Governing AI-Assisted Radiotherapy Auto-Contouring in Viet Nam: Organ-at-Risk Contour Verification, Clinical Approval and Treatment Safety

Governing AI-Assisted Radiotherapy Auto-Contouring in Viet Nam: Organ-at-Risk Contour Verification, Clinical Approval and Treatment Safety

IMPLEMENTATION INSIGHTS

Last reviewed: 28 October 2026

A patient with head and neck cancer is preparing to receive radiotherapy. Following treatment-planning computed tomography (CT), an artificial intelligence (AI) system automatically generates contours for several organs at risk, including the spinal cord, brainstem, parotid glands, and oral cavity.

At first glance, the contours appear reasonable. During a closer review, however, the radiation oncologist notices that the spinal cord contour does not accurately follow the anatomy across several CT slices near the intended high-dose region.

The discrepancy is relatively small, but its clinical significance may not be. If the contour is accepted without correction, the treatment planning system may calculate dose-volume information using an inaccurate representation of the spinal cord. The reported dose could therefore provide misleading reassurance about the radiation exposure of the actual organ.

The finding raises several questions. Should the contour be corrected before planning proceeds? Who is responsible for approving the final structure? If the discrepancy is discovered after the treatment plan has been optimized or approved, does the plan require reassessment, modification, or renewed approval?

These situations illustrate why AI-assisted auto-contouring requires more than accurate anatomical segmentation. Automatic contouring can reduce repetitive work and provide a useful starting point, but a structure that looks convincing is not necessarily suitable for clinical use. Its acceptability depends on the organ involved, the location of any discrepancy, the treatment technique, and the potential consequences for dose evaluation.

For healthcare organizations in Viet Nam, responsible implementation means ensuring that AI-generated contours are appropriately verified, corrected when necessary, approved through the established professional process, and reliably incorporated into treatment planning.

The essential distinction is between a contour generated by software and an anatomical structure that has been reviewed and approved for clinical use.

Why organ-at-risk contouring matters for treatment safety

Radiotherapy aims to deliver a prescribed radiation dose to a defined treatment target while limiting unnecessary exposure to surrounding healthy tissues.

Modern techniques such as intensity-modulated radiation therapy (IMRT), volumetric modulated arc therapy (VMAT), and stereotactic radiotherapy can produce highly conformal dose distributions. Their clinical effectiveness depends on accurate imaging, target definition, anatomical delineation, dose calculation, treatment planning, and delivery.

Organs at risk (OARs) are normal tissues whose radiation exposure requires particular consideration because excessive dose may contribute to clinically significant toxicity.

Depending on the treatment site, these structures may include the spinal cord, brainstem, optic nerves, heart, lungs, esophagus, kidneys, bowel, bladder, rectum, salivary glands, and other tissues.

The required OARs depend on the patient's anatomy, disease site, treatment technique, prescribed dose, fractionation, and applicable clinical protocol. Their contours provide anatomical information used during treatment optimization and dose-volume evaluation.

However, an OAR contour is not itself a measurement of delivered radiation dose. It is an anatomical representation used by the treatment planning system to calculate and summarize planned dose within a defined region.

If that representation is inaccurate, the resulting dose-volume information may be misleading even when the dose calculation algorithm is functioning correctly.

For example, a spinal cord contour that omits a relevant anatomical segment may underestimate the calculated near-maximum dose to the actual organ. Conversely, a contour that unnecessarily includes surrounding tissues may lead to overly restrictive planning decisions and potentially compromise target coverage.

The consequences are not identical for every OAR. A small boundary error near a steep dose gradient may be important for a serial organ such as the spinal cord, while a larger discrepancy in a low-dose region may have little influence on treatment plan acceptability.

This is why the clinical quality of OAR contours cannot be determined by their visual appearance or geometric similarity alone. The relevant question is whether the anatomical representation is sufficiently accurate for its intended role in the patient's treatment plan.

What AI-assisted auto-contouring can and cannot establish

Auto-contouring refers to the automatic generation of anatomical contours on medical images.

Systems may use atlas-based segmentation, conventional image processing, machine learning, or deep learning. Contemporary AI applications commonly analyze planning CT images and generate contours corresponding to predefined anatomical structures.

These contours can then be imported into a radiotherapy treatment planning system for review, correction, and subsequent clinical use.

The potential benefit is straightforward. Manual delineation of multiple OARs can be time-consuming, particularly in anatomically complex regions. AI may reduce the initial workload and provide a more consistent starting point for professional assessment.

However, auto-contouring applications differ in their intended purposes and technical capabilities.

A model developed to identify normal organs on planning CT may not be validated for magnetic resonance imaging (MRI), cone-beam CT, or another imaging modality. Similarly, software trained on thoracic anatomy may not be appropriate for head and neck or pelvic structures.

There is also an important distinction between OAR contouring and tumor target delineation.

Gross tumor volume (GTV), clinical target volume (CTV), and planning target volume (PTV) have different clinical meanings. Their delineation involves decisions about visible disease, possible microscopic extension, treatment margins, and planning uncertainties.

Evidence supporting AI-assisted segmentation of normal organs should not automatically be generalized to tumor targets.

An automatically generated OAR contour also does not establish that an appropriate radiation dose has been prescribed or that the treatment plan is clinically acceptable.

