Governing AI-Drafted Diagnostic Reports in Viet Nam: Clinical Verification, Professional Sign-Off and Report Corrections
PRACTICAL FRAMEWORKS & TOOLS
Governing AI-Drafted Diagnostic Reports in Viet Nam: Clinical Verification, Professional Sign-Off and Report Corrections
PRACTICAL FRAMEWORKS & TOOLS
Last reviewed: 9 October 2026
A radiologist reviews an AI-generated draft of a chest X-ray report. The wording is clear and follows the hospital's usual reporting format, but one finding is described on the wrong side of the chest. Another draft states that an abnormality is unchanged from a previous examination, although the AI system did not have access to the earlier images or report.
Both errors can be corrected before the reports are approved. The more serious situation arises when an error goes unnoticed, the report is signed and the information reaches a treating clinician or patient. By that point, the report may already have influenced decisions about further investigation or treatment.
This is one of the practical challenges facing healthcare organizations introducing AI-assisted diagnostic reporting. A system may produce convincing medical language and reduce documentation time, but neither benefit guarantees that the final report accurately represents the examination. Even a small change in wording can alter the clinical meaning of a finding, particularly when it affects anatomical location, diagnostic certainty or recommended follow-up.
For hospitals, diagnostic centers and specialty services in Viet Nam, the important question is how to ensure that AI-generated content is clinically verified, approved by an appropriately authorized professional and corrected safely when necessary. The reporting process must protect something more important than the appearance of a well-written document: the reliability of diagnostic information that healthcare professionals and patients may use to make decisions.
An AI-generated draft is not a professionally approved diagnostic report
AI can support diagnostic reporting in different ways. Some systems analyze medical images and propose findings or impressions. Others convert structured observations entered by a clinician into narrative text, organize report sections or suggest standardized terminology.
These functions should not be treated as equivalent. A system that formats findings already verified by a radiologist performs a different task from one that interprets images and generates diagnostic conclusions. Both can introduce errors, but the sources of those errors and the verification required may differ. An AI application may also combine automated findings with clinical information retrieved from other systems, producing a report that appears complete even when the underlying information is incomplete, incorrectly associated with a patient or outside the system's intended scope.
Healthcare organizations should therefore understand what their reporting application actually does, which information it receives and which parts of the report are generated or modified by AI. This understanding is necessary to determine what the reviewing professional must verify before approving the final document.
It is equally important to distinguish an internal AI-generated draft from an authorized preliminary clinical report. An internal draft is working material awaiting professional verification and should not automatically be distributed as a confirmed diagnostic result. A preliminary report, where used, is a professional communication issued before final reporting and may already influence patient management. Its authorship, status and communication therefore require appropriate controls.
The distinction becomes particularly important when reporting applications are connected to hospital information systems, electronic medical records and patient portals. Information intended only for professional review should not inadvertently appear elsewhere as an authenticated clinical conclusion.
What recent evidence tells us about AI-assisted reporting
Research published during 2025 and 2026 suggests that AI-assisted diagnostic reporting can improve efficiency, but the findings also reinforce the importance of professional verification.
In a multicohort study published in Radiology in September 2025, Hwang and colleagues evaluated AI-generated chest radiograph reports across 1,539 examinations. Under a standard acceptability criterion allowing minor revisions, AI-generated reports performed similarly to radiologist-written reports, with acceptability rates of 88.4% and 89.2%, respectively. However, under a stricter criterion requiring no revision, AI-generated reports performed less well, at 66.8% compared with 75.7%. Most radiologists surveyed did not consider the AI reports sufficiently reliable to replace radiologist-written reports [1].
A prospective cohort study by Huang and colleagues, published in JAMA Network Open in June 2025, examined AI-generated draft reports in clinical practice. Across 11,980 model-assisted radiograph interpretations and 11,980 matched interpretations without model assistance, AI use was associated with a 15.5% improvement in documentation efficiency, corresponding to approximately 29 seconds saved per report. Peer review of 800 examinations found no statistically significant difference in the assessed clinical accuracy or textual quality of final reports [2].
