Governing AI-Assisted Cancer Histopathology in Viet Nam: Whole-Slide Image Quality, Pathologist Review and Diagnostic Safety

Governing AI-Assisted Cancer Histopathology in Viet Nam: Whole-Slide Image Quality, Pathologist Review and Diagnostic Safety

IMPLEMENTATION INSIGHTS

Last reviewed: 25 October 2026

A patient undergoes a biopsy because of a suspicious lesion. The tissue specimen is sent to the pathology laboratory, where it is processed, embedded in paraffin, sectioned and stained with hematoxylin and eosin (H&E).

The laboratory scans the glass slide to create a high-resolution digital image. An artificial intelligence system analyzes the whole-slide image and highlights a small region that may contain malignant cells. The pathologist reviews the highlighted area, examines the surrounding tissue and considers the available clinical information before reaching a diagnostic conclusion.

In another case, the AI system does not identify a suspicious region, but the pathologist subsequently finds a small focus of malignancy that the algorithm failed to highlight.

A third case presents a different problem. The AI identifies a suspicious pattern, but the relevant tissue area is partly out of focus in the digital image. The pathologist needs to review the original glass slide or obtain another suitable scan before the finding can be assessed reliably.

These situations illustrate why AI-assisted cancer histopathology requires more than software capable of recognizing abnormal tissue patterns.

The reliability of the diagnostic process depends on specimen preparation, slide quality, digitization, image management, algorithm performance and professional interpretation. Problems at any of these stages may affect the information available for diagnosis.

For pathology laboratories and hospitals in Viet Nam, introducing AI therefore involves more than purchasing a whole-slide scanner or connecting image-analysis software to an existing laboratory system.

The organization needs to ensure that the correct specimen is represented, the digital image is suitable for its intended purpose, the AI output is reviewed appropriately and clinically important discrepancies are addressed before the diagnostic result is finalized.

An accurate AI algorithm does not guarantee a safe cancer diagnosis if the underlying tissue specimen, digital image or clinical review process is unreliable.

Responsible implementation means treating AI as one component of a broader diagnostic service, with clearly assigned professional responsibilities and quality controls throughout the workflow.

From tissue specimen to digital diagnosis

Histopathology plays a central role in diagnosing many cancers. Examination of tissue architecture, cellular morphology and related microscopic features helps pathologists determine whether a lesion is benign or malignant. It may also provide information about tumor type and other characteristics relevant to subsequent clinical management.

The diagnostic process begins well before tissue is examined under a microscope. Specimen collection, identification, fixation, processing, embedding, sectioning and staining all influence the quality of the material available for assessment.

A technically satisfactory digital image cannot correct every problem introduced during these earlier stages. An inadequately sampled lesion, for example, may not contain the tissue needed to establish a diagnosis. A specimen identification error may associate an otherwise accurate microscopic finding with the wrong patient.

Whole-slide imaging (WSI) adds a digital component to this process. A scanner captures microscopic images across a glass slide and produces a digital representation of the tissue section. Pathologists can review these images using suitable software, displays and workstations.

Depending on the validated system and applicable requirements, digital pathology may support primary diagnosis, remote consultation, multidisciplinary collaboration, education and quality review.

AI introduces another function. Software may identify suspicious tissue regions, recognize selected morphological features or assist with a specific diagnostic task.

However, digital pathology and AI-assisted histopathology are not the same thing. A laboratory may use WSI for clinical diagnosis without employing AI. Conversely, an AI application may depend on a particular scanner, image format or digital pathology environment before it can operate reliably.

These components require related but separate evaluations.

The College of American Pathologists (CAP) updated its guidance on validating whole-slide imaging systems for diagnostic purposes in 2022 [3]. The guidance emphasizes that laboratories should evaluate WSI in conditions reflecting the intended clinical setting and diagnostic applications.

More recently, the Digital Pathology Association (DPA) published recommendations in June 2026 addressing the validation, implementation and clinical application of AI within pathology laboratories [1]. A central recommendation is that the slide-scanning process and AI algorithm should be validated separately.

For hospitals considering AI, this distinction provides a useful starting point. The relevant question is not simply whether an algorithm recognizes cancer in a research dataset. It is whether the laboratory has a reliable diagnostic process into which that particular application can be incorporated safely.

The intended clinical use should determine how AI is governed

AI applications in histopathology perform different functions and should not be evaluated as though they present identical clinical risks.

Some applications highlight tissue regions that may contain malignancy. Others classify selected lesions, assist with grading, quantify microscopic features or support the assessment of particular biomarkers. Additional systems are being developed for prognosis prediction, treatment-response estimation and laboratory workflow management.

These applications may rely on similar technologies, but their clinical purposes are different.

A model developed to detect suspected prostate cancer in needle biopsy specimens is not automatically suitable for identifying malignancy in breast, lung or colorectal tissue.

Similarly, an algorithm that highlights potentially suspicious regions is not necessarily validated to establish a definitive diagnosis, determine tumor grade or predict treatment response.

Before introducing AI, a laboratory should define the tissue type, clinical task, intended user, relevant patient population and expected output. It should also establish how the software will be used during examination.

For example, an application may highlight suspicious areas while the pathologist reviews a case. Another may act as an additional check after the initial professional assessment. A different system may prioritize certain slides for earlier review.

These functions can influence both the supporting evidence required and how the software should be integrated into routine practice.

This article focuses on AI-assisted detection and assessment of suspected malignancy in digitized H&E-stained histopathology sections, with review by a qualified pathologist.