This article focuses primarily on AI-generated OAR contours used during radiotherapy planning, the professional verification required before clinical use, and the relationship between contour approval and treatment safety. It does not establish the safety or effectiveness of autonomous tumor target delineation, automated dose prescription, or AI-controlled radiation delivery.

What the latest clinical evidence tells us

Evidence supporting AI-assisted auto-contouring has progressed beyond retrospective comparisons between algorithm-generated structures and expert reference contours.

An important development appeared in March 2026, when researchers published a prospective multicenter observational study in Nature Communications examining deep learning-based OAR segmentation for thoracic and breast cancer radiotherapy [3].

The study enrolled 500 patients across five centers in China and involved 37 physicians. Three patients were subsequently excluded from the final analysis because of technical problems. Two had insufficient CT scan coverage, while one experienced a model generation failure, leaving 497 evaluable patient datasets.

The investigators examined 11 thoracic OARs and compared manual, AI-generated, and AI-assisted delineation using geometric performance measures and contouring time.

AI-assisted contours achieved a mean volumetric Dice similarity coefficient of 0.902, compared with 0.857 for manual contours when assessed against the study's expert consensus reference.

The mean 95th-percentile Hausdorff distance was 5.20 mm for AI-assisted delineation and 8.01 mm for manual delineation.

Median contouring time decreased from 55 minutes with manual delineation to 10 minutes with AI assistance, representing an 81.63% improvement in time efficiency as reported by the investigators.

The study also found reduced variability in selected performance measures across participating centers and physicians with different levels of experience.

These findings provide encouraging evidence that AI-assisted contouring can improve efficiency and anatomical agreement under defined conditions. However, the study evaluated a particular AI model, selected anatomical structures, and a specific clinical population. It was not a randomized trial designed to demonstrate improved patient outcomes.

The expert consensus contours provided a useful reference for comparison, but they should not be considered an infallible representation of anatomical truth.

Most importantly, improved geometric agreement and shorter contouring time do not independently establish fewer radiation-related complications, better tumor control, or improved survival.

A different perspective comes from a 2026 study published in Radiotherapy and Oncology, which examined the dosimetric implications of deep learning-based OAR segmentation using data from the HaN-Seg challenge [4].

The researchers evaluated six segmentation approaches on 14 head and neck test cases involving 30 OARs, using paired CT and MRI data.

Their analysis extended earlier geometric comparisons by examining whether dose-related restrictions were satisfied when different contours were used for evaluation.

Across the evaluated approaches, approximately 67.7% to 70.7% of automatically segmented OARs met the study's defined dosimetric restrictions.

An important qualification is necessary. The curated reference contours also achieved approximately 70.7% compliance.

Therefore, results that did not satisfy the restrictions cannot simply be attributed to AI segmentation errors. Some restrictions were difficult to meet because of treatment geometry, target proximity, dose-distribution characteristics, and the constraints used in the evaluation.

The investigators also found no clear overall relationship between geometric performance measures and dosimetric performance.

Structures near tumor regions and steep dose gradients were particularly relevant because relatively small boundary differences could influence dose-related assessments.

These findings support the complementary use of geometric and dosimetric evaluation. They do not establish that the assessed algorithms caused clinical overdoses, radiation toxicity, or other adverse patient outcomes.

The study used a forward dosimetric analysis, sometimes described as dosimetric impact approximation, rather than prospectively measuring the clinical consequences of treatment based on AI-generated contours.

Together, these studies illustrate both the promise and limitations of AI-assisted auto-contouring. AI may improve anatomical consistency and reduce repetitive work, but its clinical usefulness cannot be established through contour overlap measures alone.

Why geometric accuracy does not guarantee clinical acceptability

AI auto-contouring studies commonly report metrics such as Dice similarity coefficient, surface Dice, Hausdorff distance, and volume differences.

These measures describe how closely an automatically generated contour resembles a reference structure. They are useful for model development and evaluation, but each has limitations.

The Dice similarity coefficient measures volumetric overlap. It provides an overall indication of agreement, but its interpretation depends on the size and shape of the anatomical structure.

A relatively small boundary discrepancy may have little influence on the Dice score of a large organ. The same discrepancy may have a much greater effect on a small structure.

Hausdorff distance measures separation between contour boundaries and may help identify localized errors. However, maximum boundary-distance measures can be sensitive to isolated outliers.

Neither measure independently determines whether a contour is suitable for a particular treatment plan.

A small discrepancy adjacent to a high-dose target may be clinically more important than a larger difference in a region receiving little radiation.

For example, an OAR boundary error close to the PTV may affect the reported maximum or near-maximum dose. A similar error elsewhere in the same organ may have little effect on the final planning decision.

The Royal College of Radiologists (RCR) guidance statements published in Clinical Oncology in 2026 emphasize the importance of combining geometric, qualitative, dosimetric, and time-based assessments during clinical commissioning [1].

The joint European Society for Radiotherapy and Oncology (ESTRO) and American Association of Physicists in Medicine (AAPM) guideline published in 2024 also addresses the development, clinical validation, reporting, and quality assurance of AI models in radiotherapy [6].

These professional recommendations support an approach in which technical performance is evaluated alongside clinical relevance and intended use.

For healthcare organizations, contour validation should therefore consider both anatomical agreement and potential consequences for treatment planning.

A high Dice score may support confidence in an algorithm's general performance, but it cannot replace professional review of clinically important structures in individual patients.