These findings are encouraging, but they should be interpreted within the study's particular clinical setting and workflow. The absence of a statistically significant difference does not, by itself, establish that AI-assisted and unassisted reporting are equivalent in safety or quality across other settings.
Hong and colleagues reported further findings in the February 2026 issue of the Journal of the American College of Radiology. In a multireader study involving 756 chest radiographs and five radiologists, reading times decreased over successive batches, while acceptance of AI-generated drafts without modification increased. Agreement and report quality were more variable for abnormal examinations, highlighting the continuing importance of professional assessment when findings are complex or clinically significant [3].
More recent evidence from Singapore provides another practical perspective. In a study published in the Journal of Medical Internet Research in August 2026, Sridharan and colleagues evaluated an AI-triaged worklist combined with AI-assisted report generation across 1,054 chest radiographs. Median report generation time decreased from 2 minutes to 0.53 minutes under the AI-assisted workflow. Full agreement between AI-suggested findings and finalized radiologist reports occurred in 51.1% of cases, while radiologists removed at least one AI-suggested finding in 46.7% of cases [4].
These findings demonstrate substantial workflow benefits and active professional editing in the evaluated setting. However, the intervention combined worklist prioritization with reporting assistance, so its overall efficiency benefits cannot be attributed solely to AI drafting. Differences between AI-suggested and finalized findings should also not be interpreted directly as an AI diagnostic error rate. The study was not designed to establish downstream patient-safety outcomes.
Evidence from more complex reporting tasks raises additional concerns. In a 2026 study published in European Journal of Radiology Artificial Intelligence, Singh and colleagues compared impressions written by specialist radiologists with those generated by two large language models using existing oncologic CT and MRI reports. Across 200 imaging cases, the authors reported major errors in 6.3% and 5.1% of the AI-generated impressions, compared with 1.0% of radiologist-authored impressions. Problems included fabricated findings, inaccurate interpretation of interval changes and statements expressing greater diagnostic certainty than the available evidence supported [5].
These studies used different technologies, clinical tasks and research designs. Their results should not be directly compared or generalized to every diagnostic service. In particular, findings from chest radiography or oncologic imaging cannot automatically establish the performance of AI reporting applications in gastrointestinal endoscopy, pathology or other specialties.
Taken together, the evidence supports a balanced conclusion. AI can assist with preparing diagnostic reports, sometimes with meaningful efficiency gains. However, the final clinical document still requires verification appropriate to the examination, the system's intended purpose and the potential consequences of reporting errors.
A polished report can still contain clinically significant errors
One of the difficulties with AI-generated diagnostic reports is that the wording may appear more reliable than the underlying content. A report can use correct medical terminology, follow a familiar structure and reach a plausible conclusion while still containing a clinically significant mistake.
The problem may involve an abnormality that is not present, a finding that has been missed, an incorrect measurement or a statement that changes the clinical meaning of the examination. Laterality is a straightforward example. Describing a left-sided lesion as right-sided may affect subsequent assessment or intervention. A similar problem occurs when a negative finding becomes positive, or when a report states that an abnormality is absent despite evidence to the contrary.
Errors may also arise from information that the AI did not actually receive. A draft might refer to comparison with a previous examination, describe a procedure that was not performed or recommend follow-up without access to the relevant clinical context. The wording can appear entirely reasonable while communicating information that has not been verified.
The impression or conclusion deserves particular attention because it often communicates the most clinically important interpretation. An AI-generated impression may omit an abnormality documented in the findings section, overstate its significance or express certainty where the underlying evidence supports only suspicion. For example, an imaging feature may be suspicious for malignancy without establishing a definitive diagnosis. If AI converts that uncertainty into a confirmed diagnosis, the resulting report could influence treatment decisions beyond what the original evidence justifies.
Omissions can be equally important. Missing examination limitations, relevant negative findings or information about an incomplete procedure may leave the treating clinician with a misleading understanding of what was assessed.
These are not simply writing problems. They affect the integrity of diagnostic communication and illustrate why fluent language and a professional-looking report cannot substitute for clinical accuracy.
