workflow, or the patient population.

Healthcare professionals should have a practical means of reporting concerns, including unexpected AI–radiologist disagreements, changes in recall patterns, technically unreliable outputs, or difficulties completing recommended diagnostic assessment.

Corrective action may involve improvements to image acquisition, reader training, workflow design, software configuration, or data handling.

Where patient safety cannot be adequately maintained, restricting or suspending a particular application may be appropriate.

These activities should support an understanding of whether AI contributes to more reliable screening, rather than simply generating additional performance indicators.

The measures discussed are proposed organizational evaluation considerations. They are not a mandatory Vietnamese AI mammography scorecard or an independently validated Digital Medicine Vietnam assessment instrument.

A practical direction for healthcare organizations in Viet Nam

For hospitals and breast screening services in Viet Nam, responsible implementation should begin with an understanding of the existing clinical pathway.

Before selecting AI software, the organization should examine how mammograms are acquired, interpreted, and reported, how previous images are retrieved, and how patients with abnormal findings receive further assessment.

It should also identify where important delays or failures currently occur.

If patients experience difficulties completing diagnostic imaging or biopsy, improving these processes may provide greater immediate benefit than introducing an additional detection algorithm.

If the principal challenge is limited radiologist reading capacity, an appropriately evaluated AI-supported reading strategy may deserve consideration. However, the proposed reading configuration must be supported by relevant evidence and consistent with applicable professional and regulatory requirements.

Clinical leaders should participate in defining the intended use and the responsibilities of radiologists and other professionals.

Local evaluation should consider the hospital's actual equipment, patient characteristics, breast imaging practices, and capacity to provide diagnostic follow-up.

Where appropriate and legally permissible, supervised introduction may help determine whether the technology behaves reliably in the intended clinical environment.

The organization should also establish how disagreements are resolved, how recall decisions are documented, and how subsequent diagnostic findings contribute to quality review.

Longitudinal outcomes deserve particular attention because cancers diagnosed after an apparently negative screening examination may reveal limitations that are not visible in immediate detection statistics.

AI should be selected to address an identified clinical or operational problem, not simply because a vendor reports high accuracy or because historical mammography images are available.

A safe screening service requires reliable interpretation, appropriate recall decisions, effective patient communication, timely diagnostic assessment, and continuing evaluation.

AI may support these activities, but it cannot compensate for the absence of a functioning clinical pathway.

From AI detection to safer breast screening

AI-assisted mammography screening is becoming an increasingly evidence-supported application of clinical imaging technology.

The MASAI, PRAIM, and AITIC studies provide important evidence that AI can improve selected screening performance measures under defined conditions. At the same time, they illustrate that different reading strategies can have different consequences for cancer detection, recall rates, radiologist workload, and professional involvement.

The 2026 MASAI interval cancer findings add valuable information about the evaluated reading strategy beyond the original screening encounter. However, they do not establish that all mammography AI systems are equally effective or that improved detection automatically leads to better long-term outcomes.

For healthcare organizations, the clinical value of AI depends on the specific software, reading configuration, patient population, and ability to provide appropriate follow-up.

Radiologist interpretation remains important for understanding suspicious findings and determining how AI information should influence clinical assessment.

Recall decisions must balance the possibility of earlier cancer detection against the consequences of unnecessary investigations. Patients with abnormal findings need a reliable route to diagnostic evaluation, while those receiving negative results should understand that no screening examination can exclude every future cancer.

Interval cancer monitoring offers a longer-term perspective by examining what happens after the original screening encounter and identifying opportunities to improve the service.

For hospitals and clinics in Viet Nam, responsible implementation means connecting AI-supported image analysis with professional interpretation, clinically justified recall decisions, timely diagnostic assessment, and meaningful evaluation of subsequent outcomes.

The value of AI-assisted mammography screening should not be judged only by how many abnormalities an algorithm highlights or how many additional cancers are detected. It should be judged by whether the technology helps healthcare professionals make better decisions, supports timely diagnosis, and contributes to safer care without avoidable harm.

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 breast cancer screening, diagnostic imaging, oncology, medical, legal, regulatory, or patient-safety advice.

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

The discussion focuses on AI-assisted interpretation of screening mammography, radiologist review, recall decisions, diagnostic follow-up, and interval cancer monitoring. It does not establish the safety, effectiveness, regulatory status, or clinical suitability of any particular commercial AI product.

