Artificial intelligence in radiation treatment planning: a narrative review from automation to clinical decision support

Introduction

Radiotherapy is a cornerstone of modern cancer treatment, with more than half of patients with cancer receiving radiation therapy at some point during their disease course (1). Advances in imaging, treatment delivery, and dose modulation have significantly improved the precision of radiotherapy, enabled higher tumor control while reducing normal tissue toxicity (2-4). Central to these advances is the radiation treatment planning process, which translates clinical intent into a deliverable radiation dose distribution tailored to an individual patient. However, as radiotherapy techniques have evolved toward greater complexity, treatment planning has become increasingly time-consuming, variable, and cognitively demanding (5,6).

Conventional radiation treatment planning involves multiple sequential steps, including image acquisition, target and organ-at-risk (OAR) delineation, dose optimization, and plan evaluation. Each step requires substantial expert input from radiation oncologists, medical physicists, and dosimetrists. Despite standardized guidelines, significant inter- and intra-institutional variability persists, particularly in contouring and plan quality (7,8). Moreover, the growing adoption of advanced techniques such as intensity-modulated radiotherapy (IMRT), volumetric modulated arc therapy (VMAT), stereotactic radiotherapy, and adaptive radiotherapy has further increased the dimensionality of the planning problem (9). These challenges place considerable pressure on clinical workflows and may limit the consistent delivery of high-quality radiotherapy across institutions. Beyond summarizing existing applications, this review aims to synthesize these developments into a clinically meaningful framework that reflects the evolving role of artificial intelligence (AI) in treatment planning.

In recent years, AI, encompassing machine learning and deep learning techniques, has emerged as a promising approach to address many of these challenges (10,11). AI methods are particularly well suited to radiation treatment planning, given the availability of large retrospective datasets, the structured nature of planning tasks, and the need to balance multiple competing objectives (12). Early applications of AI in radiotherapy focused primarily on automation, aiming to reduce manual effort and improve efficiency through tasks such as automated contouring or knowledge-based planning (KBP). These approaches demonstrated that AI could replicate, and in some cases exceed, human-level performance on well-defined tasks, while significantly reducing planning time. Figure 1 summarizes a fully automated treatment-planning workflow (13). The framework begins with the prescription and computed tomography (CT) dataset, after which the system automatically identifies the primary and secondary targets. The volume of interest is then cropped and processed by independently trained deep‑learning models that generate two sets of auto‑contours for each patient. Segmented structures undergo a secondary quality‑check stage, where comparative contour metrics are evaluated using an ensemble random forest classifier. Once the contours are validated, beam parameters and the treatment plan are assigned through a scripted planning module. Together, these steps illustrate an end-to-end AI pipeline capable of generating a complete treatment plan with minimal manual intervention.

Figure 1 Workflow of the automated AI-based treatment-planning framework, showing sequential steps from image input and target localization through auto segmentation, quality checking, and scripted plan generation (13). Reproduced from Jones et al. (13) 2024, licensed under the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0). AI, artificial intelligence; CT, computed tomography; VOI, volume of interest.

As AI capabilities have matured, their role in radiation treatment planning has expanded beyond task-level automation. Contemporary AI systems increasingly function as clinical decision-support tools, assisting clinicians in navigating complex trade-offs between tumor coverage and normal tissue sparing, identifying suboptimal plans, and promoting consistency across planners and institutions (14,15). Rather than replacing human expertise, these systems are designed to augment clinical judgment by providing data-driven insights that would be difficult to derive manually. This shift reflects a broader conceptual transition in radiation oncology, in which AI is viewed not merely as a tool for efficiency, but as an enabler of more informed, transparent, and reproducible decision-making (16).

Despite substantial progress, the clinical translation of AI-driven treatment planning remains uneven. While some AI tools have been integrated into commercial treatment planning systems and routine clinical workflows (17), others remain confined to retrospective studies or single-institution evaluations (18). Key challenges include limited generalizability across patient populations and institutions, susceptibility to dataset bias, lack of interpretability, and concerns regarding safety and regulatory oversight (19). Addressing these issues is essential to ensure that AI technologies improve, rather than compromise, the quality and equity of radiotherapy care. Previous reviews have largely focused on individual AI applications such as auto-segmentation, dose prediction, or automated planning. In contrast, this review integrates these developments into a unified framework centered on clinical decision support.

The objective of this review is not only to provide a comprehensive overview of current AI applications in radiation treatment planning, but also to introduce a conceptual framework that characterizes the evolution of AI from task-based automation to clinically integrated decision support. Unlike prior reviews that primarily catalog individual techniques, we synthesize developments across the planning workflow to highlight how AI is increasingly used to support complex clinical decision-making, including trade-offs between tumor control and normal tissue sparing. This perspective reframes the role of AI from an efficiency-driven tool to a clinically embedded partner in treatment planning. We further discuss implementation challenges, safety considerations, and future directions necessary for responsible translation into practice. This article is presented in accordance with the Narrative Review reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0339/rc).

Methods

This narrative review was conducted to summarize current developments in the application of AI in radiation treatment planning, with a focus on clinical relevance, translational readiness, and implementation challenges.

A literature search was carried out using PubMed, Scopus, and Web of Science to identify relevant English-language publications from approximately 2015 to 2026. Search terms included “artificial intelligence”, “machine learning”, “deep learning”, “radiotherapy”, “treatment planning”, “auto-segmentation”, “dose prediction”, and “knowledge-based planning”.

Relevant original studies and review articles were selected based on their relevance to AI applications in radiotherapy planning. The literature was synthesized qualitatively and organized according to key stages of the treatment planning workflow. This review also incorporates a conceptual framework describing the transition from task-based automation to clinical decision support, with emphasis on comparative analysis and practical implementation considerations. Summary of the search strategy, including databases, search terms, and selection criteria, is shown in Table 1.

Overview of the radiotherapy treatment planning workflow

Radiation treatment planning is a multidisciplinary and iterative process that aims to deliver a prescribed radiation dose to the tumor while minimizing exposure to surrounding normal tissues. The workflow integrates clinical decision-making, medical imaging, physics-based dose calculation, and optimization algorithms, and serves as the primary interface between treatment intent and technical delivery (20). Understanding this workflow is essential for contextualizing current and emerging applications of AI in radiotherapy planning.

Simulation and image acquisition

The treatment planning process begins with patient simulation, during which imaging data are acquired in the treatment position. CT remains the foundation of radiotherapy planning due to its role in electron density estimation for dose calculation. Increasingly, multimodal imaging, including magnetic resonance imaging (MRI) (21) and positron emission tomography (PET) (22), is incorporated to improve soft-tissue contrast and biological target definition. Image quality, patient positioning, and motion management at this stage have a direct impact on downstream planning accuracy and complexity. Variability in imaging protocols and patient anatomy presents an early source of uncertainty that propagates through the planning workflow (23).

Target and OAR delineation

Following image acquisition, target volumes and OARs are delineated. This step is widely recognized as one of the most time-consuming and variable components of treatment planning. Target definition relies on clinical judgment, anatomical interpretation, and consensus guidelines, while OAR delineation often involves numerous structures with complex shapes and variable visibility across imaging modalities. Inter-observer variability in contouring has been well documented and can lead to clinically meaningful differences in dose distribution and treatment outcomes (24). As such, contouring represents a critical bottleneck in planning efficiency and consistency and has been a primary focus of early AI applications. Figure 2 shows the impact of AI-assisted contouring on inter-observer agreement. Figure 2A presents representative CT images in transverse, sagittal, and coronal views as anatomical reference. Figure 2B overlays the two oncologists’ manual contours, illustrating clear inter‑observer variability. In contrast, Figure 2C overlays the AI-assisted adjusted contours, demonstrating markedly improved spatial overlap and reduced variability between observers (25).

Figure 2 Impact of AI-assisted contouring on inter-observer agreement. (A) Representative transverse, sagittal, and coronal CT images used as the anatomical baseline for comparison. (B) Overlay of two independent manual contours drawn by different radiation oncologists, demonstrating substantial inter-observer variability. (C) Overlay of AI-assisted adjusted contours from the same observers, showing markedly improved spatial overlap and reduced variability with the introduction of AI support (25). Reproduced from Arjmandi et al. (25) 2026, licensed under the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0). AI, artificial intelligence; CT, computed tomography.

Dose optimization and treatment planning

Once contours are finalized, dose optimization is performed to generate a treatment plan that satisfies prescribed target coverage and OAR constraints. Modern radiotherapy techniques such as IMRT (26) and VMAT (27) rely on inverse planning, in which desired dose objectives are specified and optimization algorithms iteratively adjust beam parameters to achieve these goals (28). This process involves balancing competing objectives, often requiring multiple optimization cycles and manual adjustments by dosimetrists and physicists. Plan quality is influenced by planner experience, institutional preferences, and available computational tools, contributing to variability across planners and centers.