It does not examine immunohistochemical biomarker scoring, molecular testing, prognosis prediction, autonomous cancer diagnosis or automated pathology report generation as separate applications. Frozen-section diagnosis, cytopathology and other specimen types also require separate assessment when they fall outside a system's validated intended use.

In September 2026, Rakha and colleagues published guidance in Virchows Archiv on the implementation, governance and continuous assurance of AI in histopathology [2].

Developed by members of the European Working Group for Breast Screening Pathology, the paper recommends that implementation arrangements reflect the application's intended clinical function and potential consequences for patients.

This is particularly relevant because software that highlights suspicious regions for professional review should not automatically be governed in the same way as a system whose output directly determines a consequential diagnostic classification.

The extent of verification, clinical review and monitoring should reflect what the application actually does and how its result may influence patient care.

Whole-slide image quality is a diagnostic safety issue

The reliability of AI-assisted histopathology begins with the quality of the tissue section and its digital representation.

A whole-slide image must capture the tissue required for the intended examination with sufficient detail. Several factors can interfere with this process.

The glass slide may contain tissue folds, tears, air bubbles, contamination, staining variation or regions that are difficult to bring into focus. The scanner may fail to capture the entire tissue section or may produce areas with inadequate sharpness.

Some malignant abnormalities occupy only a small part of the specimen. A relevant focus may therefore be missed if that area is blurred, omitted or represented inaccurately.

The problem affects both AI and professional review.

An algorithm may fail to recognize an abnormality because the necessary visual information is incomplete. A pathologist examining the same digital image may also be unable to assess the relevant morphology confidently.

Image quality cannot therefore be reduced to whether a file opens successfully or whether automated processing finishes without an error message.

The laboratory needs to determine whether the image adequately represents the tissue for the diagnostic task being performed. This includes considering tissue coverage, focus, staining quality, artifacts and compatibility between the scanner and viewing environment.

The location of an image defect matters.

A small area of blur outside the diagnostically important tissue may have limited consequences. A similar defect affecting a suspicious epithelial focus or a region where invasive carcinoma is suspected may be clinically significant.

A scan that appears satisfactory at low magnification may also contain local defects that become apparent only when detailed cellular features are examined.

The laboratory should therefore establish how diagnostically important image-quality problems are identified and what happens when a scan is inadequate.

Depending on the circumstances, the appropriate response may involve rescanning, examining the original glass slide or preparing another tissue section.

The CAP guidance on WSI validation specifically recognizes the importance of confirming that tissue present on a glass slide is represented in the digital image [3].

Quality assurance should not rely solely on automated image-quality checks. A system may pass a technical assessment while still containing an omission or artifact that matters for diagnosis.

A whole-slide image is suitable for clinical use only when it adequately supports the examination for which it is intended.

The correct digital slide must be linked to the correct patient and specimen

One of the most serious patient-safety risks in histopathology is incorrect identification or attribution of tissue.

A specimen may pass through several stages before a digital image is created, including accessioning, gross examination, tissue processing, block preparation, sectioning and staining.

Digital pathology adds scanning, file management, image transfer and potential AI analysis. At each stage, the material and associated information must remain correctly linked.

An AI algorithm may accurately identify malignant tissue while the digital image is associated with the wrong patient or specimen. In that situation, algorithmic accuracy provides no protection against a potentially serious diagnostic error.

The laboratory should maintain reliable identification and traceability across the diagnostic process.

Depending on its workflow, this may involve the accession number, specimen designation, tissue block identifier, slide identifier, staining method and examination date.

A whole-slide image should remain traceable to its original glass slide and source material.

When several tissue blocks or additional levels are examined, the digital record should make clear which material each image represents.

The same principle applies when a slide is rescanned or a replacement image is created. An AI result produced from an earlier image should not automatically be associated with a later scan or different tissue section without appropriate verification.

This matters because adjacent sections from the same tissue block may not contain identical microscopic features. A small malignant focus may be present in one section but absent from another.

The laboratory should therefore be able to establish which image was analyzed, which version of the software produced the result and whether the relevant material was subsequently replaced or reexamined.

These considerations are particularly important in cancer care because pathological findings may influence surgery, systemic treatment, radiotherapy and other consequential clinical decisions.

An AI-highlighted region is not a cancer diagnosis

AI applications may highlight regions that resemble malignant tissue or other abnormal microscopic patterns.

These outputs can help pathologists direct attention toward areas deserving closer examination. However, a highlighted region is not necessarily malignant.

Inflammation, tissue repair, reactive cellular changes, staining artifacts and benign lesions can sometimes produce appearances resembling features associated with cancer.

A model may also respond to image characteristics unrelated to the clinical question. For example, staining differences or scanner-specific patterns may influence the result even though they do not represent meaningful biological abnormalities.

The pathologist must interpret suspicious findings within the broader morphological and clinical context.

This involves examining tissue architecture, cellular features and the relationship between the suspicious area and surrounding tissue.

Where appropriate, additional tissue levels, special stains, immunohistochemistry or other investigations may be necessary. An AI output does not independently establish that these investigations should be performed or omitted.

The pathologist may conclude that a highlighted area is malignant, represents a benign process or requires further assessment.

The purpose of AI assistance is to provide information that contributes to professional evaluation. It is not to convert a computational classification into an unquestioned diagnostic conclusion.

A negative AI result should not create false reassurance

False-negative findings deserve particular attention because a missed malignant lesion may have important consequences for patient care.

An AI system may fail to identify a small focus of cancer, an uncommon tumor pattern or a lesion whose microscopic appearance differs from the examples represented in its development data.

