A Real-Time Patient Orientation Verification System using Pose Estimation in CT Simulation

A Real-Time Patient Orientation Verification System using Pose Estimation in CT Simulation | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Real-Time Patient Orientation Verification System using Pose Estimation in CT Simulation Hyunji Amy Kim, Elliott Choi, Estella Choi, Youngjin Lee, Daegun Kim, and 3 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9284874/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Patient orientation mis-entry during computed tomography (CT) simulation represents a critical failure mode that corrupts the digital imaging and communications in medicine (DICOM) coordinate system, necessitating patient recall for re-scanning or complex manual corrections. We developed a tool called just another secondary evaluation (JASE), an automated real-time verification system to prevent these errors. JASE uses a standard webcam to classify patient orientation into four standard categories: head-first-supine (HFS), head-first-prone (HFP), feet-first-supine (FFS), and feet-first-prone (FFP). We compared two approaches: transfer learning with MobileNetV2 and a multilayer perceptron (MLP) using MediaPipe pose landmarks. The system displays real-time orientation on an LCD screen adjacent to the CT console. The MLP model achieved 98.3% accuracy (1/49 misclassification) versus 96.2% for MobileNetV2 (3/80 misclassifications) on an independent test set. Real-time implementation demonstrated stable performance, which indicates the potential for future clinical deployment pending validation in actual CT simulation workflows. JASE provides a low-cost, effective solution for preventing orientation errors at the point of entry. The system’s applicability extends beyond radiation oncology to radiology and nuclear medicine departments using similar imaging protocols. Patient safety CT simulation Orientation verification Pose estimation Quality assurance Figures Figure 1 Figure 2 Figure 3 Figure 4 I. INTRODUCTION Modern radiation therapy relies on precise geometric accuracy throughout the treatment workflow. The foundation of this precision begins with a computed tomography (CT) simulation, where patient spatial information is first established. However, a critical yet often overlooked failure mode occurs when therapists manually enter incorrect patient orientation at the CT console. This human error – recording head-first-supine (HFS) as head-first-prone (HFP), for instance – fundamentally corrupts the digital imaging and communications in medicine (DICOM) coordinate system by incorrectly defining spatial tags. The consequences of orientation mis-entry are severe. In Eclipse treatment planning system (TPS) (Varian Medical System, Palo Alto, CA), this error causes complete geometric misinterpretation: anterior-posterior axes invert, left-right orientations reverse, and the entire 3D coordinate system becomes corrupted. Treatment beams intended for tumors may instead target healthy tissue, rendering plans clinically useless and potentially catastrophic. Such errors have been documented to cause wrong-site treatments and near-miss events. These events, along with others such as incorrect protocol selection or scanning unintended anatomy, pose risks to patient safety, increase healthcare costs, and have potential legal and regulatory implications. Current remediation strategies are reactive and resource intensive. Most institutions resort to patient recall for complete re-scanning, exposing patients to additional radiation, delaying treatment starts, and consuming valuable scanner time and resources. The alternative – manual DICOM header correction – requires specialized expertise and carries its own risks of introducing further errors. Our institution’s correction protocol (see supplementary material) illustrates the complexity: correcting HFS to HFP requires sequential patient position changes, image reorientation with 180° rotation, and new 3D dataset generation. For HFS to FFS errors, the process involves position correction, slice order inversion, and beta-axis rotation – each step introducing potential for additional errors. The American Association of Physicists in Medicine (AAPM) Task Group 100 (TG–100) [1] advocates for proactive, risk-based quality management that identifies and mitigates potential failures before they occur. From this perspective, orientation mis-entry represents a high-risk failure mode due to its catastrophic potential combined with low detectability using current procedures. The AAPM has a proven track record of proactively addressing patient safety concerns, such as the collaboration with Medical Imaging and Technology Alliance (MITA) that resulted in the dose alert feature now standard on all CT scanners. Despite clear risks, no proactive verification systems exist that utilize standard clinical hardware. We developed a tool named just another secondary evaluation (JASE) to fill this critical gap. By