Plan evaluation and quality assurance (QA)

After optimization, treatment plans are evaluated to ensure clinical acceptability and safety. Evaluation typically includes assessment of dose-volume histograms, spatial dose distributions, and compliance with institutional or protocol-based constraints. This step relies heavily on expert judgment and may involve subjective trade-offs, particularly in complex cases. QA procedures are then performed to verify the integrity of the plan and its deliverability on the treatment machine (29). Errors introduced at any stage of the planning process can compromise treatment quality, underscoring the need for robust and systematic evaluation mechanisms.

Adaptive and online radiotherapy planning

Advances in imaging and treatment delivery have enabled adaptive radiotherapy, in which treatment plans are modified over the course of therapy to account for anatomical or biological changes (30). Adaptive planning introduces additional complexity, as it may require rapid contour updates, re-optimization, and QA under time constraints. Online adaptive radiotherapy demands efficient and reliable planning solutions that can operate within minutes (31). These emerging workflows highlight both the limitations of traditional planning approaches and the potential value of AI-driven tools capable of supporting real-time clinical decision-making.

AI applications in radiation treatment planning

AI has been applied across nearly every stage of the radiotherapy treatment planning workflow, with the overarching goals of improving efficiency, reducing variability, and enhancing plan quality (32). Early efforts focus primarily on automating discrete tasks, while more recent approaches aim to provide integrated support for complex clinical decision-making (33). In this section, we review major AI applications in radiation treatment planning, organized according to their role within the planning workflow.

Auto-segmentation and contouring

Automated segmentation of target volumes and OARs represents the most mature and widely adopted application of AI in radiation treatment planning (34). Early approaches relied on atlas-based methods, which propagated contours from previously annotated patients using deformable image registration (35). While effective in anatomically consistent regions, atlas-based techniques were limited by sensitivity to anatomical variability and often required substantial manual correction.

The advent of deep learning, particularly convolutional neural networks (CNNs), has led to significant improvements in segmentation accuracy and robustness (36). Deep learning-based models have demonstrated strong performance across a range of disease sites, including head and neck (H&N), prostate, lung, and breast cancers (37-39). These models can substantially reduce contouring time while achieving agreement with expert-defined contours that are comparable to inter-observer variability. As a result, AI-assisted contouring has been incorporated into commercial treatment planning systems and is increasingly used in routine clinical practice.

Despite these advances, challenges remain. Model performance can vary across imaging modalities, institutions, and patient populations, and rare or anatomically ambiguous structures remain difficult to segment reliably. Figure 3 illustrates the evolution of deep-learning architectures used for medical image segmentation. Figure 3A shows a basic CNN used for pixel- or voxel-wise classification, highlighting how layered convolutional operations extract hierarchical image features. Figure 3B presents a fully convolutional network (FCN), which introduces an encoder-decoder structure to enable dense, image-wide segmentation without fully connected layers. Figure 3C depicts the U-Net architecture, a specialized FCN with skip connections that link encoder and decoder pathways, allowing for improved localization and segmentation accuracy in medical imaging tasks (40). Moreover, segmentation accuracy metrics do not always capture clinically meaningful errors, reinforcing the need for continued human oversight (41). Nonetheless, auto-segmentation has become a foundational component of AI-enabled treatment planning and serves as a gateway for broader AI integration.

Figure 3 Deep-learning architectures commonly used for medical image segmentation. (A) Basic CNN architecture performing pixel- or voxel-wise classification through stacked convolutional, pooling, and fully connected layers. (B) FCN introducing an encoder-decoder structure to enable dense, end-to-end image segmentation without fully connected layers. (C) U-Net architecture, a specialized FCN incorporating skip connections between encoder and decoder pathways to improve feature localization and segmentation accuracy in medical imaging (40). Reproduced from Kalantar et al. (40) 2021, licensed under the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0). CNN, convolutional neural network; FCN, fully convolutional network.

While AI-based auto-segmentation has demonstrated strong performance for well-defined OARs, such as the bladder, lungs, or parotid glands, progress has been more limited for target volume delineation in anatomically complex regions, particularly in H&N cancers. Target segmentation in these cases remains challenging due to poor soft-tissue contrast, variability in tumor extent, and the need for clinical interpretation that extends beyond purely anatomical boundaries. As a result, advanced deep learning models may produce contours that achieve high quantitative agreement with training data but are clinically suboptimal or inconsistent with physician intent. This example highlights a broader limitation of current AI approaches: tasks that require contextual clinical judgment and interpretation remain difficult to fully automate.

Compared with atlas-based segmentation, deep learning approaches generally demonstrate superior accuracy and robustness, particularly in anatomically variable regions. However, atlas-based methods retain advantages in interpretability and may perform adequately in anatomically consistent structures with limited training data. Deep learning models, while more flexible, are highly dependent on large, high-quality annotated datasets and may exhibit reduced performance in out-of-distribution cases. In clinical practice, hybrid approaches that combine atlas initialization with deep learning refinement have been explored to balance robustness and accuracy.

Knowledge-based planning (KBP) and dose prediction

KBP leverages historical treatment plans to guide dose optimization for new patients (42). Traditional KBP methods rely on handcrafted features and statistical models to estimate achievable dose-volume relationships based on patient anatomy. These approaches have been shown to improve planning consistency and reduce dependence on individual planner experience.

More recently, deep learning-based dose prediction models have emerged as an extension of KBP (43). These models directly predict three-dimensional dose distributions or dose-volume histograms from patient anatomy, often using imaging and contour information as inputs. Deep learning approaches can capture complex, nonlinear relationships between anatomy and dose, enabling rapid estimation of high-quality plans without explicit optimization (44).

Clinically, AI-based dose prediction has demonstrated potential to reduce planning time, improve plan quality consistency, and facilitate plan benchmarking. Predicted dose distributions can serve as optimization objectives, quality references, or decision-support tools for planners. However, their performance is highly dependent on training data quality and institutional practice patterns, raising concerns about generalizability and bias when models are applied across centers (45).

Traditional KBP and deep learning-based dose prediction represent two complementary paradigms. KBP methods are typically more interpretable and easier to integrate into existing clinical workflows, as they rely on established statistical relationships and handcrafted features. In contrast, deep learning models can capture more complex, nonlinear relationships between anatomy and dose, enabling improved predictive performance in some settings. However, this increased flexibility comes at the cost of reduced interpretability and greater sensitivity to training data bias. As a result, KBP remains widely used in clinical practice, while deep learning approaches are emerging as more powerful but less transparent alternatives. For example, deep learning-based dose prediction has shown strong performance in relatively standardized treatment sites, such as prostate cancer, where anatomical variability is limited and planning objectives are well defined. However, its performance is less consistent in more complex scenarios, such as H&N or adaptive radiotherapy, where competing clinical objectives and anatomical variability are greater. In these settings, predicted dose distributions may fail to capture clinically nuanced trade-offs, reinforcing the need for human interpretation and adjustment.

Automated and end-to-end treatment plan generation

Building on dose prediction and KBP, several groups have explored automated or end-to-end treatment planning systems (46-48). These approaches aim to generate clinically acceptable treatment plans with minimal or no manual intervention by integrating segmentation, optimization, and evaluation steps. Such systems have shown promise in standardized treatment sites, where planning objectives are well defined and anatomical variability is limited.

Automated planning has been associated with reductions in planning time and improvements in plan consistency, particularly in high-volume clinical settings (49). Some commercial systems now offer automated planning modules that incorporate predefined clinical protocols and AI-driven optimization strategies (50). However, fully automated planning remains challenging in complex or atypical cases, where nuanced clinical judgment is required.

Importantly, end-to-end automation raises questions about transparency, error detection, and accountability. While automated plans may meet conventional dosimetric criteria, they may fail to capture patient-specific clinical considerations that are not encoded in optimization objectives (51). Consequently, current clinical implementations typically position automated planning as an assistive tool rather than a replacement for expert review.

Fully automated planning systems offer clear advantages in efficiency and consistency, particularly for standardized treatment sites. However, compared with human-in-the-loop (HITL) approaches, they may lack the flexibility to account for patient-specific clinical nuances and evolving treatment priorities. Human-guided or semi-automated planning frameworks allow clinicians to iteratively refine AI-generated outputs, potentially achieving higher clinical acceptability at the cost of increased interaction time. The optimal approach may therefore depend on the clinical context, with automation favored in high-volume, standardized scenarios and HITL strategies preferred for complex or atypical cases.