Organ-at-risk verification should prioritize clinically important boundaries

A practical OAR verification process begins by identifying which anatomical structures are important for the patient's treatment.

The reviewer needs to understand how each required organ should be delineated and which boundaries may be particularly relevant to dose evaluation.

In head and neck radiotherapy, careful attention may be required around the spinal cord, brainstem, optic structures, salivary glands, oral cavity, and swallowing-related tissues.

Thoracic treatment may require assessment of the lungs, heart, esophagus, spinal cord, and other nearby structures. Pelvic radiotherapy introduces different considerations involving the bladder, rectum, bowel, femoral heads, and other organs.

The appropriate list depends on the treatment site and clinical protocol.

An AI application may generate many contours, but not all are necessarily required for every patient. Conversely, a clinically important OAR may be missing from the automatically generated set and require manual delineation.

Reviewers should examine whether the contours are anatomically plausible, complete, and correctly positioned.

Potential problems include omitted anatomical segments, incorrect superior or inferior boundaries, discontinuities across image slices, inappropriate inclusion of adjacent tissues, and incorrect left–right identification.

The review should also distinguish an organ contour from related planning structures, such as a planning organ-at-risk volume (PRV), where applicable.

The RCR guidance recommends a structured approach that includes reviewing contours in multiple imaging planes and paying particular attention to structures and regions of critical dosimetric importance [1,2].

Reviewing axial, sagittal, and coronal views can reveal errors that may not be apparent on a single image slice. The appropriate imaging window and modality also matter, particularly for structures that are difficult to distinguish on standard planning CT.

The clinical significance of a discrepancy depends on the organ involved, fractionation, prescribed dose, dose distribution, and patient-specific circumstances.

For that reason, applying the same geometric acceptance threshold to every structure may not be appropriate.

A standardized review process should be consistent while recognizing that different anatomical errors have different clinical consequences.

Difficult anatomy and automation bias require attention

Auto-contouring performance can vary considerably across organs and clinical situations.

Large structures with relatively clear boundaries may be easier for an algorithm to identify than small organs, elongated structures, or tissues with poorly defined margins.

Postoperative changes, abnormal anatomy, large tumors, medical implants, motion artifacts, and differences in imaging protocols may further affect performance.

For example, dental implants can introduce CT artifacts that obscure head and neck anatomy. Previous surgery may alter normal tissue relationships, while large tumors can displace nearby organs.

Patients undergoing reirradiation may have anatomical changes that differ from those represented in the model's development data.

An AI-generated contour may still look smooth and convincing in these circumstances, even when it is inaccurate.

This creates a risk of automation bias. A reviewer may assume that a visually coherent contour is correct without examining the anatomy closely enough.

The problem may become less noticeable as clinicians become familiar with software that usually performs well. Repeated exposure to plausible contours can influence expectations and potentially reduce attention to unusual errors.

The RCR guidance recommends a structured review process, appropriate training, and awareness of common model failures [1,2].

Reviewers should understand the software's limitations and examine whether the patient's anatomy and imaging conditions fall within its validated scope.

Some cases may require substantial manual correction. Others may be more appropriately contoured without AI assistance.

A reliable service should preserve the ability to choose another method when the patient's circumstances require it.

Clinical contour approval is a distinct safety decision

Generating an OAR contour, reviewing it, and approving it for clinical use are separate activities.

AI software may produce the initial structure. An appropriately trained professional may review and edit it. Final clinical approval must occur through the healthcare organization's established professional process.

The RCR's 2024 guidance and subsequent 2026 consensus publication emphasize that the healthcare professional approving auto-contours retains responsibility for their clinical use [1,2].

The guidance also recognizes that specific tasks may be performed by different appropriately trained professionals according to their roles and relevant regulatory arrangements.

This approach should be interpreted within the applicable Vietnamese professional framework rather than treated as automatic authorization to delegate clinical responsibilities.

In radiotherapy practice, the radiation oncologist plays a central role in determining treatment objectives and accepting relevant anatomical and clinical considerations.

Medical physicists provide essential expertise in treatment planning systems, dose calculation, technical verification, and radiation safety.

Dosimetrists, radiation therapists, and other qualified professionals may contribute to contour preparation, planning, and review according to their training and authorized responsibilities.

The exact allocation of tasks should reflect the institution's approved clinical pathway and applicable professional requirements.

The approval process should identify which structures have been reviewed, whether corrections were required, and which versions are authorized for planning.

It should also distinguish clinical contour approval from treatment plan approval.

An anatomically acceptable contour does not establish that the resulting dose distribution is appropriate. Likewise, a technically acceptable dose calculation cannot compensate for an inaccurate representation of an important organ.

Clinical contour approval and treatment plan approval are connected, but they should remain distinct and identifiable decisions within the radiotherapy workflow.

The approved contour must remain connected to the treatment plan

A clinically important safety concern arises when several versions of a contour exist within the treatment planning system.

A structure may first be generated by AI, modified by a clinician, and then changed again during optimization.

If the team cannot reliably identify which version has been approved, a treatment plan may be evaluated using an outdated or incorrect structure.

This risk becomes more significant when images and contours are transferred between separate software platforms.

The organization should maintain reliable links between the patient, planning image, structure set, clinical approval, and treatment plan.

The approved contour set should be clearly identifiable to professionals responsible for optimization and dose evaluation.