Tissue artifacts, unusual morphology and inadequate image quality can also affect detection.

The risk becomes more significant when pathologists begin to treat the absence of an AI highlight as evidence that the slide contains no clinically important abnormality.

An application validated for detecting a particular cancer pattern should not automatically be used to exclude every type of malignancy or other important pathology.

Likewise, a system designed to identify suspicious regions may not be validated to issue a definitive negative diagnosis.

The laboratory should establish how AI-negative results are used within the intended clinical workflow. Where AI is introduced as an additional detection aid, it should not reduce professional examination below what is appropriate for the diagnostic task.

A clinically important finding identified by the pathologist must remain relevant even when the algorithm has not recognized it.

The organization should also understand whether the product has limitations concerning minimum tissue area, tumor content, supported tissue types or image quality.

A result generated outside these conditions may be unreliable even when the system does not display an explicit warning.

The DPA's 2026 recommendations emphasize defining minimum usable sample requirements and relevant artifact limitations for pathology AI applications [1].

For clinical implementation, the important question is not simply whether the algorithm produces a negative output. It is whether that output is valid for the material examined and whether the complete diagnostic process remains sufficiently reliable.

Clinically important disagreements need proportionate review

Differences between AI and pathologist findings are not always evidence that one is correct and the other is wrong.

A discrepancy may result from different detection thresholds, definitions of the target lesion, image-quality problems or limitations in the reference standard.

Some differences have little clinical significance. Others may influence whether a lesion is considered malignant, whether further examination is necessary or whether the case requires specialist consultation.

The laboratory should establish a practical way to manage clinically important discrepancies.

When AI identifies a suspicious region that the pathologist considers benign, additional examination of that area may be appropriate.

When the pathologist identifies a malignant focus missed by AI, the case may provide useful information about the system's limitations.

However, not every disagreement should automatically trigger extensive documentation, additional testing or referral to another institution.

The response should reflect the diagnostic uncertainty and its potential consequences for patient care.

Depending on the case, the pathologist may consider additional tissue sections, glass-slide review or a second professional opinion.

Recurring discrepancies may also reveal problems associated with particular tissue types, artifacts, staining methods or scanners. Such findings can support quality improvement and local performance monitoring.

The objective is to strengthen diagnostic reliability, not to require agreement between AI and pathologists in every case.

What clinical implementation studies tell us

Research has demonstrated that AI can assist pathologists with selected cancer detection tasks. However, diagnostic performance varies across tissue types, products, study designs and clinical settings.

Evidence from selected research datasets should not automatically be treated as evidence of improved routine clinical practice.

Studies in which AI is incorporated into actual diagnostic workflows are therefore particularly relevant.

In March 2025, Flach and colleagues published the prospective CONFIDENT P study in JCO Clinical Cancer Informatics [4].

The investigators evaluated AI assistance for detecting prostate cancer in H&E-stained needle biopsy specimens.

The study compared an AI-assisted workflow with a control workflow involving routine histopathological assessment and immunohistochemistry.

The analysis included 237 slides from 82 patients. Of these, 143 slides from 64 patients contained prostate cancer according to the study's reference diagnoses.

AI assistance was associated with reduced immunohistochemistry use. At the patient level, the reported relative risk of immunohistochemistry use per detected prostate cancer case was 0.55, with a 95% confidence interval of 0.39 to 0.72.

The investigators also reported greater diagnostic confidence among pathologists using AI assistance.

However, the median assessment time per H&E slide did not differ significantly between the study groups.

These findings show why implementation outcomes need to be examined individually. An AI application may reduce the use of additional tests without necessarily shortening every stage of professional diagnostic review.

A more recent study by Udoh and colleagues, published in the Journal of Pathology Informatics in 2026, evaluated the implementation of an AI-assisted prostate biopsy decision-support system in a fully digital academic pathology practice in the United States [5].

Following institutional validation, the system was incorporated into routine diagnostic work.

The researchers compared 139 prostate biopsy cases across periods before and after implementation, comprising 80 cases before AI introduction and 59 afterward. A transition month was excluded from the analysis.

The study reported that weighted median diagnostic turnaround time decreased from 36.3 hours to 25.3 hours, a reduction of approximately 30.3%.

Immunohistochemistry utilization also decreased, from an average of 2.37 stains per case to 1.47, representing a reduction of approximately 38%.

These results suggest that a narrowly defined AI application may improve aspects of workflow efficiency within an established digital pathology service.

However, the definition of turnaround time is important.

In this study, diagnostic turnaround time was measured from completion of whole-slide image scanning to final pathologist sign-out.

The reported improvement should therefore not be interpreted as a 30% reduction in the entire process from specimen collection or laboratory receipt to delivery of the final diagnosis.

The study also involved one institution and compared relatively short periods before and after implementation. Changes in staff experience, case characteristics or other aspects of laboratory practice may have influenced the findings.

Neither study established that AI independently improved cancer treatment outcomes or patient survival.

Both studies concerned defined applications for prostate biopsy assessment. Their results should not automatically be generalized to histopathology involving other organs or cancer types.

For hospitals in Viet Nam, the practical lesson is that the value of AI should be demonstrated within a particular diagnostic task and workflow rather than assumed from algorithm performance alone.

Algorithm validation and local verification are different responsibilities

An important distinction in international guidance concerns the difference between validating an AI algorithm and verifying its performance within the laboratory where it will be used.

A developer may evaluate an application using selected datasets and clinical studies. These evaluations provide evidence that the system performs its intended function under defined conditions.