Different mammography AI applications may have different intended purposes, supported imaging modalities, target populations, reading configurations, and clinical outputs. Evidence supporting one application should not automatically be generalized to another.

The MASAI, PRAIM, and AITIC findings should be interpreted according to their respective study designs, populations, and clinical protocols. Improvements in cancer detection, sensitivity, recall performance, or radiologist workload do not independently establish reduced breast cancer mortality or justify unrestricted implementation.

The referenced international guidelines and professional standards provide evidence-based guidance. They should not automatically be interpreted as legally binding requirements that supersede applicable Vietnamese laws, regulations, or professional provisions.

The Ministry of Health documents discussed in this article provide relevant national professional guidance on breast cancer screening and early detection. They should not be interpreted as official Vietnamese validation or authorization of any particular AI mammography system.

The proposed organizational approaches, including AI–radiologist disagreement review, recall pathway monitoring, local evaluation, and interval cancer surveillance, are educational considerations for responsible healthcare AI implementation and quality improvement. They are not mandatory Vietnamese AI screening protocols, individual legal compliance determinations, or independently validated DMVN assessment instruments.

References

1. Gommers J, Hernström V, Josefsson V, et al. Interval Cancer, Sensitivity, and Specificity Comparing AI-Supported Mammography Screening With Standard Double Reading Without AI in the MASAI Study: A Randomised, Controlled, Non-Inferiority, Single-Blinded, Population-Based, Screening-Accuracy Trial. The Lancet. 2026;407(10527):505–514. DOI: 10.1016/S0140-6736(25)02464-X.

2. Hernström V, Josefsson V, Sartor H, et al. Screening Performance and Characteristics of Breast Cancer Detected in the Mammography Screening With Artificial Intelligence Trial (MASAI): A Randomised, Controlled, Parallel-Group, Non-Inferiority, Single-Blinded, Screening Accuracy Study. The Lancet Digital Health. 2025;7(3)–e183. DOI: 10.1016/S2589-7500(24)00267-X.

3. Eisemann N, Bunk S, Mukama T, et al. Nationwide Real-World Implementation of AI for Cancer Detection in Population-Based Mammography Screening. Nature Medicine. 2025;31:917–924. DOI: 10.1038/s41591-024-03408-6.

4. Elías-Cabot E, Romero-Martín S, Raya-Povedano JL, et al. AI-Based Triage and Decision Support in Mammography and Digital Tomosynthesis for Breast Cancer Screening: A Paired, Noninferiority Trial. Nature Medicine. 2026;32:1296–1305. DOI: 10.1038/s41591-026-04277-x.

5. European Commission Initiative on Breast Cancer. European Guidelines on Breast Cancer Screening and Diagnosis: Use of Artificial Intelligence in Mammography Screening. European Commission Joint Research Centre. Recommendations issued July 2023.

6. Newell MS, Destounis SV, Leung JWT, DeMartini WB, Lee CH, Eby PR. ACR BI-RADS v2025 Manual. Reston, VA: American College of Radiology; 2025.

7. Galvan JL, Nourmohammadi N, Milch HS, Yu TT. Artificial Intelligence and Interval Cancers in Screening Mammography. Journal of Breast Imaging. Published August 18, 2026. DOI: 10.1093/jbi/wbag042.

8. World Health Organization. Global Breast Cancer Initiative Training Manual: Package for Breast Cancer Programme Managers. Geneva: World Health Organization; 2026. ISBN: 978-92-4-012026-6.

9. Ministry of Health of Viet Nam. Decision No. 3898/QĐ-BYT Promulgating Professional Guidance on Breast Cancer Screening and Early Detection. Issued December 18, 2025.

10. Ministry of Health of Viet Nam. Decision No. 2507/QĐ-BYT Promulgating Technical Guidance on Breast Cancer Screening and Early Detection Through Clinical Breast Examination. Issued August 6, 2026.

11. 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.

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

13. 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.

14. 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.

15. 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.

16. 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.

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

18. 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.

19. World Health Organization. Global Breast Cancer Initiative Implementation Framework: Assessing, Strengthening and Scaling Up of Services for the Early Detection and Management of Breast Cancer. Geneva: World Health Organization; 2023. ISBN: 978-92-4-006598-7.