If a clinically significant contour is corrected after planning has begun, the team needs to determine whether dose-volume metrics, treatment optimization, or plan acceptability require reassessment.

For example, correcting a spinal cord contour may reveal that the previously calculated near-maximum dose did not adequately represent the actual organ. The revised information may change whether the plan is considered clinically acceptable.

In another situation, correction of an overly restrictive OAR contour may allow the team to improve target coverage while maintaining appropriate normal-tissue protection.

The necessary response depends on the affected structure, the location and extent of the correction, and the resulting dosimetric consequences.

An important technical distinction should be preserved. Changing a contour does not necessarily change the physical dose distribution already calculated for an unchanged beam arrangement. However, it may change the dose-volume statistics and clinical interpretation of that distribution.

If the plan is reoptimized or other planning parameters are modified, dose recalculation and additional verification may be necessary.

A corrected contour should therefore not simply replace an earlier version without considering whether previous dose evaluations and approvals remain valid.

If a clinically important change occurs after plan approval, reassessment and renewed approval may be required before treatment proceeds.

If treatment has already started, the responsible clinical and medical physics team must assess the potential implications for delivered and remaining treatment, including whether further clinical review, replanning, or other corrective action is necessary.

The objective is to ensure that the treatment delivered to the patient reflects anatomical information and clinical decisions that have actually been reviewed and approved.

Medical physics review provides a complementary safeguard

Clinical contour verification and medical physics plan review address related but different risks.

Clinical review evaluates whether the anatomical structures are appropriate for their intended use.

Medical physics review examines relevant technical and dosimetric aspects of the treatment plan, including calculation integrity, dose evaluation, data consistency, and treatment delivery considerations.

AAPM Task Group 275 provides recommendations for effective physics plan and chart review [9].

The report emphasizes standardized, risk-informed verification before the first treatment fraction and appropriate checks during treatment.

However, physics review should not be interpreted as replacing professional clinical assessment of anatomical structures. Similarly, approval of OAR contours does not replace technical plan verification.

AAPM Task Group 100 provides a broader framework for identifying weaknesses in radiotherapy processes through process mapping and failure mode and effects analysis [10].

For AI-assisted contouring, potential failures may include incorrect patient identification, unsupported imaging inputs, missing structures, anatomical errors, inappropriate approval, or use of an outdated contour set.

These failures occur at different stages and require different safeguards.

The clinical and technical teams need to understand where their responsibilities intersect, particularly when an anatomical discrepancy is discovered after planning.

A reliable process connects clinical and technical verification without assuming that any single approval step can identify every possible problem.

Image registration and structure naming are part of contour safety

Auto-contouring may involve information from several imaging studies.

Planning CT is commonly used for dose calculation, while MRI, positron emission tomography (PET), or other imaging may provide additional information about anatomy and disease.

When images or structures are transferred between imaging modalities or different time points, registration becomes important.

Depending on the clinical purpose, registration may involve rigid alignment or deformable transformation.

AAPM Task Group 132 provides recommendations for quality assurance of image registration and fusion in radiotherapy [8].

The report emphasizes understanding registration uncertainty and verifying patient-specific results before using them for clinical decisions.

This is relevant to AI-assisted contouring because an anatomically reasonable contour on one image may become inaccurate if transferred incorrectly to another.

A structure can also be associated with the wrong image series, orientation, coordinate system, or patient record. Such errors may not be obvious during a superficial visual review.

The clinical team should therefore verify that transferred structures correspond to the correct patient and imaging dataset and remain anatomically appropriate after registration.

Structure naming is another important consideration. An accurately drawn contour may still create problems if its name, laterality, or intended meaning is incorrect.

For example, confusion between left and right organs, between an organ and its PRV, or between different composite structures may affect automated planning objectives and dose-volume reporting.

AAPM Task Group 263 recommends standardized nomenclature for radiation oncology structures and dosimetric information [7].

Consistent naming can improve communication between systems and reduce ambiguity in planning. This becomes particularly important when hospitals use automated scripts or planning templates that assign objectives according to structure names.

A naming error may influence downstream calculations even when contour geometry is accurate.

Hospitals should establish how AI-generated structure names are mapped to approved planning terminology and how missing, duplicate, or incorrectly labeled structures are identified.

Correct naming supports safe planning, but it does not independently establish anatomical or dosimetric acceptability.

Local commissioning should reflect the actual radiotherapy service

Performance demonstrated in an international study may not transfer directly to a Vietnamese radiotherapy department.

Differences in CT equipment, acquisition parameters, contrast protocols, reconstruction methods, anatomical characteristics, and local contouring conventions can influence model performance.

The RCR guidance recommends clinical commissioning using representative local data and assessment of the intended anatomical structures and clinical protocols [1,2].

Its 2026 consensus recommendations include a suggested commissioning requirement for a local dataset, independent of model development, containing at least 20 representative cases per relevant clinical protocol.

This is a professional recommendation from the RCR, not a universally mandated Vietnamese legal threshold.

The appropriate commissioning approach should reflect the particular software, clinical application, available evidence, patient population, and applicable requirements.

Commissioning should establish where the AI system may be used and where additional restrictions are appropriate.

A hospital needs to understand which organs the model supports, which imaging inputs are acceptable, and which patients or clinical situations fall outside its validated scope.

Evaluation should include anatomical review, relevant geometric measures, clinically meaningful assessment of dosimetric consequences, and the time required for review and correction.