providing real-time, automated orientation verification before DICOM data finalization, JASE transforms a low-detectability/high-impact failure into a preventable error, aligning with modern patient safety principles. II. MATERIALS AND METHODS 1. System architecture and dataset JASE consists of a webcam positioned to view the CT simulation couch, classification algorithm running on a standard computer, and an LCD display mounted adjacent to the CT console. The webcam captures continuous video of patients during positioning, which is processed in real-time to determine orientation before the therapist finalizes the DICOM header information. Figure 1 demonstrates the system architecture, showing the data flow from webcam capture through pose extraction and classification to the final LCD display, along with examples of the four standard patient orientations detected by the system. To develop and validate the classification algorithm, we created the JASE dataset by filming volunteers positioned on a CT simulation couch in each of the four standard orientations: HFS, HFP, FFS, and FFP. A total of eight healthy volunteers (4 female, 4 males; age range 16–35 years) were recruited for dataset creation. All participants were non-patients, and recordings were performed with volunteers fully clothed. The webcam was mounted inside the CT control room along the x-ray glass window, positioned approximately 1.5 meters above couch height and angled downward at roughly 35–40 degrees toward the patient to ensure full-body visibility during positioning. The videos were then converted to MP4 format and processed using open-source computer vision library (OpenCV) to extract individual frames. Every 10th frame was extracted to reduce redundancy while maintaining sufficient variation. This process yielded 1,539 original images distributed across the four orientation classes. The dataset was expanded through manual augmentation including vertical flipping, horizontal mirroring, and rotation to introduce greater positional variation. Additional fine-tuning data was collected using the same protocol and integrated into the dataset. For testing, we reserved 20 images per orientation class (10 from original data, 10 from fine-tuning data) creating an 80-image test set separate from training data. The final dataset contained 6,442 images organized into folders by orientation class, processed using Keras ImageDataGenerator with automatic assignment of class labels (FFP=0, FFS=1, HFP=2, HFS=3). 2. Model development Both models used in this project—MobileNetV2 and multilayer perceptron (MLP)—are forms of artificial neural networks (ANNs), computational systems inspired by the human brain [2]. ANNs consist of interconnected layers: input, hidden, and output layers [2]. For JASE, inputs are either raw image pixels (convolutional neural network (CNN) model) or body joint coordinates from pose estimation (MLP model). The output layer contains four neurons corresponding to the orientation classes (FFP, FFS, HFP, HFS). Models learn by adjusting weights and biases through backpropagation to minimize loss functions, enabling them to learn complex patterns required for accurate classification [3]. We implemented two distinct approaches. The first utilized transfer learning with MobileNetV2, a lightweight CNN optimized for embedded applications. MobileNetV2 employs inverted residual blocks and linear bottlenecks that maintain spatial information while reducing computation through depthwise separable convolutions. Unlike traditional CNNs, its inverted residual blocks expand input with 1×1 convolution, apply depthwise convolution, then compress back to low dimensions—ideal for real-time deployment [4]. For transfer learning, we froze the pretrained base layers and added custom classification layers consisting of a dropout layer with 30% rate, followed by a dense layer with 128 neurons using a Rectified Linear Unit (ReLU) activation, another dropout layer at 25%, and a final dense output layer with 4 neurons using softmax activation. The 128-unit dense layer was selected for sufficient representational capacity for the four-class orientation task without over-parameterizing the model relative to the size of our dataset. Freezing basic layers was necessary since the architecture for a basic MobileNetV2 is very extensive with millions of trainable parameters. If these layers were not frozen, training the model would increase the risk of overfitting, and require substantially more computational resources. Moreover, dropout layers at 25% randomly drop 25% of the neurons during training, which prevents overfitting and significantly increases the performance of a neural network by contributing to variance in the fine-tuned model [5]. To further inhibit overfitting, class weights were calculated so that the model can prioritize classes that are underrepresented in the dataset. Subsequently, during fine-tuning we unfroze top layers and trained with learning rate 0.0001 to prevent