Regulatory authorization, where applicable, may provide additional information about the product's status in the relevant jurisdiction.

However, these steps do not automatically establish that the software will perform reliably in every pathology laboratory.

Local conditions may differ from those represented during development. Hospitals may use different scanners, staining procedures, tissue preparation methods, image management systems and diagnostic workflows.

The September 2026 guidance by Rakha and colleagues distinguishes three complementary stages: algorithm validation, local verification and continuous assurance [2].

Algorithm validation establishes the evidence supporting the application's intended performance.

Local verification examines whether the previously validated system performs acceptably under the laboratory's actual conditions of use.

Continuous assurance involves monitoring, incident management, change control and other activities necessary to maintain acceptable performance after deployment.

These stages serve different purposes and should not be treated as interchangeable.

For a Vietnamese hospital, local verification should consider the specimens, scanners, staining methods and clinical tasks expected in routine practice.

It should also examine whether images are transferred and processed reliably, whether results are presented at an appropriate point in the diagnostic workflow and whether pathologists can review them effectively.

Acceptance criteria should reflect the intended use and potential patient consequences. The evaluation should also specify what will happen when an image or result does not meet those criteria.

A system that highlights regions for optional review may require a different extent of evaluation from one whose output directly influences a consequential diagnostic classification.

However, even an assistive application can create risks if staff begin to rely on it beyond its validated purpose.

Local verification should therefore consider not only the accuracy of individual outputs but also how the software changes professional practice.

Validating digital pathology does not automatically validate AI

Another important distinction concerns the whole-slide imaging system itself.

A laboratory may have validated digital pathology for primary diagnosis. This provides evidence that the digital imaging process is suitable for defined professional diagnostic activities.

However, successful interpretation of whole-slide images by pathologists does not automatically demonstrate that an AI algorithm will analyze those images reliably.

The algorithm may respond differently to changes in resolution, staining characteristics, scanner optics, compression or image format.

The CAP guidance addresses validation of WSI systems for diagnostic use [3]. The DPA's 2026 recommendations go further by emphasizing that the scanning process and AI application require separate validation considerations [1].

They also stress the need to establish appropriate evidence of algorithm performance with the scanner models used in clinical practice.

Where an organization wishes to use an algorithm with an additional scanner model, suitable compatibility evidence is necessary.

In some circumstances, well-designed interscanner concordance studies may support this assessment without requiring complete repetition of the original clinical validation.

However, such evidence must be relevant to the intended use and the differences between the devices involved.

This distinction is particularly useful for laboratories transitioning from conventional microscopy to digital pathology.

The first decision concerns whether the digital pathology system is suitable for the intended diagnostic activities.

The second concerns whether a particular AI application can be added safely to that environment.

Completing the first evaluation does not automatically resolve the second.

Clinical utility requires more than agreement with a reference diagnosis

AI studies frequently report diagnostic accuracy, agreement with pathologists or the ability to identify abnormalities in digital images.

These measures are useful, but their interpretation depends on the study design.

Research datasets may exclude difficult tissue sections, poor-quality scans, rare tumor variants or cases with unusual morphology.

If those exclusions are not considered, reported performance may appear more favorable than the results achievable during routine clinical care.

The reference standard also matters.

Some studies compare AI findings with the opinion of one pathologist. Others use multiple reviewers, consensus interpretation or additional diagnostic investigations.

These approaches do not necessarily provide equivalent levels of certainty.

Agreement with a reference diagnosis does not automatically establish that an AI application improves the accuracy of the pathologist who uses it.

For an assistive system, the clinically relevant comparison may involve how pathologists perform with and without AI support.

The evaluation should consider whether AI helps identify clinically important findings that would otherwise be missed, whether it introduces unnecessary investigations and whether it changes diagnostic decisions appropriately.

Another concern is automation bias.

A pathologist may become more likely to accept an AI-highlighted region as suspicious or to overlook an area that the algorithm has not highlighted.

These effects depend on the application and how it is integrated into the workflow.

A system that generates more highlights or processes images faster is not necessarily safer.

Clinical utility requires evidence that the additional information makes a meaningful contribution to diagnostic practice.

The DPA's 2026 recommendations emphasize that clinical applications should demonstrate appropriate analytical and clinical validation, with evidence of utility relevant to the intended patient-care benefit [1].

For pathology laboratories, this means looking beyond the algorithm's technical performance to examine how the complete diagnostic process changes.

International laboratory standards provide a quality foundation

AI-assisted histopathology should be integrated into the laboratory's quality management arrangements rather than managed solely as a technology project.

ISO 15189:2022, Medical laboratories — Requirements for quality and competence, provides an internationally recognized framework for medical laboratory quality and professional competence [6].

ISO/TS 23824:2024 provides guidance on applying ISO 15189 to anatomic pathology laboratories [7].

These documents are relevant because digital pathology and AI depend on reliable pre-examination, examination and post-examination processes.

Specimen identification and preparation affect the material available for diagnosis. Scanning and image quality influence the digital representation of that material. Professional interpretation and communication of findings determine how the diagnostic information is used.

A laboratory's quality management system can help coordinate these activities.

However, ISO 15189 accreditation should not automatically be interpreted as approval of every AI algorithm used within an accredited laboratory.

The relevant accreditation scope, validated activities and applicable requirements remain important.

Another development is ISO/DIS 24051-2:2026, Medical laboratories — Part 2: Digital pathology and artificial intelligence (AI)-based image analysis [8].

As of October 25, 2026, this document remains a Draft International Standard under development. The ISO enquiry ballot began on October 23, 2026.