A model that generates contours rapidly but requires extensive editing may provide less operational benefit than expected.

Similarly, good performance for common anatomical structures does not establish that the system is equally reliable for small organs or unusual anatomy.

Technical integration also matters. Accurate segmentation may provide little benefit if the software introduces errors during image transfer, structure naming, or interaction with the treatment planning system.

Local commissioning should therefore examine the complete workflow rather than assess the algorithm in isolation.

Acceptance criteria should be established by appropriately qualified professionals, informed by intended use and relevant evidence.

The joint ESTRO–AAPM guideline also supports transparent validation, documented model limitations, and appropriate consideration of implementation and model changes [6].

The purpose of commissioning is to determine whether a particular system can be used reliably within the hospital's actual clinical environment.

Time savings should include professional review and correction

Reducing contouring time is one of the most frequently reported benefits of AI-assisted segmentation.

However, the time required to generate contours is not equivalent to the time needed to produce clinically approved structures.

The complete workflow includes image preparation, contour generation, anatomical review, correction, confirmation of required structures, and approval.

A system may produce an initial set of contours in seconds while the clinical team still requires substantial time to assess difficult anatomy.

Additional verification tasks may also offset some of the time saved during initial delineation.

A 2025 study in Physics and Imaging in Radiation Oncology examined the clinical use of deep learning-based OAR auto-segmentation over 18 months [5].

The audit included 1,255 patients from a single institution.

The investigators reported an increase in mean surface Dice similarity between initial and edited contours, from 0.87 to 0.97, and an increase in the proportion of OARs left unedited, from 21.5% to 40%.

The analysis identified changes associated with vendor model updates, evolving local contouring practices, and reduced editing in areas considered to have limited clinical significance.

These findings illustrate how routine audit data can reveal changes in clinical use.

However, fewer edits do not automatically establish better contour quality.

A lower editing rate may reflect improved model performance, greater familiarity with the software, changes in local conventions, or inadequate attention to errors.

Hospitals should therefore interpret editing rates alongside clinical review findings and relevant dosimetric consequences.

The objective is not to minimize corrections. It is to ensure that necessary corrections are made while avoidable manual work is reduced.

Adaptive radiotherapy introduces additional considerations

Adaptive radiotherapy adds complexity because anatomical conditions may change during a course of treatment.

Changes in tumor size, organ position, patient weight, and bladder or bowel filling can affect the relationship between the planned dose distribution and the patient's current anatomy.

AI-assisted auto-contouring may support adaptive workflows by generating or updating structures on newly acquired images.

However, evidence supporting auto-contouring on conventional planning CT should not automatically be applied to adaptive radiotherapy.

Adaptive workflows may involve different imaging modalities, image quality, registration methods, and time constraints.

Online adaptive radiotherapy can create particular pressure because contour review, replanning, and approval may need to occur within a short clinical window.

The need for rapid decisions does not eliminate the importance of verifying clinically significant anatomical boundaries. Instead, it increases the importance of clearly defined responsibilities, appropriate commissioning, and reliable safeguards.

AAPM Task Group 384 is developing guidance on the clinical implementation of automated segmentation for adaptive radiotherapy, including commissioning, validation, patient-specific quality assurance, and management of AI-related risks.

As of October 2026, the task group's stated objectives should not be treated as equivalent to a completed published consensus report.

Healthcare organizations considering adaptive applications should rely on applicable published guidance and evaluate the specific workflow independently.

The same principle applies to automated contour propagation and dose accumulation, which introduce additional uncertainties requiring appropriate validation.

Monitoring performance after implementation

An AI auto-contouring system may perform acceptably during initial commissioning but behave differently as clinical conditions change.

Software updates, changes in CT equipment, reconstruction protocols, image processing, or patient populations may influence its outputs.

New treatment techniques may also change which OARs are important and how accurately they need to be delineated.

Ongoing evaluation should therefore examine whether the system remains suitable for its intended use.

Hospitals may monitor clinically significant corrections, repeatedly inaccurate structures, unexpected model failures, and cases in which auto-generated contours cannot be safely used.

The consequences of errors matter more than their number alone.

Frequent minor adjustments in regions with little dosimetric relevance may be less concerning than a single important error affecting a critical organ near the high-dose region.

Monitoring should also identify changes associated with software updates.

A vendor update may improve segmentation for some structures while changing performance for others. The organization should determine whether additional verification is necessary before using updated software routinely.

Healthcare professionals should have a practical process for reporting concerns when contours appear unreliable or when an error influences treatment planning.

Corrective action may involve changes to data processing, clinical procedures, training, software configuration, or the scope of use.

Where reliable clinical use cannot be maintained, temporarily restricting or suspending the affected function may be appropriate.

A manual contouring pathway should remain available when AI is unsuitable or unavailable.

The purpose of monitoring is to maintain a dependable clinical service rather than assume that software remains safe indefinitely after initial approval.

The Vietnamese radiotherapy context

AI-assisted auto-contouring may be relevant to Vietnamese hospitals providing radiotherapy for head and neck, thoracic, breast, abdominal, pelvic, and other cancers.

Its potential value depends on the capabilities of each service.

Some radiotherapy departments have established planning systems, experienced radiation oncologists, medical physicists, and structured quality assurance processes.

Others may face limitations in staffing, specialized software, imaging integration, or resources for extensive commissioning.