In practice, automated planning systems perform best in standardized, high-volume treatment sites but may struggle in atypical or complex cases. For instance, while automated prostate planning workflows have demonstrated high efficiency and consistency, similar approaches in H&N planning often require substantial manual refinement due to the complexity of anatomical structures and competing dose constraints. This illustrates that the benefits of automation are highly context-dependent.

AI for plan evaluation and QA

Beyond plan generation, AI has been increasingly applied to plan evaluation and QA (52). Machine learning models have been developed to automatically assess plan quality, identify deviations from institutional standards, and detect potential errors. These tools can flag suboptimal plans early in the workflow, reducing the risk of downstream corrections or treatment delays.

AI-based QA systems have also been explored for treatment delivery verification, including detection of anomalous machine parameters and inconsistencies between planned and delivered dose (53). In the context of increasingly complex and time-sensitive workflows, such as adaptive radiotherapy, these tools offer the potential to enhance safety and efficiency.

As with other AI applications, integration into clinical practice requires careful validation and clear definition of roles and responsibilities. AI-driven QA tools are most effective when used to augment, rather than replace, existing safety processes, providing an additional layer of oversight that complements human expertise.

Compared with conventional rule-based or physics-based QA methods, AI-driven QA systems offer the ability to detect complex patterns and subtle deviations that may not be captured by predefined criteria. However, traditional QA approaches remain more transparent and are grounded in well-established physical principles, making them more readily interpretable and trusted in safety-critical settings. AI-based QA is therefore best positioned as a complementary layer that augments, rather than replaces, existing QA processes. A comparison of major AI approaches across the treatment planning workflow is summarized in Table 2.

From automation to clinical decision support

This section represents the central conceptual contribution of this review. We propose that the role of AI in radiation treatment planning is undergoing a fundamental transition from task-based automation toward clinically integrated decision support. While early applications of AI in radiation treatment planning were primarily motivated by efficiency gains through task automation, the role of AI in contemporary radiotherapy has evolved substantially. As treatment planning has become more complex and data-rich, there is growing recognition that the greatest value of AI lies not in replacing human expertise, but in supporting clinical decision-making (54). This shift reflects a broader transition from automation-oriented tools toward AI systems designed to assist clinicians in navigating complex trade-offs, improving consistency, and enhancing transparency in treatment planning.

Limitations of automation-centric approaches

Automation-focused AI tools have demonstrated clear benefits, including reduced planning time, decreased inter-planner variability, and improved standardization. However, automation alone is insufficient to address the full spectrum of clinical challenges encountered in radiation treatment planning. Fully automated systems typically operate within predefined objectives and constraints, which may not adequately capture patient-specific considerations, evolving clinical priorities, or nuanced trade-offs between competing dosimetric goals (55).

Moreover, automated outputs can obscure underlying assumptions and failure modes, potentially creating overreliance on algorithmic recommendations (56). In complex or atypical cases, automated plans that satisfy standard dosimetric criteria may still be clinically suboptimal. These limitations highlight the need for AI systems that enhance, rather than constrain, human decision-making and allow clinicians to interrogate, contextualize, and adapt algorithmic outputs. The distinction between automation-oriented and decision-support AI is summarized in Table 3.

AI as a tool for navigating clinical trade-offs

Radiation treatment planning inherently involves balancing competing objectives, such as maximizing tumor coverage while minimizing toxicity to surrounding normal tissues. These trade-offs are often patient-specific and influenced by clinical factors that extend beyond numerical dose constraints. AI-driven decision-support systems can assist clinicians by exploring a broader solution space than is feasible manually and by presenting alternative planning scenarios that illuminate trade-offs between competing goals (57). For example, in complex H&N cases, AI-generated dose predictions may suggest achievable trade-offs, but clinicians must still interpret these recommendations in the context of patient-specific priorities and uncertainties.

For example, AI-based dose prediction and plan evaluation tools can provide reference benchmarks that help clinicians assess whether a plan is near optimal or if further improvement is achievable. By contextualizing a given plan within a distribution of prior outcomes, AI systems can support more informed decision-making without prescribing a single “optimal” solution (58). This paradigm shifts the role of AI from plan generator to analytical partner in the planning process.

Human-AI interaction in treatment planning

Effective integration of AI into radiotherapy planning depends critically on human-AI interaction (59,60). Rather than functioning as black-box systems, AI tools must be designed to facilitate clinician engagement, interpretation, and oversight. Interfaces that allow users to visualize predicted dose distributions, understand sources of uncertainty, and compare alternative plans are essential for building trust and ensuring safe adoption.

Clinicians retain responsibility for treatment decisions, and AI systems should be positioned to support, not supplant, clinical judgment. This supervisory role is particularly important in cases where AI outputs conflict with clinical intuition or patient-specific considerations. Transparent AI systems that provide interpretable outputs and highlight potential limitations are more likely to be integrated effectively into routine practice (61). Figure 4 illustrates a human-centric intelligent treatment planning (HCITP) workflow that integrates automated planning with human oversight (60). The process begins with the physician’s prescription, after which the Execution Module autonomously operates the treatment planning system (TPS) to generate a plan. A human evaluator then reviews the proposed plan and determines whether it meets clinical requirements. If revisions are needed, feedback is provided through the Conversation Module, which uses a large‑language-model conversational interface to communicate with the system. This feedback informs the Evaluation Module, which incorporates clinical notes, guidelines, and physician preferences to guide the next planning iteration. Together, these modules form an interactive loop that combines AI-driven automation with human judgment to ensure plan quality and safety.

Figure 4 Human-centric intelligent treatment planning workflow. The system autonomously generates a treatment plan through the Execution Module, after which a human evaluator reviews the plan. If revisions are needed, feedback is exchanged through the Conversation Module and incorporated by the Evaluation Module to guide the next planning iteration (60). Reproduced from Jafar et al. (60) 2026, licensed under the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0). TPS, treatment planning system.

In this context, human oversight extends beyond passive review to active validation, where clinicians critically assess AI-generated outputs for clinical plausibility and safety. The integration of formal HITL protocols ensures that AI systems augment, rather than bypass, expert judgment.

Consistency, equity, and standardization of care

One of the most promising aspects of AI-enabled decision support is its potential to reduce unwarranted variability in treatment planning and promote more equitable care (62). Differences in planner experience, institutional resources, and local practice patterns can lead to significant disparities in plan quality. AI-based decision-support tools can help establish consistent benchmarks for plan evaluation and facilitate knowledge transfer across institutions (63).

However, this potential benefit must be balanced against the risk of perpetuating existing biases embedded in historical training data (64). Decision-support systems trained on data from a limited set of institutions or patient populations may inadvertently encode local practice patterns or inequities. Ensuring diversity and representativeness in training datasets is therefore essential for realizing the equity-promoting potential of AI in radiotherapy planning (65). Importantly, equity considerations extend beyond reducing variability between institutions in high-resource settings and must also account for disparities in infrastructure and access to technology across global healthcare systems, as discussed further in “Infrastructure constraints and global equity” section.

Toward trustworthy AI-enabled decision support

For AI-driven decision-support systems to be clinically useful, they must be trustworthy, robust, and aligned with clinical values (66). Trustworthiness encompasses not only technical performance, but also transparency, reliability, and accountability. Clinicians must be able to understand when and why an AI system may fail, and safeguards must be in place to detect and mitigate such failures.

Regulatory and ethical considerations further underscore the importance of framing AI as a decision-support tool rather than an autonomous decision-maker (67). Clear delineation of responsibility, rigorous validation, and ongoing performance monitoring are essential components of safe clinical deployment. By emphasizing human-centered design and continuous evaluation, AI can be integrated into treatment planning workflows in a manner that enhances both efficiency and clinical confidence (68).

Clinical evidence and translational readiness

The clinical impact of AI in radiation treatment planning depends not only on technical performance, but also on evidence demonstrating safety, effectiveness, and generalizability in real-world settings. While a large body of AI research in radiotherapy remains retrospective or methodological in nature, an increasing number of studies have evaluated AI-based planning tools in clinically relevant contexts. In this section, we review the current state of clinical evidence supporting AI-enabled treatment planning and assess its readiness for broader clinical translation.

Retrospective validation studies

Most AI applications in radiation treatment planning have been evaluated using retrospective datasets, often drawn from single institutions. These studies have demonstrated that AI-assisted contouring, dose prediction, and automated planning can achieve performance comparable to expert planners while substantially reducing planning time (69,70). In contouring tasks, deep learning models frequently achieve agreement metrics within the range of inter-observer variability, supporting their use as time-saving aids rather than definitive ground truth generators (71).