Clinical verification must return to the original diagnostic evidence
Reviewing an AI-generated diagnostic report should involve more than correcting grammar or improving the wording.
For diagnostic imaging, the authorized interpreting professional needs to assess the relevant images, clinical indication and available supporting information in accordance with their professional responsibilities. The AI draft may assist with documentation, but it should not become the sole source used to determine what the examination shows.
A practical verification process begins by confirming that the report belongs to the correct patient and examination. Patient identifiers, examination dates, anatomical regions, procedures performed and relevant clinical details should be consistent with the source records. Where contrast administration, specimen information or procedural details are relevant, the report should accurately reflect what occurred.
The reviewer then needs to determine whether the described findings are supported by the examination. This includes checking important abnormalities, anatomical locations, measurements, laterality, relevant negative findings and examination limitations. Where comparison with previous examinations is necessary, the reviewer should establish which earlier studies or reports were actually available. An AI-generated statement about interval change should not be accepted merely because it uses familiar reporting language.
The findings and impression should also communicate a consistent clinical interpretation. An important finding should not disappear from the conclusion without appropriate reason, and the conclusion should not introduce a diagnosis unsupported by the documented examination. The level of diagnostic certainty must reflect the available evidence.
The method of verification should reflect the AI system's function. When AI converts clinician-verified structured observations into narrative text, the review may focus on faithful transcription, completeness and preservation of clinical meaning. When AI proposes findings directly from diagnostic data, professional verification must also determine whether those findings are supported by the original evidence.
An additional concern is that an incorrect AI-generated statement may influence how the reviewer interprets the examination. The reporting workflow should therefore support genuine professional assessment rather than encourage the clinician simply to confirm the proposed wording.
The purpose is not to require an unnecessarily burdensome second reporting process. It is to ensure that the final document accurately reflects the clinical evidence and the professional's interpretation.
Professional sign-off must represent meaningful clinical approval
Electronic reporting systems can make it easy to accept, edit and sign AI-generated drafts. This may improve efficiency, but the approval process becomes less reliable if the workflow encourages rapid acceptance without sufficient verification.
A signature or electronic confirmation authenticates a document through the applicable process. However, the presence of a signature does not, by itself, demonstrate that every clinically important statement was adequately assessed.
Healthcare organizations should ensure that the person approving the report has the necessary professional authority and access to the information required for verification. That person should be able to correct findings, revise the impression, add relevant limitations, remove unsupported statements or reject the AI-generated draft entirely.
The design of the reporting interface also matters. If accepting a draft is much easier than correcting it, the system may unintentionally encourage acceptance. Similar concerns arise when professionals are expected to approve more reports without sufficient time or access to the original diagnostic information.
Some examinations require additional consideration, particularly when findings are complex, uncertain or likely to influence significant treatment decisions. A reporting workflow should allow appropriate consultation, additional review or further information where necessary.
The professional signing the final report remains responsible within the scope of their legally assigned professional duties. However, professional sign-off does not automatically eliminate the healthcare organization's or technology provider's responsibilities under applicable laws and contractual arrangements.
For clinical governance, a signed report should represent a genuine professional decision about its clinical content, not simply the completion of an electronic approval step.
What Viet Nam's legal framework means for AI-drafted diagnostic reports
Viet Nam's existing healthcare legislation provides the principal legal foundation for professional accountability and the management of diagnostic information.
The Law on Medical Examination and Treatment No. 15/2023/QH15, effective from 1 January 2024, is particularly relevant. Article 45 establishes professional duties, including compliance with professional and technical requirements and responsibility for the medical examination and treatment provided. Article 60 addresses the responsibilities of healthcare establishments, including responsibility for examination and treatment results produced by practitioners under their management. Article 69 governs medical records, including their creation, updating, storage and confidentiality, and recognizes the equivalent legal value of paper and electronic medical records. Article 71 addresses the prevention of medical incidents through identification, reporting, analysis and preventive measures [7].