These differences matter because introducing AI redistributes work rather than eliminating professional responsibilities.

Automatic contour generation may reduce manual delineation time, but clinical review, approval, and technical verification remain necessary.

For a department with limited planning capacity, AI may help reduce repetitive work. However, introducing the technology without adequate review capacity could create a new problem: more generated structures entering the planning process without sufficient verification.

The organization should therefore identify where delays or quality concerns arise in its existing workflow.

If inconsistent contouring practice is the principal problem, standardizing anatomical guidance and strengthening peer review may provide immediate benefits.

If errors occur during image transfer or plan approval, improving those processes may be more important than introducing automatic segmentation.

AI should be selected to address an identified clinical or operational need rather than deployed simply because the technology is available.

Its role should support the existing responsibilities of qualified healthcare professionals, not create a parallel process in which automatically generated structures are assumed to be ready for treatment planning.

Vietnamese legal and professional responsibilities

Introducing AI-assisted auto-contouring into radiotherapy requires consideration of Vietnamese laws and professional requirements governing medical examination and treatment, radiation safety, artificial intelligence, medical devices, and personal data protection.

These areas are related but distinct.

Evidence supporting a particular algorithm does not automatically establish that a product may be used for every proposed clinical purpose in Viet Nam.

The relevant assessment depends on the actual software, its intended medical purpose, deployment arrangements, the data it processes, and the clinical decisions its outputs may influence.

Medical examination, treatment, and professional accountability

Viet Nam's Law on Medical Examination and Treatment No. 15/2023/QH15, effective January 1, 2024, as amended, provides an important foundation for healthcare services, professional practice, patient rights, and clinical responsibilities [12].

Hospitals considering AI-assisted radiotherapy planning should ensure that its proposed use is consistent with their authorized services, the professional scope of the practitioners involved, and applicable technical requirements.

An AI-generated OAR contour does not independently establish that a structure is suitable for treatment planning. Nor does it expand a practitioner's legally permitted professional scope or remove the responsibilities associated with clinical approval and treatment decisions.

The Ministry of Health issued Decision No. 1585/QĐ-BYT on June 3, 2026, promulgating Guidance on Radiotherapy Technical Procedures – Volume 1, comprising 34 technical procedures [13].

The decision provides national professional guidance for the procedures within its scope and took effect upon issuance.

Decision No. 1586/QĐ-BYT, also issued on June 3, 2026, promulgated Guidance on Radiotherapy Technical Procedures – Volume 2, comprising eight technical procedures [14].

An important subsequent amendment must be recognized.

Decision No. 1986/QĐ-BYT, issued on July 1, 2026, changed the specified application date of Radiotherapy Volume 2 to January 1, 2028 [15].

Accordingly, Volume 2 should not be described as a set of technical procedures already required to be applied from July 2026.

Healthcare organizations should determine which technical procedures and professional requirements are currently applicable to their authorized radiotherapy services.

These Ministry of Health documents should not automatically be interpreted as validation or authorization of any particular AI auto-contouring product.

Where hospitals develop internal procedures for AI-assisted contouring, those procedures should be assessed against the relevant legal framework and professional requirements.

Radiation safety and atomic energy regulation

Radiotherapy involves the medical use of ionizing radiation and therefore requires consideration of radiation protection and safety requirements.

Viet Nam's Law on Atomic Energy No. 94/2025/QH15 took effect on January 1, 2026, replacing the previous atomic energy law subject to applicable transitional provisions [16].

Decree No. 332/2025/NĐ-CP, also effective January 1, 2026, provides implementing provisions concerning radiation safety, nuclear safety and security, licensing, inspection, incident response, and related responsibilities [17].

Healthcare organizations must assess the requirements applicable to their actual radiation activities, equipment, and authorized services.

AI-assisted contouring does not itself deliver radiation, but inaccurate contours may affect the information used to optimize and evaluate a treatment plan.

The technology should therefore be considered within the radiotherapy quality and safety system rather than treated solely as administrative software.

International guidance from the International Atomic Energy Agency (IAEA), including Specific Safety Guide SSG-46, provides an additional reference on radiation protection and safety in medical uses of ionizing radiation [11].

Such guidance can inform professional practice but should not be described as automatically replacing applicable Vietnamese legal requirements.

Artificial intelligence regulation

Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [18].

Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides implementing provisions concerning AI activities, including risk classification and management [19].

Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, promulgates a list of high-risk AI systems [20].

The legal classification of a radiotherapy auto-contouring system depends on its actual functions, intended purpose, deployment arrangements, and applicable criteria.

Software that generates draft OAR contours for professional review may raise different regulatory and clinical considerations from a system intended to approve structures autonomously or make treatment planning decisions.

However, the absence of a particular auto-contouring function from a named high-risk list does not independently establish that the software is legally low-risk or exempt from other relevant obligations.

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

A system may contribute to important patient-safety consequences if its outputs result in inaccurate anatomical representations, inappropriate planning decisions, or treatment delays.

Hospitals should therefore evaluate both regulatory compliance and the potential clinical consequences of using the software.

Medical device regulation

AI software intended to delineate anatomical structures for radiotherapy planning may fall within Vietnamese medical device regulation, depending on its intended medical purpose and functions.

Decree No. 98/2021/NĐ-CP on medical device management, as amended, including by Decrees No. 07/2023/NĐ-CP and No. 04/2025/NĐ-CP, forms part of the relevant framework [21].

Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [22].

Healthcare organizations should determine the regulatory status applicable to the specific AI software before routine clinical deployment.

The regulatory status of a treatment planning system or radiotherapy machine does not automatically establish that separately integrated AI software may be used for every proposed medical purpose.

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

The assessment should reflect the actual software version, intended use, and clinical decisions it is expected to support.

Commercial, open-source, and internally developed systems may involve different arrangements, but none should automatically be assumed exempt from relevant professional or medical device requirements solely because of its development model.

Personal data protection

Radiotherapy planning images, structure sets, and associated clinical records may contain or be linked to identifiable patient information.

These records can include medical images, examination dates, diagnoses, treatment details, patient identifiers, and other health information.

Viet Nam's Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [23,24].

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

Particular attention is necessary when CT images or radiotherapy planning information are transferred to external AI platforms or processed through cloud infrastructure.

Hospitals should understand where information is processed and stored, who can access it, and whether it may be used beyond the intended clinical service.

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

Secondary use of radiotherapy images for model development, retraining, commercial product improvement, or research requires separate consideration of applicable legal and institutional requirements.

Removing patient names or identification numbers does not necessarily establish that the underlying imaging data and associated metadata are fully anonymous.

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

These responsibilities should inform implementation planning rather than be addressed only after data have been transferred or reused.

Measuring whether AI improves radiotherapy planning

A hospital considering AI-assisted OAR contouring should define the improvement it expects before introducing the system.

The objective may be to reduce contouring time, improve consistency, support clinical staff, or improve the reliability of anatomical information used during treatment planning.

Time savings are relevant, but they should reflect the complete contouring process rather than only the time required for automatic generation.

Geometric performance is also important, but it should be interpreted alongside clinical acceptability and potential dosimetric consequences.

The organization should examine whether AI-generated contours require clinically significant corrections and whether these corrections influence treatment optimization or plan evaluation.

Relevant considerations include missing structures, substantial anatomical discrepancies, incorrect laterality, inappropriate structure naming, and clinically important differences in dose-volume metrics.

The evaluation should also examine whether the correct structure versions are used and whether relevant review and approval steps are completed before treatment delivery.

The number of automatically generated contours is not a meaningful measure of success on its own.

A system may produce many structures while failing to identify an OAR that is important for a particular patient. Likewise, a low editing rate may reflect improved model performance or inadequate attention to errors.

Operational evaluation should examine whether AI reduces delays without weakening professional review.

Longer-term clinical outcomes, including radiation-related toxicity and treatment effectiveness, remain important. However, demonstrating improvements in these outcomes requires suitable study designs, adequate follow-up, and careful consideration of other factors influencing treatment.

Routine quality monitoring should not be presented as proof that AI-assisted contouring independently reduces complications or improves survival.

The purpose is to determine whether the technology contributes clinically useful anatomical information while preserving the safeguards necessary for safe treatment.

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

A practical direction for healthcare organizations in Viet Nam

Responsible implementation should begin with a clear understanding of the existing radiotherapy planning pathway.

Before selecting AI software, the organization should examine how planning images are acquired, how OARs are delineated, who reviews the contours, and how final structures are approved.

It should also understand how contour changes affect dose evaluation and how the team confirms that approved structures are used in the treatment plan.

Radiation oncologists, medical physicists, and other relevant professionals should participate in defining the intended use.

The organization should review supporting evidence, technical requirements, local imaging conditions, patient characteristics, and applicable legal obligations.

Local commissioning should determine whether the software reliably generates the required structures and whether the clinical team can recognize and correct significant errors.

The hospital should establish how contours are reviewed, how disagreements are resolved, and how approved structures remain linked to the final treatment plan.

A reliable fallback should remain available when AI is unsuitable or unavailable.

The organization should also recognize when improving conventional contouring protocols, peer review, image registration, or treatment planning procedures may provide greater immediate benefit than introducing new software.

AI should be selected to address an identified clinical or operational problem rather than deployed solely because automated segmentation is technically available.

The objective is not to eliminate manual contouring. It is to reduce unnecessary work while maintaining appropriate anatomical verification, professional accountability, and treatment safety.

From automated contours to safer radiotherapy

AI-assisted auto-contouring is becoming an increasingly important application of artificial intelligence in radiotherapy.

Recent evidence indicates that it can reduce delineation time and improve selected measures of contour consistency.

The 2026 multicenter study in Nature Communications provides encouraging evidence of efficiency gains, while the HaN-Seg dosimetric analysis demonstrates why favorable geometric agreement does not automatically establish clinical acceptability.

International guidance from the RCR, ESTRO, AAPM, and IAEA provides useful principles for clinical commissioning, anatomical verification, professional approval, treatment planning review, and radiation safety.

However, evidence that an algorithm performs well does not independently establish that an entire radiotherapy service has become safer.

For healthcare organizations in Viet Nam, the value of AI-assisted auto-contouring depends on how reliably the technology is integrated into clinical practice.

Automatically generated OAR contours need to be appropriate for the patient, verified according to their clinical significance, and approved through the applicable professional process.

The final treatment plan must be evaluated using the appropriate anatomical structures, with clinical and technical responsibilities remaining clear.

A contour that looks accurate is not necessarily suitable for treatment planning. Equally, approval of an anatomical contour does not automatically mean that the resulting treatment plan is acceptable.