Similarly, retrospective evaluations of KBP and deep learning-based dose prediction have shown improved plan consistency and reduced dependence on planner experience (72,73). These studies have been particularly informative in highlighting the potential of AI to standardize planning quality across cases with similar anatomical characteristics. However, retrospective designs limit the ability to assess clinical impact, as they do not capture downstream effects on workflow, decision-making, or patient outcomes.

Prospective and clinical implementation studies

Fewer studies have examined AI tools in prospective or routine clinical settings, but this body of evidence is steadily growing (74,75). Prospective evaluations of AI-assisted contouring have demonstrated meaningful reductions in contouring time, with clinicians maintaining final approval and oversight. In these settings, AI tools are typically used as initial contour generators, allowing clinicians to focus on review and refinement rather than manual delineation (76).

Clinical implementation studies of automated or semi-automated planning systems have reported improvements in planning efficiency and consistency, particularly in high-volume treatment sites such as prostate (77) and H&N cancers (78). Importantly, these studies emphasize the role of institutional adaptation, protocol customization, and user training in achieving successful deployment. AI tools that align closely with existing clinical workflows are more likely to be adopted and sustained over time.

Impact on workflow and resource utilization

One of the most consistently reported benefits of AI-assisted treatment planning is improved workflow efficiency (79). Reductions in contouring and planning time can alleviate workload pressures on clinical staff and enable faster treatment initiation, which may be particularly relevant in resource-constrained settings. By reducing reliance on individual planner expertise, AI tools may also facilitate task redistribution and support training of less experienced staff.

However, efficiency gains must be balanced against the time required for validation, QA, and ongoing monitoring of AI systems (80). In some cases, initial integration of AI tools may increase workload due to the need for extensive testing and user training. Long-term efficiency benefits are most likely to be realized when AI systems are robust, well-integrated, and supported by institutional governance structures (81).

A critical but often underemphasized aspect of AI-assisted workflows is the need for independent human validation of model outputs. While many studies report strong quantitative performance using metrics such as Dice similarity coefficients or dose-volume histogram comparisons, these metrics may fail to capture clinically unacceptable or non-physical errors. As a result, formal HITL workflows, in which clinicians independently review, validate, and, if necessary, correct AI-generated outputs, are essential for safe clinical deployment. Importantly, the clinical utility of AI must be evaluated in the context of a verification trade-off: if an AI tool reduces generation time but requires substantially longer manual correction or validation, the net benefit may be negative. This highlights the need for workflow-aware evaluation metrics that incorporate both efficiency and validation burden.

Structured evaluation frameworks, such as failure mode and effects analysis (FMEA), have been proposed to systematically assess risks associated with AI integration into clinical workflows. These approaches can help identify potential failure points, quantify risk severity, and guide the design of robust validation and QA processes for AI-assisted treatment planning.

Generalizability and multi-institutional performance

Generalizability remains a key challenge for the clinical translation of AI in radiation treatment planning (82). Models trained on data from a single institution may perform suboptimally when applied to external datasets due to differences in imaging protocols, contouring conventions, and treatment techniques. Multi-institutional studies and external validations are therefore essential for establishing the reliability of AI tools across diverse clinical environments (83,84).

Recent efforts to evaluate AI models across institutions have highlighted both the potential and limitations of current approaches. While some models demonstrate robust performance across sites, others require retraining or fine-tuning to maintain accuracy. These findings underscore the importance of transparent reporting, standardized evaluation frameworks, and collaborative data-sharing initiatives to support safe and effective translation (85).

Current state of clinical readiness

Taken together, available evidence suggests that certain AI applications in radiation treatment planning, particularly auto-segmentation and KBP, are approaching or have reached clinical maturity. These tools are increasingly incorporated into commercial treatment planning systems and routine workflows (86). In contrast, more advanced applications, such as fully automated end-to-end planning and adaptive decision-support systems, remain at earlier stages of translation and require further validation (87).

Clinical readiness should be viewed as a continuum rather than a binary threshold. Factors such as task complexity, clinical risk, and availability of human oversight influence the appropriate level of automation and decision support. Continued prospective evaluation, post-deployment monitoring, and iterative refinement will be essential to ensure that AI technologies deliver meaningful and equitable benefits in clinical practice (88). Table 4 summarizes the current clinical evidence supporting AI applications in radiation treatment planning, categorized by study design and translational maturity. While most published studies remain retrospective and single institution in nature, emerging prospective and implementation studies demonstrate meaningful improvements in efficiency and consistency with maintained clinician oversight. Overall, the evidence suggests that auto-segmentation and KBP are approaching clinical maturity, whereas fully automated and adaptive decision-support systems require further validation before widespread adoption.

Challenges and limitations

Despite rapid advances and growing clinical adoption, the integration of AI into radiation treatment planning faces several important challenges. These limitations span technical, clinical, ethical, and organizational domains and must be addressed to ensure that AI technologies improve the safety, quality, and equity of radiotherapy care rather than introduce new risks or disparities.

Data quality, annotation variability, and bias

Beyond recognizing the presence of bias, it is critical to consider practical strategies for mitigating its impact, which may differ depending on the specific AI application within the treatment planning workflow. AI models for radiation treatment planning rely heavily on large, high-quality datasets for training and validation. However, clinical radiotherapy data are inherently heterogeneous, reflecting variations in imaging protocols, contouring practices, treatment techniques, and institutional preferences (89). Contouring datasets often encode inter-observer variability and local practice patterns, which may be inadvertently learned and propagated by AI models.

Bias in training data represents a critical concern (90). Datasets that underrepresent certain patient populations, disease sites, or anatomical variations may result in models that perform poorly or unpredictably when applied more broadly. These biases can exacerbate existing inequities in care if AI tools preferentially benefit well-represented populations. Addressing these challenges requires deliberate efforts to curate diverse and representative datasets and to evaluate model performance across clinically relevant subgroups (91).

Several general strategies have been proposed to address bias in AI models for radiation treatment planning. First, the use of multi-institutional and demographically diverse datasets can improve representativeness and reduce the risk of models encoding local practice patterns. Second, standardized contouring guidelines and consensus-driven annotations can help reduce inter-observer variability that may otherwise be learned by models. Third, external validation across independent datasets is essential to assess model robustness under domain shift. In addition, emerging approaches such as federated learning enable collaborative model development without centralized data sharing, thereby increasing data diversity while preserving patient privacy. Finally, subgroup performance analysis (e.g., across anatomical sites, disease stages, or patient demographics) is necessary to identify and quantify potential biases before clinical deployment.

Importantly, the impact and mitigation of bias may vary across different categories of AI applications in treatment planning. In auto-segmentation, bias is often driven by annotation variability and differences in contouring guidelines; strategies such as multi-observer training datasets, consensus contours, and uncertainty-aware models can help address this issue. In KBP and dose prediction, bias may reflect institutional planning preferences embedded in historical datasets; techniques such as cross-institutional model training, domain adaptation, and normalization of planning objectives can reduce these effects. For automated planning systems, bias may arise from predefined optimization templates that do not generalize across patient populations; incorporating HITL feedback and adaptive learning mechanisms can help mitigate this limitation. In AI-based plan evaluation, bias may manifest as systematic preference for specific planning styles; therefore, benchmarking against multi-institutional standards and incorporating clinically interpretable evaluation metrics are important for ensuring fairness and generalizability.

Generalizability and domain shift

Generalizability remains one of the most significant barriers to widespread clinical deployment of AI in radiation treatment planning (92). Models trained under specific institutional conditions may experience performance degradation when applied to new settings, a phenomenon often referred to as domain shift. Differences in scanner hardware, imaging protocols, contouring guidelines, and planning techniques can all contribute to this effect.

While strategies such as transfer learning and multi-institutional training have shown promise, they introduce additional complexity and resource requirements (93). From a clinical perspective, uncertainty regarding model performance in local settings can hinder adoption. Robust external validation, transparent reporting of training data characteristics, and standardized benchmarking frameworks are essential for building confidence in AI tools across institutions.

Safety, robustness, and failure detection

Radiation treatment planning is a safety-critical process, and errors can have serious clinical consequences. AI systems, like all complex software, are susceptible to failure modes that may not be immediately apparent to users. These failures may arise from atypical patient anatomy, poor image quality, or scenarios that are insufficiently represented in training data.

Ensuring robustness requires not only high average performance, but also mechanisms for identifying when AI outputs may be unreliable. Confidence estimation, uncertainty quantification, and automated failure detection are active areas of research but are not yet widely implemented in clinical systems. Until such safeguards are routinely available, human oversight remains indispensable for mitigating risk (94).