catastrophic forgetting, using early stopping to monitor validation loss [6]. The second approach used MediaPipe to extract 33 anatomical landmarks from each image, generating normalized x, y, z coordinates. These coordinates represent relative positions—for instance, (0.388, 0.303, -0.463) indicates 38.8% from left, 30.3% from top, with negative z showing camera proximity. This created 99–dimensional feature vectors capturing body configuration independent of visual appearance. The MLP followed a funnel architecture [7] with dense layers progressively decreasing from 512 to 256 to 128 to 64 units, incorporating batch normalization and a 25% dropout between layers. This configuration balanced model capacity and regularization relative to the low-dimensional 99–point pose input. Larger or wider architectures resulted in overfitting during development, whereas smaller networks underperformed due to insufficient representational capacity [8]. The final classification layer used softmax activation across four outputs. Both models employed class weight balancing to handle minor dataset imbalances and prevent overfitting. 3. Real-time implementation and evaluation Real-time deployment integrated video capture, model inference, and result display using OpenCV for image acquisition and preprocessing. The system captured continuous video from the USB webcam, extracting every 10th frame to balance processing load with responsiveness. Each frame underwent resizing and normalization before classification. Hardware display utilized an Arduino UNO R3 microcontroller connected to a 2-inch TFT LCD module (ST7789-based) via serial peripheral interface (SPI) interface. The Python classification script transmitted predicted orientation labels to the Arduino through USB serial communication at 9600 baud using PySerial. The LCD initially displayed "READY" and updated dynamically as predictions arrived, with text color indicating confidence levels: green for high confidence, yellow for moderate, and red for low confidence predictions. Model evaluation employed standard classification metrics including accuracy, precision, recall, and F1-score calculated using scikit-learn. Confusion matrices visualized error patterns between orientation classes. Real-time performance assessment measured inference latency, end-to-end system delay, and operational stability over extended periods. Figure 2 illustrates the complete hardware setup including Arduino wiring configuration and example real-time orientation display. 4 . Evaluation metrics Model performance was evaluated using standard classification metrics: balanced accuracy, precision, recall, and F1-score, calculated using scikit-learn. These metrics provide a comprehensive assessment of classification performance, with F1-score and balanced accuracy offering a balanced measure particularly important for datasets with minor class imbalances. 5 . Ethical Considerations All videos used in this study were captured on an iPhone as MP4 recordings during volunteer simulations conducted by members of the research team and colleagues from the same laboratory. The data did not originate from clinical scanners, hospital information systems, or DICOM-based workflows and contained no PHI beyond visible facial features. According to institutional policy, such non-clinical simulation activities are classified as not human subjects research and are exempt from IRB review and written consent requirements. All participants were informed of the project purpose and verbally agreed to the use of their images for research and publication. Data collection was limited to volunteer simulations rather than the clinical patients because RGB video capture may reveal identifiable facial features. This constrain influenced the scope of the present study, which focuses on proof-of-concept validation rather than clinical generalization. III. RESULTS The fine-tuned transfer learning model achieved a balanced accuracy of 96.2%, a precision of 96.7%, a recall of 96.3%, and an F1-score of 96.3%. The MLP model had a better performance, as it achieved a balanced accuracy of 98.3%, a precision of 98.1%, a recall of 98.0%, and an F1-score of 98.0%. Based on the confusion matrix below, the fine-tuned transfer learning model misclassified the HFP class once and the FFP class twice. The MLP model only misclassified the FFP class as the HFP class once. These performances indicate that the MLP model outperforms the fine-tuned transfer learning model in all evaluated metrics. Given the limited size of the hold-out test set (80 images for the image-based model and 49 landmark-valid images for the MLP model), these metrics should be interpreted as indicative of technical feasibility rather than statistically powered performance estimates. Results are visualized in Figure 3 (performance metrics comparison) and Figure 4 (confusion matrices for both models), where diagonal entries represent correct predictions and off-diagonal values indicate specific misclassification