With the advancement of precision radiotherapy, the management of setup errors has become a critical component in ensuring treatment safety and accurate dose delivery. In current clinical practice, image-guided radiotherapy (IGRT) and cone-beam computed tomography (CBCT) are commonly employed for periodic setup corrections. Although these approaches improve positioning accuracy, frequent imaging introduces additional radiation exposure and procedural workload, posing further challenges to clinical efficiency and staffing (1–3). Existing quality assurance (QA) procedures largely rely on manual experience or fixed threshold criteria, which are highly subjective, lack real-time responsiveness, and often fail to balance sensitivity and scalability. The coupling among six-dimensional (6D) setup errors is complex, and threshold-based methods are prone to missed detections or false alarms. Achieving high sensitivity under a low false-alarm rate remains a significant challenge in clinical quality control.

In recent years, artificial intelligence (AI) has achieved substantial progress in medical image analysis and radiotherapy workflow optimization (4–6). In particular, unsupervised learning methods can automatically identify abnormal patterns without the need for manual annotation, offering an alternative strategy for clinical quality assurance. For example, AI-based automatic organ segmentation systems have significantly improved contouring efficiency and consistency in clinical practice (7).

Although AI has demonstrated outstanding performance in treatment planning and quality assurance (QA) tasks, there is currently relatively limited systematic research, either domestically or internationally—on real-time quality control of six-dimensional (6D) setup error detection. This research gap leaves radiotherapy QA largely in the “offline detection” stage, without achieving a closed-loop system of “real-time monitoring and intelligent alerting.”

In this study, a total of 7,539 CBCT-based 6D setup error datasets collected between May 2022 and March 2025 were used to construct and validate an unsupervised learning–driven framework for real-time abnormality detection. The Isolation Forest (IF) algorithm was employed as the primary model, with the Local Outlier Factor (LOF) serving as a reference. Potential outliers were defined using a 3σ threshold for performance evaluation.

Methodologically, this study integrates stratified analyses (across different immobilization methods and treatment sites), interpretability analyses (via principal component analysis [PCA] and Spearman correlation), and longitudinal stability assessment to comprehensively evaluate the model’s performance and clinical feasibility. Finally, a closed-loop “setup–monitoring–alert” framework is proposed to facilitate multicenter translation and promote the clinical implementation of AI-driven radiotherapy quality assurance.

Based on a large-scale dataset of six-dimensional (6D) setup errors derived from CBCT, this study systematically evaluated the feasibility of applying unsupervised machine learning for anomaly detection and quality control (QC) in radiotherapy. The results demonstrated that the Isolation Forest (IF) model significantly outperformed the Local Outlier Factor (LOF) in overall performance, maintaining high recall and precision even under low false alarm conditions (see Table 3, Figure 5). These findings validate the suitability and clinical potential of IF for real-time anomaly alerting in radiotherapy workflows (12–15).

The moderate PR-AUC reflects the inherent trade-off between sensitivity and false alerts in low base-rate detection tasks, underscoring that alert frequency must be clinically calibrated to avoid alert fatigue.

Importantly, a statistically abnormal geometric setup does not necessarily translate into a clinically meaningful dosimetric impact. In routine image-guided radiotherapy, most setup deviations are corrected before dose delivery, and their dosimetric relevance depends on factors such as target margins, beam geometry, and the spatial relationship between targets and organs at risk.

In the absence of direct dosimetric recalculation or clinical outcome endpoints, we reported geometric tolerance exceedance rates based on recorded six-dimensional couch corrections as a minimal translational proxy. While translational exceedance beyond common clinical thresholds was rare, a small subset of fractions exhibited larger rotational deviations, which may warrant prioritized quality assurance review rather than direct inference of dose error. Accordingly, the proposed framework is positioned as an auxiliary quality control alerting tool rather than a surrogate for dosimetric evaluation or clinical decision-making.From the perspective of immobilization methods (Table 2, Figure 1), all subgroups achieved AUC ≥ 0.92, indicating consistent model performance under diverse fixation conditions. The SRT dual-face thermoplastic mask and neck–thorax mask combined with a vacuum cushion achieved near-ideal performance (AUC = 0.994 and 1.000, respectively), suggesting that higher fixation stability enhances the separability between normal and abnormal samples, thereby improving detection accuracy. Conversely, fixation methods with greater adaptability but lower constraint—such as the vacuum cushion and U-shaped mask—still maintained robust AUC values (0.925–0.930), demonstrating the framework’s cross-strategy robustness. However, smaller subgroups, such as the SBRT fixation frame (N = 51) and abdominal mask with vacuum cushion (N = 6), exhibited greater uncertainty in AUC estimation due to limited sample size (16). The consistency of results under both global and site-stratified normalization further suggests that the reported performance was not driven by the choice of normalization strategy.This also implies that a proportion of statistically defined anomalies may not trigger alerts at low false-positive operating points, further supporting the role of this framework as a screening and prioritization tool rather than an exhaustive detector.

Across different treatment sites (Table 1), the model performed best in head and neck patients (AUC ≈ 0.973), followed by thoracic (AUC ≈ 0.954) and abdominal (AUC ≈ 0.946) cohorts. These variations may be attributed to differences in immobilization adaptability and the physiological motion of the respiratory and digestive systems. The findings suggest that, for thoracic and abdominal radiotherapy, reliance on a single surface or image registration modality may be insufficient. Future QC frameworks should incorporate respiratory gating or multimodal monitoring approaches (e.g., SGRT combined with pressure sensors) to enhance overall control accuracy (17, 18).