Interpretability, transparency, and trust

The opacity of many AI models, particularly deep learning-based systems, poses challenges for clinical acceptance. Clinicians may be reluctant to rely on recommendations that cannot be readily explained or interrogated, especially when AI outputs conflict with clinical intuition (95). Lack of interpretability can also complicate error investigation and accountability.

Improving transparency does not necessarily require full interpretability of model internals, but rather the ability to provide clinically meaningful explanations and contextual information (96). Visualization of predicted dose distributions, comparison with reference plans, and clear communication of model limitations can all contribute to more informed and confident use of AI tools.

To address the “black box” nature of many AI models, a range of technical approaches for improving interpretability and trustworthiness have been proposed. Visualization-based methods, such as saliency maps, highlight regions of input images that contribute most strongly to model predictions, providing clinicians with intuitive insight into model behavior. More advanced techniques, including layer-wise relevance propagation (LRP), attempt to decompose model outputs and attribute contributions across network layers, offering a more detailed explanation of decision pathways. In addition, uncertainty quantification methods, such as Bayesian neural networks or Monte Carlo dropout, enable estimation of prediction confidence, allowing identification of cases where model outputs may be less reliable. These approaches can support clinical decision-making by not only providing predictions but also conveying the degree of confidence and rationale behind them. Although many of these methods remain under active development, their integration into clinical workflows represents an important step toward building trust and facilitating safe adoption of AI in radiation treatment planning.

Regulatory, legal, and workflow considerations

The deployment of AI in radiation treatment planning raises important regulatory and legal questions. AI tools used for clinical decision support may be subject to regulatory oversight, requiring evidence of safety, effectiveness, and ongoing performance monitoring. The dynamic nature of AI systems, particularly those that evolve over time, poses challenges for traditional regulatory frameworks (97).

From a practical standpoint, successful integration of AI into clinical workflows requires alignment with existing processes, roles, and responsibilities (98). Poorly integrated tools may disrupt workflows or create additional burdens rather than alleviating them. Institutional governance structures, clear delineation of accountability, and clinician engagement are critical for ensuring that AI adoption enhances, rather than undermines, clinical practice.

In practical terms, integration of AI into clinical workflows is largely mediated through TPS, which serve as the central platform for plan generation and evaluation. Most modern TPS platforms, such as Eclipse (Varian), RayStation (RaySearch Laboratories), and Monaco (Elekta), provide application programming interfaces (APIs) and user-scripting environments that enable customization and integration of external tools. For example, Eclipse supports the Eclipse Scripting API (ESAPI), which allows developers to access patient data, automate planning steps, and interface with third-party software. Similarly, RayStation offers a Python-based scripting environment that facilitates the development and deployment of AI-driven workflows, including automated contouring, plan generation, and evaluation. These scripting capabilities have become the primary mechanism for integrating commercial and research AI tools, such as third-party solutions (e.g., Radformation), into clinical practice. However, despite the availability of these technical frameworks, challenges remain in ensuring interoperability, validation, version control, and regulatory compliance, which contribute to the persistent “implementation gap” between research prototypes and routine clinical deployment.

Infrastructure constraints and global equity

While AI in radiation treatment planning has the potential to improve consistency and quality of care, its development and evaluation have largely been centered in high-resource settings. Many existing AI models assume access to high-performance computing infrastructure, including graphics processing units (GPUs), large-scale data storage, and reliable high-speed internet connectivity. However, such resources are not universally available, particularly in low- and middle-income countries (LMIC), where radiotherapy capacity is already limited and infrastructure constraints are significant.

These disparities raise important concerns regarding the equitable translation of AI technologies. Models developed and validated in high-income settings may be impractical or inaccessible in resource-constrained environments, potentially widening existing gaps in cancer care. For example, deployment of computationally intensive deep learning models may be infeasible in clinics that rely on central processing unit (CPU)-based systems or have limited network connectivity for cloud-based solutions.

Addressing these challenges requires the development of lightweight, hardware-aware AI models that can operate efficiently on limited computational resources. Approaches such as model compression, pruning, quantization, and the design of efficient network architectures can reduce computational demands while maintaining acceptable performance. In addition, CPU-optimized models and offline-capable systems may enable broader deployment in settings without reliable internet access.

Beyond technical considerations, equitable implementation also requires attention to data representation, training datasets that include diverse patient populations, and collaboration with institutions in LMIC to ensure that AI tools are adapted to local clinical workflows and needs.

Ultimately, achieving responsible translation of AI in radiation treatment planning requires moving beyond a one-size-fits-all model and ensuring that technological advances are accessible, adaptable, and beneficial across diverse global healthcare settings.

Future directions

The continued evolution of AI in radiation treatment planning will be shaped by advances in technology, changes in clinical practice, and growing emphasis on safe and equitable implementation. While current applications have focused largely on automation and efficiency, future developments are expected to further expand the role of AI as a clinically integrated decision-support system across the radiotherapy care continuum.

Adaptive and online radiotherapy planning

One of the most promising frontiers for AI in radiation oncology is adaptive radiotherapy. Anatomical and biological changes during treatment can significantly affect dose distribution, yet conventional replanning workflows are often too resource-intensive to implement routinely. AI-driven contouring, dose prediction, and plan evaluation tools have the potential to enable rapid replanning, supporting both offline and online adaptive radiotherapy (99).

In online adaptive workflows, where planning decisions must be made within minutes, AI systems capable of reliable, real-time support are particularly valuable (100). Achieving this vision will require not only technical advances, but also rigorous validation and robust safety mechanisms to ensure that time constraints do not compromise treatment quality.

Integration of multi-modal and longitudinal data

Future AI systems are likely to incorporate a broader range of data sources beyond anatomical imaging. Functional imaging, radiomics, genomics, and clinical outcome data offer opportunities to personalize treatment planning based on tumor biology and patient-specific risk profiles (101). By integrating multi-modal and longitudinal data, AI-enabled decision-support systems could move beyond dosimetric optimization toward biologically informed planning (102). Such integration also raises challenges related to data standardization, interpretability, and clinical utility. Ensuring that increasingly complex models remain actionable and understandable to clinicians will be essential for successful translation (103).

Federated and privacy-preserving learning

Concerns regarding data privacy and institutional data sharing have limited the scale and diversity of training datasets for AI in radiotherapy (104). Federated learning and other privacy-preserving approaches offer a potential solution by enabling collaborative model development without centralized data pooling (105). These approaches may improve model generalizability while respecting patient privacy and regulatory constraints.

Although promising, federated learning introduces new technical and organizational challenges, including communication efficiency, data heterogeneity, and governance structures. Addressing these issues will be critical for realizing the benefits of large-scale, collaborative AI development in radiation oncology (106).

Toward learning health systems in radiotherapy

AI has the potential to support learning health systems in radiotherapy, in which treatment planning practices evolve continuously based on accumulated data and outcomes. By linking planning decisions to clinical results, AI systems could facilitate feedback loops that inform guideline refinement, protocol optimization, and personalized care (107).

Implementing such systems will require careful consideration of data quality, causal inference, and ethical oversight. Importantly, learning health systems must be designed to benefit all patients equitably, avoiding reinforcement of existing disparities (108).

Human-centered and trustworthy AI design

As AI becomes more deeply integrated into treatment planning workflows, human-centered design principles will be increasingly important. AI systems should be developed with input from clinicians, patients, and other stakeholders to ensure that they address real clinical needs and align with professional values. Emphasis on transparency, interpretability, and user control will be essential for building trust and promoting responsible adoption (109).

Ultimately, the success of AI in radiation treatment planning will depend not only on technical innovation, but also on thoughtful integration into clinical practice. By prioritizing safety, equity, and collaboration between humans and machines, AI has the potential to enhance decision-making and improve the quality of radiotherapy care (110). Table 5 outlines key future directions for AI in radiation treatment planning, highlighting emerging technological domains, anticipated clinical impact, and associated challenges. These directions extend beyond efficiency-focused automation toward adaptive workflows, biologically informed personalization, collaborative model development, and human-centered system design. Together, they underscore that future progress will depend not only on technical innovation, but also on safe, equitable, and clinically integrated implementation.

Conclusions

AI has rapidly become an integral component of radiation treatment planning, with applications spanning contouring, dose optimization, plan evaluation, and QA. Early successes in task-level automation have demonstrated clear benefits in efficiency and consistency, supporting the growing clinical adoption of AI-assisted planning tools. However, as radiotherapy continues to evolve toward greater complexity and personalization, the limitations of automation-centric approaches have become increasingly apparent.