Both anatomical verification and dosimetric evaluation are essential.

The value of AI-assisted radiotherapy auto-contouring should not be judged only by how quickly an algorithm generates structures or how closely they resemble reference contours. It should be judged by whether the technology supports reliable anatomical verification, appropriate clinical approval, and safer treatment planning without avoidable harm.

Professional Education Notice

This article is provided by Digital Medicine Vietnam (DMVN) for professional education, evidence interpretation, and knowledge exchange. It does not constitute individualized radiation oncology, radiotherapy planning, medical physics, radiation safety, medical device, legal, regulatory, or patient-safety advice.

The opening clinical scenario is illustrative and does not describe a documented patient case.

The discussion focuses primarily on AI-assisted delineation of organs at risk for radiotherapy planning, professional contour verification, clinical approval, and relevant treatment safety considerations. It does not establish the safety, effectiveness, regulatory status, or clinical suitability of any particular commercial, open-source, or internally developed AI system.

Different auto-contouring applications have different intended purposes, imaging requirements, supported anatomical structures, patient populations, and clinical outputs. Evidence supporting one application should not automatically be generalized to another.

The cited studies should be interpreted according to their respective designs, populations, software systems, and limitations. Improvements in contouring time, geometric agreement, or workflow efficiency do not independently establish reduced radiation toxicity, improved tumor control, or better survival.

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

The Vietnamese laws, decrees, circulars, and Ministry of Health documents discussed in this article should be interpreted according to their actual scope and applicable implementation dates. They should not be treated as official validation or authorization of any particular radiotherapy AI product.

The organizational approaches discussed, including local commissioning, OAR contour verification, professional approval, dosimetric evaluation, and continuing quality monitoring, are proposed considerations for responsible healthcare AI implementation and quality improvement.

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

References

1. Mackay K, Banfill K, Bernstein D, et al. Royal College of Radiologists Guidance Statements on the Use of Auto-contouring in Radiotherapy. Clinical Oncology. 2026;50:104004. DOI: 10.1016/j.clon.2025.104004.

2. The Royal College of Radiologists. Guidance on Auto-contouring in Radiotherapy. London: The Royal College of Radiologists; 2024.

3. Niu G, Guan Y, Zhang Y, et al. A Prospective Multicenter Trial of Deep Learning Auto-Segmentation for Organs at Risk in Thoracic Radiotherapy. Nature Communications. 2026;17:4633. DOI: 10.1038/s41467-026-70863-9.

4. Podobnik G, Ibragimov B, Peterlin P, Strojan P, Vrtovec T. Dosimetric Assessment of Deep Learning Based Organ-at-Risk Segmentation: Insights From the HaN-Seg Challenge. Radiotherapy and Oncology. 2026;217:111387. DOI: 10.1016/j.radonc.2026.111387.

5. Mason J, Doherty J, Robinson S, de la Bastide M, Miskell J, McLauchlan R. Auditing the Clinical Usage of Deep-Learning Based Organ-at-Risk Auto-Segmentation in Radiotherapy. Physics and Imaging in Radiation Oncology. 2025;33:100716. DOI: 10.1016/j.phro.2025.100716.

6. Hurkmans C, Bibault JE, Brock KK, et al. A Joint ESTRO and AAPM Guideline for Development, Clinical Validation and Reporting of Artificial Intelligence Models in Radiation Therapy. Radiotherapy and Oncology. 2024;197:110345. DOI: 10.1016/j.radonc.2024.110345.

7. Mayo CS, Moran JM, Bosch W, et al. American Association of Physicists in Medicine Task Group 263: Standardizing Nomenclatures in Radiation Oncology. International Journal of Radiation Oncology, Biology, Physics. 2018;100(4):1057–1066. DOI: 10.1016/j.ijrobp.2017.12.013.

8. Brock KK, Mutic S, McNutt TR, Li H, Kessler ML. Use of Image Registration and Fusion Algorithms and Techniques in Radiotherapy: Report of the AAPM Radiation Therapy Committee Task Group No. 132. Medical Physics. 2017;44(7)–e76. DOI: 10.1002/mp.12256.

9. Ford E, Conroy L, Dong L, et al. Strategies for Effective Physics Plan and Chart Review in Radiation Therapy: Report of AAPM Task Group 275. Medical Physics. 2020;47(6)–e272. DOI: 10.1002/mp.14030.

10. Huq MS, Fraass BA, Dunscombe PB, et al. The Report of Task Group 100 of the AAPM: Application of Risk Analysis Methods to Radiation Therapy Quality Management. Medical Physics. 2016;43(7):4209–4262. DOI: 10.1118/1.4947547.

11. International Atomic Energy Agency. Radiation Protection and Safety in Medical Uses of Ionizing Radiation. IAEA Safety Standards Series No. SSG-46. Vienna: International Atomic Energy Agency; 2018.

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

13. Ministry of Health of Viet Nam. Decision No. 1585/QĐ-BYT Promulgating Guidance on Radiotherapy Technical Procedures – Volume 1. Issued June 3, 2026.

14. Ministry of Health of Viet Nam. Decision No. 1586/QĐ-BYT Promulgating Guidance on Radiotherapy Technical Procedures – Volume 2. Issued June 3, 2026.

15. 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 Technical Procedure Guidance. Issued July 1, 2026.

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

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

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

19. 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.

20. 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.

21. 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.

22. 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.

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

24. 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.