The interpretability analysis further elucidated the principal factors driving anomaly detection. PCA-based dimensionality reduction revealed that abnormal samples formed relatively distinct clusters within the first two principal components, which together explained approximately 45% of total variance (Figure 2). This supports the statistical independence of abnormal patterns in multidimensional space. The feature importance analysis indicated that the anterior–posterior (AP) and Pitch directions contributed most strongly to anomaly scoring, consistent with the dominant axes of positioning errors in clinical thoracoabdominal radiotherapy (Figure 3). These findings align with clinical observations—that rotational deviations are often more difficult to detect through conventional surface inspection—and provide quantitative support for the model’s clinical interpretability (19, 20). In the present study, feature importance was evaluated at the global level; subgroup-specific importance patterns were not separately modeled and warrant future investigation.

The longitudinal trend analysis (Figure 4) showed that performance metrics (ROC-AUC and PR-AUC) (21) remained stable across all temporal cohorts, with no evidence of systematic drift or degradation. This finding indicates that the model maintains long-term discriminative stability, supporting its potential use for continuous quality monitoring in radiotherapy.

Compared with the traditional z-score thresholding method (22), the Isolation Forest (IF) achieved an approximately 0.21 improvement in ROC-AUC (0.960 vs. 0.754). Under a clinically acceptable false positive rate (FPR ≈ 5%), both recall and precision were markedly higher (0.52/0.71 vs. 0.25/0.52), indicating superior usability of the proposed approach in real-time clinical alerting scenarios (see Table 3).

Major Innovations of This Study:

1. Introduction of Unsupervised Learning into Radiotherapy

Quality Control:

To the best of our knowledge, this study is the first to integrate unsupervised machine learning into radiotherapy quality assurance (QA), moving beyond traditional statistical thresholds and human experience–based approaches. The proposed method enables efficient, real-time, and scalable anomaly detection. The image-guided QA workflow was designed with reference to the AAPM TG-179 report on QA for CT-guided radiotherapy systems, emphasizing a “human–AI collaborative” QC model in which AI serves only as an auxiliary tool for detection and alerting, while final decisions remain the responsibility of clinicians and medical physicists.

As volumetric modulated arc therapy (VMAT) has largely replaced conventional IMRT in multi-site radiotherapy due to its higher efficiency and superior organ-at-risk (OAR) sparing, the introduction of AI can further improve workflow efficiency. However, AI cannot fully replace human QA, as excessive reliance could introduce new safety risks (23–26).

2. Multi-level Validation: The framework was validated from five complementary perspectives—overall performance, immobilization methods, treatment sites, interpretability, and longitudinal stability—ensuring systematic and generalizable results across diverse clinical contexts.

3. Enhanced Interpretability: Through principal component analysis (PCA) and feature importance analysis, key error dimensions contributing to anomalies were identified, providing clinically meaningful insights that improve understanding and practical adoption of AI-driven QC models.

4. Longitudinal Drift Detection: To the best of our knowledge, AI-derived anomaly scores were applied to track potential systematic drifts over time, offering valuable references for early QC alerting and adaptive radiotherapy (ART) decision-making.

Despite its promising results, this study has several limitations.First, the definition of anomalies was based on a statistical threshold (|z| > 3) rather than true clinical mistreatment events. Future research should incorporate dosimetric validation and annotations of actual adverse clinical events to refine the labeling process.Second, this study is subject to center-dependent limitations. All data were collected from a single institution using a specific combination of linear accelerator platform (Varian VitalBeam), on-board CBCT registration system, and local IGRT decision rules. Variations across centers in CBCT image characteristics (e.g., noise properties), ROI definition, registration algorithms, availability and utilization of 6D couch corrections, and operator-dependent manual review practices may influence setup error distributions and anomaly patterns. Therefore, direct generalization of the reported performance should be interpreted cautiously, and further multicenter validation is warranted.Finally, this work compared only the Isolation Forest (IF) and Local Outlier Factor (LOF) algorithms. Future studies should explore more advanced models—such as autoencoders, variational autoencoders (VAE), and graph neural networks (GNN)—to enhance the detection of more complex and subtle anomalies (27–30).

In summary, this study shows the feasibility and practical value of an unsupervised learning framework for detecting six-dimensional setup error anomalies in radiotherapy. The multi-level validation results support its potential integration into future intelligent QA systems, particularly when combined with emerging technologies such as surface-guided radiotherapy (SGRT) and adaptive radiotherapy (ART). Such integration could help establish an automated, real-time clinical quality control workflow, further improving the precision and safety of radiotherapy.

Based on a large-scale dataset of six-dimensional (6D) setup errors, this study proposed and validated an unsupervised learning–driven framework for identifying statistically unusual setup patterns in radiotherapy. The results demonstrated that the Isolation Forest (IF) model exhibited excellent discriminative performance overall and across different immobilization methods and treatment sites.Interpretability analysis identified the anterior–posterior (AP), Pitch, and left–right (LR) directions as the major contributors to anomaly detection. The longitudinal analysis further revealed the framework’s ability to identify potential system drifts, providing an early-warning reference for individualized quality control and adaptive radiotherapy (ART).

Methodologically, this study verified the effectiveness of unsupervised learning in anomaly detection and proposed a closed-loop “setup–monitoring–alert” quality control framework. Future work should incorporate dosimetric effect analysis, multicenter validation, and deep learning models to facilitate the transition of this framework from proof-of-concept to standardized clinical implementation.