Its proposed scope addresses digitization of stained tissue sections, processing of whole-slide images and AI-based image analysis supporting anatomical pathology examinations, particularly involving paraffin-embedded tissue.

The draft reflects growing international attention to the technical and quality requirements associated with digital pathology and AI.

However, it should not be described as a published International Standard, an existing certification scheme or a mandatory legal requirement for pathology laboratories in Viet Nam.

Healthcare organizations should distinguish established standards from documents that remain under development.

Pathologist review remains central to diagnostic safety

The pathologist's role is not limited to confirming whether an AI-highlighted region contains malignant cells.

Cancer diagnosis may require examination of the complete specimen, interpretation of tissue architecture and integration of relevant clinical information.

A pathologist may need to compare several slides, review additional tissue levels or request further investigations.

An algorithm developed for a particular detection task may not address all these requirements.

The DPA's 2026 recommendation statement emphasizes that clinical AI used in pathology should remain under qualified pathologist oversight [1].

AI may perform selected tasks effectively, but it does not replace the breadth of expertise required for comprehensive histopathological interpretation.

This distinction matters when hospitals introduce AI to improve efficiency.

A system may reduce the time needed to identify certain suspicious regions, but appropriate professional review still needs to take place.

The extent and form of that review may differ according to the validated application and applicable requirements.

However, an automated output does not independently remove professional responsibility for the final diagnosis.

Pathologists should be able to examine the underlying image, question the AI finding and use appropriate alternative methods when necessary.

If a digital image is inadequate, the laboratory should have a practical way to review the original glass slide or obtain another suitable image.

If the AI output conflicts with important morphological or clinical findings, the discrepancy should be addressed through professional judgment and, where appropriate, additional review.

The objective is not to require pathologists to agree with AI. It is to ensure that AI contributes to diagnostic care without weakening professional accountability.

AI should not create false efficiency gains

One reason hospitals consider digital pathology and AI is the possibility of improving efficiency.

Software may help identify selected suspicious regions, support case review or reduce time spent on particular repetitive tasks.

These functions may be valuable, especially in laboratories with growing workloads.

However, efficiency should not be measured only by how quickly an algorithm processes an image.

The complete diagnostic workflow includes specimen handling, processing, staining, scanning, quality checks, professional interpretation and communication of final findings.

If AI reduces image review time but increases rescanning, troubleshooting or the need to resolve unclear outputs, the overall benefit may be smaller than expected.

Similarly, an application may appear to improve turnaround time while generating unnecessary additional investigations or increasing the risk of missed findings.

Hospitals should therefore examine whether improvements in one part of the process create additional work elsewhere.

The CONFIDENT P study illustrates this distinction. AI assistance was associated with reduced immunohistochemistry use, but it did not significantly reduce the median time spent assessing each H&E slide [4].

The organization also needs to consider how AI affects case prioritization.

If software highlights certain cases as potentially malignant, those cases may receive earlier attention.

However, examinations not prioritized by AI must still receive appropriate professional review within clinically acceptable periods.

An AI-based prioritization system should not create excessive delays for cases considered lower priority, particularly when the model may fail to recognize some clinically important abnormalities.

The relevant measure is therefore not simply faster image analysis. It is whether AI improves the overall diagnostic service without compromising safety or appropriate access to care.

What laboratories should evaluate after introducing AI

A pathology laboratory should define the improvement it expects before introducing AI.

An organization may wish to improve detection of selected malignant findings, reduce clinically important diagnostic discrepancies or support more efficient examination of a defined group of cases.

These objectives require different evaluation measures.

Image quality is an important starting point. The laboratory should understand how often digital slides require rescanning and how frequently focus problems, incomplete tissue coverage or other artifacts affect professional or automated review.

AI processing reliability also matters.

The organization should identify how many eligible images are processed successfully, how many cannot be analyzed and what causes the failures.

Diagnostic performance should be evaluated according to the intended clinical task.

Depending on the application, relevant measures may include sensitivity, specificity, clinically important false-negative and false-positive findings, and agreement with a suitable reference standard.

However, agreement alone is not sufficient.

A system may agree frequently with routine diagnoses while failing to detect uncommon but clinically consequential lesions.

Where appropriate, the laboratory should consider whether AI changes the performance of pathologists compared with established practice.

The diagnostic workflow should also be evaluated. Relevant measures may include time to professional review, time to final diagnosis, additional tissue preparation, immunohistochemistry utilization, rescanning workload and time spent resolving AI-related discrepancies.

The denominator should be clear.

For example, a model may be evaluated per digital slide, tissue block, specimen or patient. These measures are not interchangeable.

A system may perform well at identifying malignant regions on individual slides while producing different results when evaluated across all specimens belonging to one patient.

Similarly, turnaround time should be defined precisely. The interval from slide scanning to pathologist sign-out is not equivalent to the full interval from specimen collection or receipt to final diagnosis.

Unexpected consequences deserve attention.

An AI application may produce numerous suspicious highlights that increase review time without identifying additional clinically meaningful lesions.

Conversely, a model may perform well with common tumor patterns but miss rare variants or small malignant foci.

These findings should inform whether the application remains suitable for continued use and whether its scope needs to be restricted.

Performance may also vary according to scanner, staining batch, tissue type or other clinically relevant characteristics.

Where diagnostic changes affect subsequent cancer care, longer-term evaluation may consider whether improved detection or workflow performance contributes to more appropriate clinical management.

However, patient outcomes are influenced by many factors beyond pathology AI, including disease stage, treatment availability and multidisciplinary care.