This review highlights a conceptual shift in the role of AI in radiation treatment planning, from tools designed primarily to reduce manual workload toward systems that support clinical decision-making. By assisting clinicians in navigating complex trade-offs, benchmarking plan quality, and promoting consistency across patients and institutions, AI has the potential to enhance both the effectiveness and equity of radiotherapy care. Importantly, these benefits are most likely to be realized when AI systems are implemented as decision-support tools that augment, rather than replace, human expertise.

Despite substantial progress, important challenges remain. Issues related to data quality, generalizability, interpretability, and safety continue to limit the widespread and equitable deployment of AI in clinical practice. Addressing these challenges will require rigorous validation, transparent reporting, and ongoing performance monitoring, as well as close collaboration among clinicians, researchers, industry partners, and regulators. Human-centered design and explicit consideration of ethical and equity implications are essential to ensure that AI technologies align with clinical values and patient needs. Ensuring equitable access to AI technologies across diverse healthcare settings, including resource-constrained environments, will be critical to realizing the full global impact of AI in radiotherapy.

Looking forward, advances in adaptive radiotherapy, multi-modal data integration, and privacy-preserving learning are expected to further expand the role of AI in treatment planning. As these technologies mature, the focus must remain on building trustworthy systems that enhance clinical insight and support informed decision-making. By embracing this paradigm, AI can serve as a powerful enabler of high-quality, consistent, and patient-centered radiotherapy in the era of precision oncology.

The primary contribution of this review is the identification of a paradigm shift from automation to clinical decision support, which reframes how AI systems should be developed, evaluated, and integrated into radiation treatment planning workflows.

Acknowledgments

None.

Footnote

Reporting Checklist: The author has completed the Narrative Review reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0339/rc

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0339/prf

Funding: This work was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant (No. RGPIN-2025-06991).

Conflicts of Interest: The author has completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0339/coif). The author reports the support from the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant (No. RGPIN-2025-06991). The author has no other conflicts of interest to declare.