Circular No. 13/2025/TT-BYT, effective from 21 July 2025, provides guidance on implementing electronic medical records. Article 3 specifies recognized methods of electronic signing or confirmation for relevant medical-record content, including lawful electronic signatures, biometric techniques and other electronic confirmation methods permitted under the Law on Electronic Transactions [8].
These provisions are relevant when diagnostic reports are incorporated into electronic medical records. However, they should not be interpreted as requiring every diagnostic report to use one particular form of digital signature or as establishing a universal clinical verification checklist specifically for AI-generated drafts. The appropriate method of authentication depends on the applicable documentation requirements, professional activity and information system.
Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 has been effective since 1 March 2026. Article 4 establishes human-centered principles, including maintaining human authority, control and responsibility in AI activities. Article 6 requires stricter risk management in essential sectors directly affecting life and health. For healthcare, it emphasizes patient safety, reliability under real-world conditions of use and protection of health data [9].
Decree No. 142/2026/NĐ-CP, effective from 1 May 2026, further details the AI regulatory framework, including risk classification, transparency, conformity assessment and obligations applicable to relevant participants [10].
Decision No. 33/2026/QĐ-TTg, effective from 15 August 2026, establishes Viet Nam's current list of high-risk AI systems. Its two healthcare-specific entries concern defined AI-assisted surgery and robotic-surgery systems under particular conditions. An application that drafts diagnostic reports should therefore not automatically be described as legally classified high-risk solely because it contributes to clinical documentation [11].
However, the absence of diagnostic report drafting from those healthcare-specific entries does not establish that the activity carries little clinical risk. Healthcare organizations still need to assess the system's actual regulatory classification, intended functions and potential consequences for patient care.
Other requirements may apply depending on the technology and how it is used. Where AI-enabled software is regulated as a medical device, the relevant medical-device framework, including Decree No. 98/2021/NĐ-CP, as amended, and Circular No. 24/2026/TT-BYT, should be considered as applicable [12,13]. Where patient information is processed through AI reporting applications, the organization must also assess its obligations under the Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP, according to the actual processing activities [14,15].
These legal instruments provide an important foundation for responsible implementation. They do not, however, establish one identical AI-draft verification, professional sign-off or report-amendment procedure for every diagnostic specialty.
Healthcare organizations should therefore develop appropriate reporting controls within the applicable legal and professional requirements, while recognizing that additional practical safeguards may be justified by the potential consequences of diagnostic errors.
Keep unverified AI drafts separate from reports released for clinical use
A diagnostic report may pass through several information systems before reaching the treating clinician or patient. In imaging services, an AI-generated draft may appear within a reporting application connected to a radiology information system and picture archiving and communication system. The final report may subsequently become available through the hospital information system, electronic medical record or another clinical interface.
The organization should understand how report status is represented throughout this process. An internal AI draft awaiting professional verification should be distinguishable from an approved final report. Where the service also uses authorized preliminary clinical reports, their status should be identifiable and appropriately managed.
Without these distinctions, an unverified AI draft may become visible to another clinician who assumes that the examination has already been interpreted and approved. The same concern applies to patient-facing systems, where a document may appear complete and official even though professional verification has not occurred.
Healthcare organizations should therefore determine when draft content becomes available, who can access it and what approval is needed before the information is released for clinical use.
This does not mean urgent clinical information must be withheld until every administrative step is completed. Where immediate communication is necessary, the service should use an appropriate authorized professional communication process. The objective is to support timely patient care without allowing an unverified AI draft to be mistaken for an authenticated diagnostic conclusion.
Corrections before and after report sign-off require different handling
Not every reporting error has the same consequences, and the appropriate response depends partly on when the error is identified.
Before professional sign-off, correcting an AI-generated draft is generally part of the normal review process. The reviewer can amend inaccurate wording, remove unsupported findings, correct laterality or revise the impression before approving the report.
Once a report has been finalized and released, the situation changes. The information may already have entered the patient's clinical record, been viewed by the treating team or influenced decisions about further investigation or treatment. A clinically important change to an issued report should therefore be managed through a controlled correction or amendment process consistent with applicable requirements and the organization's approved procedures.