A reduction in turnaround time or an increase in suspicious regions identified should not automatically be interpreted as evidence of improved survival or treatment outcomes.

The success of AI-assisted histopathology should be assessed through diagnostic reliability, appropriate professional review, workflow consequences and patient safety—not simply algorithm accuracy.

Monitoring must continue after deployment

Initial validation does not guarantee that an AI application will continue performing identically under all future conditions.

Changes in scanners, staining procedures, image management systems and software versions may affect performance.

Even when the algorithm itself remains unchanged, the images supplied to it may become different.

A new staining reagent or modified tissue preparation procedure, for example, may alter the appearance of microscopic structures.

A replacement scanner may generate images with different color characteristics, optical resolution, compression or file formats.

These changes may affect the reliability of a model developed or validated under different conditions.

The laboratory should therefore establish when additional verification is necessary.

Not every minor technical modification requires complete repetition of the original validation. The extent of review should reflect the nature of the change, the intended clinical use and its potential consequences for patients.

Monitoring should also include clinically important concerns raised by pathologists.

Repeated false-positive highlights, suspected missed malignancies, unexpected changes in detection rates or increasing processing failures may indicate that further investigation is necessary.

The organization should determine how these findings are reviewed and what corrective actions are appropriate.

Depending on the circumstances, responses may include additional training, scanner maintenance, workflow modification, further verification or temporary suspension of AI use.

The laboratory should also consider how findings from routine quality review can identify recurring weaknesses without unnecessarily burdening staff.

This continuing assurance approach is consistent with the September 2026 guidance by Rakha and colleagues [2].

The purpose is to maintain acceptable diagnostic performance throughout the operational life of the application rather than rely indefinitely on evidence obtained before installation.

Digital pathology and AI in the Vietnamese healthcare context

Digital pathology and AI are receiving increasing professional attention in Viet Nam.

An article published on the Ministry of Health's information portal on April 15, 2026, discussed the application of AI in pathology and cancer screening and included perspectives from Vietnamese pathology specialists [9].

The discussion reflects growing interest in how AI might support diagnostic quality and laboratory efficiency.

However, professional interest, pilot activities and technology demonstrations should not be treated as proof that AI-assisted cancer histopathology is already widely deployed, independently validated or improving patient outcomes throughout Viet Nam.

Readiness may differ substantially between hospitals.

Some pathology laboratories may have digital scanners, image management systems and experience with remote consultation. Others may still rely mainly on conventional glass-slide microscopy.

The investment required for digital pathology extends beyond purchasing a scanner.

A laboratory needs appropriate image storage, information systems, workstations, technical support, staff training and arrangements for maintaining diagnostic services during equipment or software failures.

AI introduces additional requirements involving product selection, clinical verification, data handling, cybersecurity and continued monitoring.

These considerations are particularly important when a hospital intends to incorporate AI into routine diagnosis rather than use it for research or retrospective quality review.

A facility with limited digital pathology infrastructure may achieve more immediate benefit by strengthening specimen identification, tissue preparation, diagnostic review or access to specialist consultation before introducing AI.

Another laboratory with an established digital pathology service may be better positioned to evaluate a narrowly defined application.

The appropriate approach depends on the actual clinical need and organizational capabilities.

The question is not whether digital pathology AI is technologically advanced. It is whether a healthcare organization can use the selected application reliably within its existing diagnostic responsibilities.

Vietnamese professional and legal responsibilities

AI-assisted cancer histopathology must be considered within Viet Nam's legal framework governing medical examination and treatment, pathology services, medical devices, artificial intelligence and personal data protection.

These areas involve related but distinct responsibilities.

Authorization or successful clinical evaluation in another country does not automatically establish that an AI application may be used for every intended diagnostic purpose in Viet Nam.

Medical examination, treatment and pathology practice

The Law on Medical Examination and Treatment No. 15/2023/QH15, effective January 1, 2024, as subsequently amended, provides an important legal foundation for healthcare services and professional practice in Viet Nam [10].

The Law establishes responsibilities relevant to professional scope, clinical practice, patient rights and the quality and safety of healthcare services.

For pathology laboratories, a relevant Ministry of Health document is Decision No. 3123/QĐ-BYT, issued on October 2, 2025 [11].

The Decision promulgates Volume 1 of the Guidance on Technical Procedures in Pathology, comprising 69 technical procedures, and was originally scheduled to take effect on July 1, 2026.

However, Decision No. 1986/QĐ-BYT, issued by the Ministry of Health on July 1, 2026, amended the effective date of Decision No. 3123/QĐ-BYT to January 1, 2028 [12].

Accordingly, as of October 25, 2026, Decision No. 3123/QĐ-BYT should not be described as a technical procedure document already in force.

Its content may provide useful information for future preparation, but healthcare organizations must distinguish that future framework from the professional requirements currently applicable to their services.

The postponement also does not mean that pathology laboratories are exempt from existing requirements concerning technical procedures, authorized services, professional qualifications or patient safety.

A laboratory proposing to introduce WSI and AI should assess the technology against the applicable legal and professional framework in effect at the time of deployment.

Decision No. 3123/QĐ-BYT should not be interpreted as blanket authorization for autonomous AI cancer diagnosis or as an AI-specific validation protocol.

The use of AI does not independently expand the legally permitted scope of practice of healthcare personnel or remove the responsibilities associated with diagnostic interpretation.

A system introduced for research, education or retrospective quality review should also not automatically be considered authorized for routine clinical diagnosis.

The actual intended use must be assessed before software outputs influence patient care.