Ethical Statement: The author is accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Delaney G, Jacob S, Featherstone C, et al. The role of radiotherapy in cancer treatment: estimating optimal utilization from a review of evidence-based clinical guidelines. Cancer 2005;104:1129-37. [Crossref] [PubMed]
2. Parchur AK, Yadav P. Editorial: Innovative approaches in precision radiation oncology. Front Med (Lausanne) 2026;13:1776567. [Crossref] [PubMed]
3. Chow JCL. Hybrid Artificial Intelligence–Monte Carlo Methods for Radiotherapy Dose Calculation. Encyclopedia 2026;6:55.
4. Bhide SA, Nutting CM. Recent advances in radiotherapy. BMC Med 2010;8:25. [Crossref] [PubMed]
5. Van Dyk J. Quality Assurance of Radiation Therapy Planning Systems: Current Status and Remaining Challenges. International Journal of Radiation Oncology*Biology*Physics 2008;71:S23-7.
6. Gooding MJ, Aluwini S, Guerrero Urbano T, et al. Fully automated radiotherapy treatment planning: A scan to plan challenge. Radiother Oncol 2024;200:110513. [Crossref] [PubMed]
7. de Jong EEC, Guckenberger M, Andratschke N, et al. Variation in current prescription practice of stereotactic body radiotherapy for peripherally located early stage non-small cell lung cancer: Recommendations for prescribing and recording according to the ACROP guideline and ICRU report 91. Radiother Oncol 2020;142:217-23. [Crossref] [PubMed]
8. Toussaint A, Richter A, Mantel F, et al. Variability in spine radiosurgery treatment planning - results of an international multi-institutional study. Radiat Oncol 2016;11:57. [Crossref] [PubMed]
9. Webster M, Podgorsak A, Li F, et al. New Approaches in Radiotherapy. Cancers (Basel) 2025;17:1980. [Crossref] [PubMed]
10. Siddique S, Chow JCL. Artificial intelligence in radiotherapy. Rep Pract Oncol Radiother 2020;25:656-66. [Crossref] [PubMed]
11. Chow JCL, Wang T. Editorial: Machine learning in radiation therapy for lung cancer. Front Oncol 2024;14:1444543. [Crossref] [PubMed]
12. Netherton TJ, Cardenas CE, Rhee DJ, et al. The Emergence of Artificial Intelligence within Radiation Oncology Treatment Planning. Oncology 2021;99:124-34. [Crossref] [PubMed]
13. Jones S, Thompson K, Porter B, et al. Automation and artificial intelligence in radiation therapy treatment planning. J Med Radiat Sci 2024;71:290-8. [Crossref] [PubMed]
14. Chow JC. Artificial intelligence in radiotherapy and patient care. In: Artificial Intelligence in Medicine. Cham: Springer International Publishing; 2021:1-13.
15. Chow JCL. Quantum Computing and Machine Learning in Medical Decision-Making: A Comprehensive Review. Algorithms 2025;18:156.
16. Landry G, Kurz C, Traverso A. The role of artificial intelligence in radiotherapy clinical practice. BJR Open 2023;5:20230030. [Crossref] [PubMed]
17. Sheng Y, Zhang J, Ge Y, et al. Artificial intelligence applications in intensity modulated radiation treatment planning: an overview. Quant Imaging Med Surg 2021;11:4859-80. [Crossref] [PubMed]
18. Ng F, Jiang R, Chow JCL. Predicting radiation treatment planning evaluation parameter using artificial intelligence and machine learning. IOP SciNotes 2020;1:014003.
19. Kiser KJ, Fuller CD, Reed VK. Artificial intelligence in radiation oncology treatment planning: a brief overview. J Med Artif Intell 2019;2:9.
20. Meyer P, Biston MC, Khamphan C, et al. Automation in radiotherapy treatment planning: Examples of use in clinical practice and future trends for a complete automated workflow. Cancer Radiother 2021;25:617-22. [Crossref] [PubMed]
21. Devic S. MRI simulation for radiotherapy treatment planning. Med Phys 2012;39:6701-11. [Crossref] [PubMed]
22. Zaidi H, Vees H, Wissmeyer M. Molecular PET/CT imaging-guided radiation therapy treatment planning. Acad Radiol 2009;16:1108-33. [Crossref] [PubMed]
23. Baier F, Koelbl O, Steger F, et al. Interobserver Variation Within Planning Target Volume and Organs at Risk in a Patient with Oropharyngeal Carcinoma: A Contouring Study with Anatomical Analysis. Curr Oncol 2026;33:39. [Crossref] [PubMed]
24. Chamunyonga C, Mengersen K, Hargrave C. Inter-observer variation in radiation oncology contouring: A scoping review of contour comparison methods, and reporting of impact on the organ at risk plan dose variations. Radiography (Lond) 2026;32:103295. [Crossref] [PubMed]
25. Arjmandi N, Sebzari AR, Molaei F, et al. Clinical validation of AI‐assisted contouring in prostate radiation therapy treatment planning: Highlighting automation bias and the need for standardized quality assurance. Journal of Applied Clinical Medical Physics 2026;27:e70425.
26. Guo J, Qin S, Guo C, et al. Desargues cloud TPS: a cloud-based automatic radiation treatment planning system for IMRT. Biomed Eng Online 2026;25:22. [Crossref] [PubMed]
27. Men C, Romeijn HE, Jia X, et al. Ultrafast treatment plan optimization for volumetric modulated arc therapy (VMAT). Med Phys 2010;37:5787-91. [Crossref] [PubMed]
28. Inoue T, Iijima K, Takatsu J, et al. Impact of the locking function of hybrid inverse planning optimization on the treatment plan quality of intracavitary/interstitial brachytherapy for locally advanced cervical cancer. Radiol Phys Technol 2026;19:284-92. [Crossref] [PubMed]
29. García-Romero A, Baeza-Trujillo M, Teijeiro-García A, et al. Guide for quality control and safety of external radiotherapy treatment planning systems and treatment plans. Phys Med 2026;141:105699. [Crossref] [PubMed]
30. Chow JCL. Machine learning in cancer imaging for enhanced precision in diagnosis and therapy. Discov Comput 2026;29:186. [Crossref] [PubMed]
31. Lim-Reinders S, Keller BM, Al-Ward S, et al. Online Adaptive Radiation Therapy. International Journal of Radiation Oncology*Biology*Physics 2017;99:994-1003.
32. Chow JCL. Internet-based computer technology on radiotherapy. Rep Pract Oncol Radiother 2017;22:455-62. [Crossref] [PubMed]
33. Sun W, Niraula D, El Naqa I, et al. Precision radiotherapy via information integration of expert human knowledge and AI recommendation to optimize clinical decision making. Comput Methods Programs Biomed 2022;221:106927. [Crossref] [PubMed]
34. Jin D, Guo D, Ge J, et al. Towards automated organs at risk and target volumes contouring: Defining precision radiation therapy in the modern era. J Natl Cancer Cent 2022;2:306-13. [Crossref] [PubMed]
35. Li Y, Wu W, Sun Y, et al. The clinical evaluation of atlas-based auto-segmentation for automatic contouring during cervical cancer radiotherapy. Front Oncol 2022;12:945053. [Crossref] [PubMed]
36. Soltaninejad A, Shahbazi-Gahrouei D, Khorasani A, et al. Evaluation of CNN-based deep learning models for auto-contouring in glioblastoma radiotherapy: a review. Radiat Oncol 2025;20:169. [Crossref] [PubMed]
37. Duan J, Bernard M, Downes L, et al. Evaluating the clinical acceptability of deep learning contours of prostate and organs‐at‐risk in an automated prostate treatment planning process. Medical Physics. 2022;49:2570-81.
38. Lustberg T, van Soest J, Gooding M, et al. Clinical evaluation of atlas and deep learning based automatic contouring for lung cancer. Radiother Oncol 2018;126:312-7. [Crossref] [PubMed]
39. van Dijk LV, Van den Bosch L, Aljabar P, et al. Improving automatic delineation for head and neck organs at risk by Deep Learning Contouring. Radiother Oncol 2020;142:115-23. [Crossref] [PubMed]
40. Kalantar R, Lin G, Winfield JM, et al. Automatic Segmentation of Pelvic Cancers Using Deep Learning: State-of-the-Art Approaches and Challenges. Diagnostics (Basel) 2021;11:1964. [Crossref] [PubMed]
41. Zhang L, Tanno R, Xu MC, et al. Disentangling Human Error from Ground Truth in Segmentation of Medical Images. Advances in Neural Information Processing Systems 2020;33:15750-62.
42. Chanyavanich V, Das SK, Lee WR, Lo JY. Knowledge‐based IMRT treatment planning for prostate cancer. Medical Physics 2011;38:2515-22.
43. Liu J, Zhang X, Cheng X, et al. A deep learning-based dose prediction method for evaluation of radiotherapy treatment planning. Journal of Radiation Research and Applied Sciences 2024;17:100757.
44. Appelt AL, Elhaminia B, Gooya A, et al. Deep Learning for Radiotherapy Outcome Prediction Using Dose Data - A Review. Clin Oncol (R Coll Radiol) 2022;34:e87-96. [Crossref] [PubMed]
45. Kandalan RN, Nguyen D, Rezaeian NH, et al. Dose prediction with deep learning for prostate cancer radiation therapy: Model adaptation to different treatment planning practices. Radiother Oncol 2020;153:228-35. [Crossref] [PubMed]
46. Yousefi A, Ketabi S, Moreno AC, et al. An innovative process for efficient automated optimizing IMRT knowledge‐based planning (KBP). Medical Physics 2025;52:e18055.
47. Fanou AM, Patatoukas G, Chalkia M, et al. Implementation, Dosimetric Assessment, and Treatment Validation of Knowledge-Based Planning (KBP) Models in VMAT Head and Neck Radiation Oncology. Biomedicines 2023;11:762. [Crossref] [PubMed]
48. Oppitz H, Eckl M, Siebenlist K, et al. Knowledge-based automated radiation therapy treatment planning utilizing dose prediction with a 2.5 DU-Net. Physica Medica 2025;139:105199.
49. Moore KL. Automated Radiotherapy Treatment Planning. Seminars in Radiation Oncology 2019;29:209-18.
50. Lou Z, Cheng C, Mao R, et al. A novel automated planning approach for multi-anatomical sites cancer in Raystation treatment planning system. Phys Med 2023;109:102586. [Crossref] [PubMed]
51. Baroudi H, Brock KK, Cao W, et al. Automated Contouring and Planning in Radiation Therapy: What Is 'Clinically Acceptable'? Diagnostics (Basel) 2023;13:667. [Crossref] [PubMed]
52. Claessens M, Oria CS, Brouwer CL, et al. Quality Assurance for AI-Based Applications in Radiation Therapy. Semin Radiat Oncol 2022;32:421-31. [Crossref] [PubMed]
53. Fan J, Xing L, Ma M, et al. Verification of the machine delivery parameters of a treatment plan via deep learning. Phys Med Biol 2020;65:195007. [Crossref] [PubMed]
54. Niraula D, Sun W, Jin J, et al. A clinical decision support system for AI-assisted decision-making in response-adaptive radiotherapy (ARCliDS). Sci Rep 2023;13:5279. [Crossref] [PubMed]
55. Ono T, Iramina H, Hirashima H, et al. Applications of artificial intelligence for machine- and patient-specific quality assurance in radiation therapy: current status and future directions. J Radiat Res 2024;65:421-32. [Crossref] [PubMed]
56. Pillai M, Adapa K, Das SK, et al. Using Artificial Intelligence to Improve the Quality and Safety of Radiation Therapy. J Am Coll Radiol 2019;16:1267-72. [Crossref] [PubMed]
57. Bijman R, Rossi L, Sharfo AW, et al. Automated Radiotherapy Planning for Patient-Specific Exploration of the Trade-Off Between Tumor Dose Coverage and Predicted Radiation-Induced Toxicity-A Proof of Principle Study for Prostate Cancer. Front Oncol 2020;10:943. [Crossref] [PubMed]