Depending on the reporting system and clinical context, this may involve an addendum, an amended report or another controlled mechanism. The process should maintain appropriate traceability of the original report, the information changed, the person authorizing the amendment and the relevant timing.
Simply replacing a finalized report without sufficient traceability can make it difficult to determine which information was available when earlier clinical decisions were made.
The clinical significance of the correction should guide the response. Correcting a typographical error is different from changing the anatomical location of a lesion, revising a diagnosis or modifying a recommendation for further investigation. For example, changing a report from “no pneumothorax” to “pneumothorax present” may have immediate implications for patient management. Such a correction should not be handled as though it were an ordinary wording change.
It is also important to distinguish the correction of an error from an amendment made because new information has become available. Additional comparison studies, new clinical findings or subsequent professional interpretation may justify supplementing a report without necessarily indicating that the original report was inaccurate when issued.
Where an error may have originated in an AI-generated draft, the organization should consider how it entered the reporting process and why it was not identified before sign-off. The aim is to preserve the clinical integrity of the record and manage the consequences of information that may already have been used in patient care.
A corrected report must reach those who may have relied on the original
Correcting the document is only part of the response.
A treating clinician may already have read the original report, discussed the findings with the patient, ordered further investigations or made a management decision. If a clinically significant correction is issued afterward, the revised information needs to reach the appropriate recipients.
Updating a report in one electronic system does not necessarily mean that everyone who accessed the original has seen the correction. Healthcare organizations should therefore establish how significant amendments are communicated and, where appropriate, how receipt or acknowledgment can be confirmed.
The urgency should reflect the clinical consequences. A correction that could change immediate management may require direct and timely professional communication. A minor nonurgent amendment may be managed through established reporting channels, provided those channels are suitable for the circumstances.
The organization should also consider situations where the original report has been printed, downloaded, sent to another healthcare facility or accessed through a patient portal. An updated version in the main reporting system may not automatically replace copies available elsewhere.
International professional guidance provides useful reference points. The American College of Radiology's Practice Parameter for Communication of Diagnostic Imaging Findings distinguishes final reports from preliminary and other professional communications and addresses the importance of communicating significant differences that may affect care [16]. The Royal College of Radiologists' 2025 third-edition Standards for Interpretation and Reporting of Imaging Investigations also provides guidance on the quality and consistency of imaging interpretation and reporting [17].
These documents are professional guidance rather than binding Vietnamese legal requirements. Their practical relevance is that diagnostic reporting does not end when a document is signed. Clinically important corrections need to be managed as changes in information that may already have influenced patient care.
Report amendments can help identify weaknesses in the reporting process
Report amendments have long been part of diagnostic quality improvement, even without AI.
A retrospective study by Patra and colleagues, published in the Indian Journal of Radiology and Imaging in 2021, examined amendments to finalized CT, ultrasound and MRI reports. Among 97,003 approved reports, 1,076 contained addenda, representing approximately 1.1%. Some amendments corrected errors, while others added clinically relevant information without necessarily indicating an original reporting mistake [6].
The study predates widespread use of current generative AI reporting tools and should not be interpreted as an estimate of AI-related error rates. Nevertheless, it illustrates how reviewing amendments can help organizations identify weaknesses in observation, interpretation and clinical communication.
When AI-assisted drafting is introduced, healthcare organizations can apply similar quality-improvement methods while examining the potential contribution of AI.
The proportion of AI drafts accepted without modification may provide useful operational information, but it is not sufficient evidence of safety. A high acceptance rate could indicate that the system produces useful drafts, or it could reflect insufficient verification and excessive reliance on AI-generated content.
Likewise, frequent editing does not automatically indicate that AI is unsafe. Professionals may appropriately revise wording, clarify uncertainty, correct terminology or add information that was not available to the system.
More meaningful measures include clinically significant discrepancies identified before sign-off, inaccuracies that remain in finalized reports, amendments affecting clinical interpretation and failures to communicate important corrections.
These measures need to be interpreted carefully. Case complexity, professional experience, reporting workload and the limitations of the original examination may influence the results.