Artificial intelligence regulation

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [13].

Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions, including requirements concerning AI risk classification and management [14].

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

The applicable legal classification of an AI histopathology application depends on the system's actual functions, intended purpose, deployment arrangements and relevant regulatory criteria.

Software that highlights suspicious tissue regions for professional review may have a different regulatory profile from a system intended to generate autonomous diagnostic conclusions.

However, the absence of a separately named category for histopathology AI from a particular high-risk list should not be interpreted as definitive confirmation that every such application is legally low-risk.

Statutory AI classification must also be distinguished from clinical risk.

An application may create important patient-safety concerns if its outputs contribute to missed malignancy, incorrect diagnosis or delayed treatment, regardless of its formal legal classification.

The use of AI does not independently transfer or remove professional authority and responsibilities established under applicable law.

The Artificial Intelligence Law also includes transitional arrangements for certain systems already operating before its effective date. Their applicability should be assessed for the individual system rather than treated as a general exemption from healthcare, medical device or data protection requirements.

Hospitals should therefore consider both the legal obligations associated with the technology and the potential consequences of its use in clinical practice.

Medical device regulation

AI software intended for medical image analysis may fall within medical device regulation depending on its intended medical purpose and functions.

Decree No. 98/2021/NĐ-CP on medical device management, as amended by subsequent regulations including Decrees No. 07/2023/NĐ-CP and No. 04/2025/NĐ-CP, provides part of the relevant framework [16].

Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [17].

Before deployment, a healthcare organization should determine the regulatory status of the specific scanner, software and relevant integrated functions proposed for clinical use.

The regulatory status of a whole-slide scanner should not automatically be assumed to establish the permitted clinical use of an AI application connected to it.

Likewise, foreign regulatory authorization or published clinical validation does not independently demonstrate compliance with all Vietnamese requirements.

The assessment must relate to the actual product, its intended purpose and the clinical function it will perform.

Personal data protection

Whole-slide images and related pathology records may be linked to identifiable patient information, including specimen identifiers, medical records, diagnoses and treatment details.

The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [18, 19].

Healthcare organizations should assess the lawful basis and purposes of processing, applicable consent requirements or exceptions, access controls, information security, retention and arrangements with technology providers.

These responsibilities become particularly important when digital pathology images are uploaded to external platforms or processed through cloud services.

The laboratory should understand where data are stored, who can access them and whether the provider intends to use patient information or images for purposes beyond the immediate diagnostic service.

Whole-slide images may remain linkable to a patient even when a visible name has been removed.

Slide labels, embedded image data, metadata and related laboratory records may contain information capable of identifying the specimen or patient.

Removing a visible identifier should therefore not automatically be treated as sufficient anonymization.

Secondary use of pathology images for algorithm development, model training or commercial product improvement requires separate assessment under the applicable legal and institutional framework.

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

These legal assessments should support responsible implementation without weakening professional practice or patient safety.

A practical direction for pathology laboratories in Viet Nam

For hospitals considering AI-assisted cancer histopathology, implementation should begin with a clearly defined diagnostic or quality problem.

A laboratory may wish to improve detection of selected malignant features, reduce clinically important diagnostic discrepancies or support more efficient review of a specific group of cases.

These objectives require different types of evidence.

The organization should first assess the reliability of its existing histopathology process.

Specimen identification, tissue processing, sectioning, staining and professional slide review must be sufficiently controlled for reliable diagnosis.

Where digital pathology is already available, the laboratory should determine whether the whole-slide imaging process has been evaluated appropriately for the intended diagnostic activities.

The proposed AI application can then be considered separately.

Clinical leaders, pathologists, laboratory quality personnel and relevant technical teams should participate in reviewing the intended use and supporting evidence.

The laboratory should understand the tissue types, imaging requirements, diagnostic outputs and known limitations of the selected software.

Local verification should reflect the scanners, staining methods, image management arrangements and clinical workflow actually used.

Pathologists should receive appropriate training and remain able to review AI outputs within the complete clinical and morphological context.

Where permitted by applicable requirements, a supervised initial evaluation may help establish whether the software contributes useful information and whether unexpected diagnostic concerns arise.

The laboratory should also examine whether the technology increases rescanning, troubleshooting or additional review workload.

A reliable fallback must remain available when AI is unavailable, the digital image is inadequate or the original glass slide requires further examination.

The objective is not to make cancer diagnosis dependent on a single software platform.

It is to determine whether AI can strengthen an existing diagnostic service while preserving quality, professional accountability and patient safety.

From AI detection to reliable cancer diagnosis

AI-assisted histopathology offers opportunities to help pathologists identify selected malignant features and improve aspects of laboratory workflow.

However, its clinical value depends on more than algorithm performance.

Reliable cancer diagnosis begins with appropriate specimen collection and identification and continues through processing, staining, slide preparation, image acquisition and professional interpretation.

A technically accurate AI system cannot compensate for every error introduced earlier in this process.

Nor can an automated output replace the need to understand the complete morphological and clinical context.

Recent international guidance emphasizes that algorithm validation, local verification and continuing quality assurance are different but complementary responsibilities.

For healthcare organizations in Viet Nam, the challenge is to apply these principles within the practical realities of pathology services, including equipment, professional competence, quality management and the applicable legal framework.

The organization needs to understand what the AI application is intended to detect, whether the underlying images are suitable, how clinically important disagreements are resolved and whether the software improves the diagnostic process.

Evaluation should also consider workload, turnaround time, unexpected consequences and the effect of AI on professional judgment.