58. Wang C, Zhu X, Hong JC, et al. Artificial Intelligence in Radiotherapy Treatment Planning: Present and Future. Technol Cancer Res Treat 2019;18:1533033819873922. [Crossref] [PubMed]
59. Chow JCL, Li K. Ethical Considerations in Human-Centered AI: Advancing Oncology Chatbots Through Large Language Models. JMIR Bioinform Biotechnol 2024;5:e64406. [Crossref] [PubMed]
60. Jafar A, Jia X. Towards human-centric intelligent treatment planning for radiation therapy. NPJ Digit Med 2026;9:155. [Crossref] [PubMed]
61. Lewandowska M, Street D, Yim J, et al. Artificial intelligence in radiation therapy treatment planning: A discrete choice experiment. J Med Radiat Sci 2025;72:S42-51. [Crossref] [PubMed]
62. Ejedegba EO. Equitable Healthcare in the Age of AI: Predictive Analytics for Closing Gaps in Access and Outcomes. International Journal of Research Publication and Reviews 2022;3:2882-94.
63. Sundaresan S, Zhang Z. AI-enabled knowledge sharing and learning: redesigning roles and processes. International Journal of Organizational Analysis 2021;30:983-99.
64. Aquino YSJ. Making decisions: Bias in artificial intelligence and data-driven diagnostic tools. Australian Journal of General Practice 2023;52:439-42.
65. Azenkot T, Rivera DR, Stewart MD, et al. Artificial Intelligence and Machine Learning Innovations to Improve Design and Representativeness in Oncology Clinical Trials. Am Soc Clin Oncol Educ Book 2025;45:e473590. [Crossref] [PubMed]
66. Labkoff S, Oladimeji B, Kannry J, et al. Toward a responsible future: recommendations for AI-enabled clinical decision support. J Am Med Inform Assoc 2024;31:2730-9. [Crossref] [PubMed]
67. Zakout G. Contextualising legal and ethical conundrums of artificial intelligence in oncology. The Lancet Oncology 2024;25:1113-6.
68. Heising LM, Verhaegen F, Scheib SG, et al. Toward a human‐centric co‐design methodology for AI detection of differences between planned and delivered dose in radiotherapy. Journal of Applied Clinical Medical Physics 2025;26:e70071.
69. Yan H, Yin FF, Willett C. Evaluation of an artificial intelligence guided inverse planning system: clinical case study. Radiother Oncol 2007;83:76-85. [Crossref] [PubMed]
70. Buccilli B, Lucke-Wold B. Performance of artificial intelligence for brain tumor segmentation, treatment planning, and outcome prediction: a comprehensive systematic review and meta-analysis. J Neurooncol 2025;176:88. [Crossref] [PubMed]
71. Wong J, Fong A, McVicar N, et al. Comparing deep learning-based auto-segmentation of organs at risk and clinical target volumes to expert inter-observer variability in radiotherapy planning. Radiother Oncol 2020;144:152-8. [Crossref] [PubMed]
72. Babier A, Zhang B, Mahmood R, et al. OpenKBP: the open‐access knowledge‐based planning grand challenge and dataset. Medical Physics 2021;48:5549-61.
73. Lou B, Doken S, Zhuang T, et al. An image-based deep learning framework for individualising radiotherapy dose: a retrospective analysis of outcome prediction. The Lancet Digital Health 2019;1:e136-47.
74. Barten DLJ, Pieters BR, Bouter A, et al. Towards artificial intelligence-based automated treatment planning in clinical practice: A prospective study of the first clinical experiences in high-dose-rate prostate brachytherapy. Brachytherapy 2023;22:279-89. [Crossref] [PubMed]
75. Pavunraj D, Kumar AM, Anbumaheshwari K. AI in personalized treatment planning. In: AI insights on nuclear medicine. IGI Global Scientific Publishing; 2025:117-42.
76. Hu Y, Nguyen H, Smith C, et al. Clinical assessment of a novel machine-learning automated contouring tool for radiotherapy planning. J Appl Clin Med Phys 2023;24:e13949. [Crossref] [PubMed]
77. Huang S, Wu J, Lin X, et al. Auto-Segmentation and Auto-Planning in Automated Radiotherapy for Prostate Cancer. Bioengineering (Basel) 2025;12:620. [Crossref] [PubMed]
78. Thongsawad S, Chamchod S, Chawengsaksopak K, et al. Quality and mechanical efficiency of automated knowledge‐based planning for volumetric‐modulated arc therapy in head and neck cancer. Journal of Applied Clinical Medical Physics 2025;26:e14588.
79. Aliotta E, Nourzadeh H, Choi W, et al. An Automated Workflow to Improve Efficiency in Radiation Therapy Treatment Planning by Prioritizing Organs at Risk. Adv Radiat Oncol 2020;5:1324-33. [Crossref] [PubMed]
80. Palazzo G, Mangili P, Deantoni C, et al. Real-world validation of Artificial Intelligence-based Computed Tomography auto-contouring for prostate cancer radiotherapy planning. Phys Imaging Radiat Oncol 2023;28:100501. [Crossref] [PubMed]
81. Kalsi S, French H, Chhaya S, et al. The Evolving Role of Artificial Intelligence in Radiotherapy Treatment Planning-A Literature Review. Clin Oncol (R Coll Radiol) 2024;36:596-605. [Crossref] [PubMed]
82. Zhang L, Liu Z, Zhang L, et al. Technical note: Generalizable and promptable artificial intelligence model to augment clinical delineation in radiation oncology. Med Phys 2024;51:2187-99. [Crossref] [PubMed]
83. Tudda A, Castriconi R, Benecchi G, et al. Knowledge-based multi-institution plan prediction of whole breast irradiation with tangential fields. Radiother Oncol 2022;175:10-6. [Crossref] [PubMed]
84. Mashayekhi M, McBeth R, Nguyen D, et al. Artificial intelligence guided physician directive improves head and neck planning quality and practice Uniformity: A prospective study. Clin Transl Radiat Oncol 2023;40:100616. [Crossref] [PubMed]
85. Aldhafeeri FM. Governing Artificial Intelligence in Radiology: A Systematic Review of Ethical, Legal, and Regulatory Frameworks. Diagnostics 2025;15:2300.
86. Conroy L, Winter J, Khalifa A, et al. Artificial Intelligence for Radiation Treatment Planning: Bridging Gaps From Retrospective Promise to Clinical Reality. Clin Oncol (R Coll Radiol) 2025;37:103630. [Crossref] [PubMed]
87. Huang K, Chung C, Ludmir EB, et al. Automatic end-to-end VMAT treatment planning for rectal cancers. J Appl Clin Med Phys 2024;25:e14259. [Crossref] [PubMed]
88. Zheng D, Hyun M, Fanning A. Ethical and regulatory considerations in artificial intelligence for adaptive radiation therapy. In: Artificial Intelligence in Adaptive Radiation Therapy. Bristol: IOP Publishing; 2025:1-21.
89. Vogelius IR, Petersen J, Bentzen SM. Harnessing data science to advance radiation oncology. Mol Oncol 2020;14:1514-28. [Crossref] [PubMed]
90. Caravatta L, Macchia G, Mattiucci GC, et al. Inter-observer variability of clinical target volume delineation in radiotherapy treatment of pancreatic cancer: a multi-institutional contouring experience. Radiat Oncol 2014;9:198. [Crossref] [PubMed]
91. Hurkmans C, Bibault JE, Clementel E, et al. Assessment of bias in scoring of AI-based radiotherapy segmentation and planning studies using modified TRIPOD and PROBAST guidelines as an example. Radiother Oncol 2024;194:110196. [Crossref] [PubMed]
92. Olin AB, Thomas C, Hansen AE, et al. Robustness and Generalizability of Deep Learning Synthetic Computed Tomography for Positron Emission Tomography/Magnetic Resonance Imaging-Based Radiation Therapy Planning of Patients With Head and Neck Cancer. Adv Radiat Oncol 2021;6:100762. [Crossref] [PubMed]
93. Liu X, Chen D, Liu Y, et al. Cross-technique transfer learning for autoplanning in magnetic resonance imaging-guided adaptive radiotherapy for rectal cancer. Phys Med 2025;129:104873. [Crossref] [PubMed]
94. Beckley J. Advanced Risk Assessment Techniques: Merging Data-Driven Analytics with Expert Insights to Navigate Uncertain Decision- Making Processes. International Journal of Research Publication and Reviews 2025;6:1454-71.
95. Huynh E, Hosny A, Guthier C, et al. Artificial intelligence in radiation oncology. Nat Rev Clin Oncol 2020;17:771-81. [Crossref] [PubMed]
96. Luo Y, Tseng HH, Cui S, et al. Balancing accuracy and interpretability of machine learning approaches for radiation treatment outcomes modeling. BJR Open 2019;1:20190021. [Crossref] [PubMed]
97. Jeong C, Goh Y, Kwak J. Challenges and opportunities to integrate artificial intelligence in radiation oncology: a narrative review. Ewha Med J 2024;47:e49. [Crossref] [PubMed]
98. Kawamura M, Kamomae T, Yanagawa M, et al. Revolutionizing radiation therapy: the role of AI in clinical practice. J Radiat Res 2024;65:1-9. [Crossref] [PubMed]
99. Dona Lemus OM, Cao M, Cai B, et al. Adaptive Radiotherapy: Next-Generation Radiotherapy. Cancers (Basel) 2024;16:1206. [Crossref] [PubMed]
100. Sibolt P, Andersson LM, Calmels L, et al. Clinical implementation of artificial intelligence-driven cone-beam computed tomography-guided online adaptive radiotherapy in the pelvic region. Phys Imaging Radiat Oncol 2021;17:1-7. [Crossref] [PubMed]
101. Chen W, Lu Y, Qiu L, et al. Designing personalized treatment plans for breast cancer. Information Systems Research 2021;32:932-49.
102. Gao Y, Tan T, Wang X, et al. Multi-Modal Longitudinal Representation Learning for Predicting Neoadjuvant Therapy Response in Breast Cancer Treatment. IEEE J Biomed Health Inform 2025;29:9041-50. [Crossref] [PubMed]
103. Yu Y, Chen D, Zhu Y. Multi-modal predictive modeling in prostate cancer radiotherapy: A radiomics and dosiomics approach. Journal of Radiation Research and Applied Sciences 2025;18:101518.
104. Skripcak T, Belka C, Bosch W, et al. Creating a data exchange strategy for radiotherapy research: towards federated databases and anonymised public datasets. Radiother Oncol 2014;113:303-9. [Crossref] [PubMed]
105. Mir BA, Abbas SR, Lee SW. Federated Learning in Healthcare Ethics: A Systematic Review of Privacy-Preserving and Equitable Medical AI. Healthcare (Basel) 2026;14:306. [Crossref] [PubMed]
106. Rauniyar A, Hagos DH, Jha D, et al. Federated learning for medical applications: A taxonomy, current trends, challenges, and future research directions. IEEE Internet of Things Journal 2023;11:7374-98.
107. Valdes G, Simone CB 2nd, Chen J, et al. Clinical decision support of radiotherapy treatment planning: A data-driven machine learning strategy for patient-specific dosimetric decision making. Radiother Oncol 2017;125:392-7. [Crossref] [PubMed]
108. Swart R, Boersma L, Fijten R, et al. Implementation Strategy for Artificial Intelligence in Radiotherapy: Can Implementation Science Help? JCO Clin Cancer Inform 2024;8:e2400101. [Crossref] [PubMed]
109. Chen H, Gomez C, Huang CM, et al. Explainable medical imaging AI needs human-centered design: guidelines and evidence from a systematic review. NPJ Digit Med 2022;5:156. [Crossref] [PubMed]
110. Karalis VD. The Integration of Artificial Intelligence into Clinical Practice. Applied Biosciences 2024;3:14-44.