Where an AI-related error reaches the finalized report, the review should examine both the generated content and the professional verification process. The important question is not simply whether AI made a mistake, but how that mistake passed through the reporting workflow and whether the existing safeguards were adequate.
This approach supports learning from reporting errors without treating every amendment as evidence of AI failure or discouraging professionals from making appropriate corrections.
Different diagnostic specialties require different verification arrangements
The principles of professional verification, authentication and correction can apply across diagnostic services, but their practical implementation needs to reflect the nature of each specialty.
In radiology, the underlying evidence may consist primarily of images, relevant clinical history and comparison examinations. In gastrointestinal endoscopy, a report may need to describe findings observed during the procedure, anatomical locations, interventions performed, specimens collected, procedural limitations and complications. In pathology, report accuracy may depend on specimen identification, microscopic interpretation, additional investigations and specialist judgment.
An AI application that detects a possible lesion during endoscopy does not necessarily have the information required to generate a complete procedural report. Similarly, a system preparing narrative text from selected pathology images may lack important clinical or laboratory information needed for the final interpretation.
Healthcare organizations should therefore determine which elements of the report the AI can reasonably support, which information is available to the system and which content must be supplied or independently verified by the relevant professional.
The same caution applies when an organization uses several AI applications across different departments. A consistent reporting-governance principle can be maintained, but the actual verification arrangements may need to differ. Evidence supporting AI-generated chest radiograph reports should not be used as proof that AI-generated endoscopy, pathology or other specialty reports are equally reliable.
A practical approach for healthcare organizations in Viet Nam
Healthcare organizations do not necessarily need to create an entirely separate governance structure for every AI reporting application. A more practical approach is to incorporate specific verification, approval and amendment controls into existing diagnostic quality-management and patient-safety processes.
The starting point should be one clearly defined reporting use case. The organization needs to establish whether the AI produces diagnostic findings, converts clinician-verified observations into narrative text or combines these functions. It should also identify which examinations are within the application's intended use and what information is available when the draft is generated.
Before routine deployment, the reporting team should evaluate how the system handles clinically important content using representative local cases. Particular attention should be given to omitted findings, incorrect laterality, altered measurements, unsupported comparisons, inappropriate certainty and inconsistencies between findings and impressions. Published research and vendor demonstrations can inform this evaluation, but they should not replace assessment within the organization's own clinical environment.
The next consideration is professional verification and approval. The organization should determine who is authorized to approve each type of report, what diagnostic information must be available for review and whether the reviewer can readily edit or reject AI-generated content. The approval method should also meet the applicable requirements for professional authentication and clinical documentation.
Information technology and clinical teams should then examine how reports move through their systems. Unverified drafts, authorized preliminary communications and final reports should have appropriately distinguishable statuses. The systems should also support suitable correction procedures, version traceability and communication of clinically significant amendments.
After implementation, the organization should evaluate both efficiency and final report quality. A manageable review process may examine selected AI drafts alongside professionally approved reports, the nature of clinically significant edits, amendments made after release and any incidents involving inaccurate diagnostic communication.
The assessment should remain proportionate to the system's clinical influence. An application formatting findings already verified by a specialist may require different safeguards from one independently interpreting images and proposing diagnostic conclusions.
Healthcare organizations should also consider whether productivity expectations remain realistic once professional verification is included. A system that reduces drafting time but places unreasonable pressure on clinicians to approve more reports may undermine the clinical safeguards on which its safe use depends.
The objective is not to add unnecessary documentation or administrative burden. It is to ensure that the organization has a workable process for moving from AI-generated text to accurate, professionally approved diagnostic information, with an appropriate mechanism for correcting significant errors after release.
From faster report drafting to safer diagnostic communication
AI-generated diagnostic reports may help healthcare professionals reduce repetitive documentation and prepare structured clinical reports more efficiently. Recent research provides encouraging evidence for selected applications, particularly in diagnostic imaging, while also identifying limitations that healthcare organizations need to understand.