The central question is not whether an algorithm can highlight a suspicious tissue region.

It is whether the pathology laboratory can use that information to support accurate, timely and clinically appropriate diagnostic care.

AI-assisted cancer histopathology creates clinical value when reliable whole-slide imaging, appropriately validated AI analysis and accountable pathologist review work together to support diagnostic safety.

Professional Education Notice

This article is intended for professional education and knowledge exchange. It does not constitute individualized medical, pathology, oncology, laboratory, legal, regulatory or patient-safety advice.

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

This article focuses on AI-assisted detection and assessment of suspected malignancy in digitized H&E-stained histopathology sections, primarily within pathology services caring for patients with possible or confirmed cancer. It does not establish the safety or effectiveness of AI systems for autonomous cancer diagnosis, biomarker scoring, molecular testing, prognosis prediction, treatment selection or automated pathology report generation.

Different AI applications have different intended uses, supported tissue types, imaging requirements, diagnostic outputs and regulatory authorizations. Evidence supporting one application should not automatically be generalized to another.

International professional recommendations, including those from the Digital Pathology Association, the College of American Pathologists and the European Working Group for Breast Screening Pathology, provide relevant implementation guidance. They should not automatically be interpreted as binding requirements for pathology laboratories in Viet Nam.

ISO 15189:2022 and ISO/TS 23824:2024 are relevant international documents concerning medical laboratory quality and anatomic pathology. ISO/DIS 24051-2:2026 remains under development as of the review date and should not be treated as a published International Standard or an existing Vietnamese legal requirement.

Decision No. 3123/QĐ-BYT promulgated professional guidance on pathology technical procedures, but Decision No. 1986/QĐ-BYT postponed its effective date to January 1, 2028. As of October 25, 2026, it should not be presented as already effective or as an AI-specific clinical validation requirement.

The organizational approaches discussed, including slide quality assurance, local verification, pathologist review, discrepancy management and continuing performance monitoring, are proposed considerations for responsible AI implementation and healthcare quality improvement. They are not mandatory Vietnamese Ministry of Health AI histopathology protocols or independently validated DMVN assessment instruments.

Evidence showing improved algorithm accuracy, agreement, diagnostic turnaround time or workflow efficiency does not automatically establish that AI improves cancer treatment outcomes or patient survival.

The applicable professional, legal and regulatory obligations associated with individual AI systems, pathology services, medical device software, personal data processing and clinical research require separate assessment under the relevant Vietnamese framework.

References

1. Silberman N, Parwani A, McClintock DS, et al. Recommendation Statement for the Validation, Implementation, and Clinical Application of Artificial Intelligence Within a Clinical Laboratory From the Digital Pathology Association. AI in Precision Oncology. 2026;3(4):71–97. DOI: 10.1177/2993091X261455975.

2. Rakha EA, Wesseling J, Kovács A, et al. Guidance for Laboratory Implementation, Governance and Continuous Assurance of Artificial Intelligence in Histopathology. Virchows Archiv. Published online September 2, 2026. DOI: 10.1007/s00428-026-04684-y.

3. Evans AJ, Brown RW, Bui MM, et al. Validating Whole Slide Imaging Systems for Diagnostic Purposes in Pathology. Archives of Pathology & Laboratory Medicine. 2022;146(4):440–450. DOI: 10.5858/arpa.2020-0723-CP.

4. Flach RN, van Dooijeweert C, Nguyen TQ, et al. Prospective Clinical Implementation of Paige Prostate Detect Artificial Intelligence Assistance in the Detection of Prostate Cancer in Prostate Biopsies: CONFIDENT P Trial Implementation of Artificial Intelligence Assistance in Prostate Cancer Detection. JCO Clinical Cancer Informatics. 2025;9. DOI: 10.1200/CCI-24-00193.

5. Udoh AI, Eyzaguirre E, Muthukumarana V, Thaker HM. Validation, Implementation, and Impact of an AI Model in Routine Practice for Pathologic Diagnosis of Prostate Cancer in an Academic Medical Center. Journal of Pathology Informatics. 2026;22:100675. DOI: 10.1016/j.jpi.2026.100675.

6. International Organization for Standardization. ISO 15189:2022. Medical Laboratories — Requirements for Quality and Competence. Fourth edition. Published December 2022.

7. International Organization for Standardization. ISO/TS 23824:2024. Medical Laboratories — Guidance on Application of ISO 15189 in Anatomic Pathology. Published May 2024.

8. International Organization for Standardization. ISO/DIS 24051-2:2026. Medical Laboratories — Part 2: Digital Pathology and Artificial Intelligence (AI)-Based Image Analysis. Draft International Standard, under development as of October 25, 2026.

9. Ministry of Health of Viet Nam. Ứng dụng trí tuệ nhân tạo trong giải phẫu bệnh, tầm soát ung thư. Ministry of Health Information Portal. Published April 15, 2026.

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

11. Ministry of Health of Viet Nam. Decision No. 3123/QĐ-BYT Promulgating the Professional Document “Guidance on Technical Procedures in Pathology,” Volume 1. Issued October 2, 2025. Effective date subsequently amended by Decision No. 1986/QĐ-BYT to January 1, 2028.

12. Ministry of Health of Viet Nam. Decision No. 1986/QĐ-BYT Amending the Effective Dates and Application Dates of Certain Ministry of Health Decisions Promulgating Professional Technical Procedure Guidelines. Issued July 1, 2026; effective July 1, 2026.

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

14. 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.

15. 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.

16. 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.

17. 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.

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

19. 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.