For healthcare organizations in Viet Nam, the central challenge is not simply whether AI can produce a convincing report. It is whether the final document accurately represents the examination, whether the appropriate professional has verified and approved its content, and whether clinically significant corrections are safely managed after release.
These responsibilities are not new to diagnostic services. However, AI introduces additional ways in which errors may enter the reporting process, and its apparent fluency can make those errors more difficult to recognize.
A reliable reporting process should make it possible to identify inaccurate AI-generated information before approval, maintain the integrity of finalized reports and ensure that important corrections reach those who may have relied on the original information.
The value of AI-assisted diagnostic reporting should therefore be judged not only by how quickly a draft is produced, but by whether the final report remains accurate, clinically meaningful and reliable throughout its use in patient care.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, medical-device, personal-data-protection, accreditation or quality-management advice.
The verification and amendment approach discussed here is a proposed organizational practice. It is not a prescribed Vietnamese Ministry of Health procedure, statutory AI reporting checklist, WHO standard or international certification requirement.
Applicable responsibilities depend on the AI system's intended purpose, actual functionality, regulatory classification, diagnostic specialty, professional scope of practice, documentation arrangements and relevant legal requirements.
The international studies cited in this article involve different technologies, clinical settings and evaluation methods. Their findings should not be generalized automatically to other diagnostic specialties, AI applications or healthcare organizations.
Recommendations concerning draft status, clinical verification, professional approval, controlled amendments, version traceability and communication of clinically significant corrections are intended to support responsible clinical practice alongside applicable legal and professional requirements.
References
Hwang EJ, Lee JH, Lim WH, et al. Clinical Validation of a Generative Artificial Intelligence Model for Chest Radiograph Reporting: A Multicohort Study. Radiology. 2025;316(3). DOI: 10.1148/radiol.250568.
Huang J, Wittbrodt MT, Teague CN, et al. Efficiency and Quality of Generative AI–Assisted Radiograph Reporting. JAMA Network Open. 2025;8(6). DOI: 10.1001/jamanetworkopen.2025.13921.
Hong EK, Suh CH, Nukala M, et al. Radiologist Interaction with Artificial Intelligence-Generated Preliminary Reports: A Longitudinal Multireader Study. Journal of the American College of Radiology. 2026;23(2):292–298. DOI: 10.1016/j.jacr.2025.09.015.
Sridharan S, Png N, Seah AXH, et al. Impact of AI-Triaged Worklists and AI-Assisted Report Generation on Radiology Turnaround Times: Prospective Real-World Study. Journal of Medical Internet Research. 2026;28. DOI: 10.2196/92181.
Singh R, Venugopal V, Daga R, et al. Clinical Readiness of LLM-Generated Oncologic CT/MRI Impressions: A Blinded Comparison with Subspecialty Oncoradiologists at a Tertiary Cancer Centre. European Journal of Radiology Artificial Intelligence. 2026;6:100088. DOI: 10.1016/j.ejrai.2026.100088.
Patra A, Premkumar M, Keshava SN, et al. Radiology Reporting Errors: Learning from Report Addenda. Indian Journal of Radiology and Imaging. 2021;31(2):333–344. DOI: 10.1055/s-0041-1734351.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment. Adopted 9 January 2023; effective 1 January 2024.
Ministry of Health of Viet Nam. Circular No. 13/2025/TT-BYT Providing Guidance on the Implementation of Electronic Medical Records. Issued 6 June 2025; effective 21 July 2025.
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Artificial Intelligence. Issued 30 April 2026; effective 1 May 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued 30 June 2026; effective 15 August 2026.
Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Medical Device Management, as amended.
Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Determining Risk Levels and Management Measures for Medical-Device Products and Amending Certain Provisions of Circular No. 05/2022/TT-BYT. Issued 30 June 2026; effective 1 July 2026.
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.
American College of Radiology. ACR Practice Parameter for Communication of Diagnostic Imaging Findings. American College of Radiology.
The Royal College of Radiologists. Standards for Interpretation and Reporting of Imaging Investigations. Third edition. London: The Royal College of Radiologists; October 2025.