patterns between orientation classes. IV. DISCUSSION The superiority of pose-based classification over raw image analysis (MLP 98.3% vs transfer learning 96.2% accuracy) reveals a fundamental insight: skeletal geometry provides more robust orientation markers than pixel-level features. While image-based models must contend with clinically irrelevant variables—ambient lighting, patient clothing, blankets—pose landmarks offer an invariant representation of patient configuration [9]. This robustness proves essential for deployment across diverse clinical environments where visual conditions vary unpredictably. JASE exemplifies the proactive quality management philosophy advocated by AAPM TG–100 [1]. Traditional approaches detect orientation errors only after they propagate downstream, often during plan review or, worse, at treatment delivery [10]. By implementing targeted process control at the point of error origin—the CT console—JASE transforms a low-detectability/high-impact failure into a preventable event. The system enhances what TG–100 terms "error detectability" without adding workflow complexity, directly reducing the risk priority number for this failure mode. It is important to note that the application of TG–100 principles in this study is conceptual, as we did not measure quantitative changes in event frequency or workflow time. Such risk-reduction analysis requires prospective deployment in clinical CT simulation workflows, which is beyond the scope of this technical feasibility study but will be a focus of future works. Development challenges illuminated important technical considerations. The transfer learning model's susceptibility to overfitting [11] and catastrophic forgetting [12] required careful regularization through early stopping, data augmentation, and learning rate optimization [13]. These challenges, while successfully addressed, underscore why the simpler pose-based approach ultimately proved superior—it inherently avoids many pitfalls of image-based learning by operating on abstracted anatomical features rather than raw pixels. The present study did not include a systematic robustness evaluation across varying patient body habitus, clothing, or immobilization devices. Model evaluation was intentionally limited to common positioning conditions encountered within our acquisition environment, consistent with the feasibility-oriented scope of this work. Current limitations center on MediaPipe's landmark detection reliability. Some images lacked pose landmark visibility, causing MediaPipe to fail in extracting key points. This limitation is a result of the setup of the webcam. In our setup, the slightly elevated and oblique angle sometimes obscured extremities such as feet or hands, which led to incomplete landmark extraction and contributed to MediaPipe’s occasional detection failures. This finding highlights the significance of the camera positioning; it is crucial that major body landmarks are visible to the camera, or the MLP will have many inaccuracies and impair its performance. As a result, a number of images were excluded from both the training set and test set, with 2282 images excluded from the training set and 31 images excluded in the testing set. Because only landmark-complete samples could be used by the MLP architecture, the effective dataset represents only a subset of the full image collection. This introduces a selection bias, as cases with occluded extremities or partial body visibility are systematically underrepresented in the landmark-valid subset. As a result, performance comparisons between the MLP and MobileNetV2 models should be interpreted within the context of this evaluable subset rather than as a fully matched dataset-wide comparison. The final hold-out test set contained only 80 images (20 per class), which limits the statistical robustness of the reported metrics. The results should therefore be interpreted as feasibility-level findings rather than fully powered performance estimates. Cross-validation was considered; however, it was not appropriate for this dataset. First, as the MLP model could only be evaluated on images in which all 99 pose landmarks were successfully extracted, incomplete landmark sets prevented a consistent division of data across folds. In addition, ensuring evaluation consistency between the MobileNetV2 model—which uses all images—and the MLP model—which uses only landmark-valid images—would not be feasible under a cross-validation scheme. Inference-time measurements, confidence distribution analysis, or statistical comparison tests (e.g., McNemar’s test) were not included in the present analysis, as the limited and imbalanced landmark-valid subset prevents sufficiently powered paired statistical evaluation. These analyses will require a larger and more diverse dataset and are identified as important directions for future work. For these reasons, we adopted a strict hold-out test set, which provides a more conservative and methodologically appropriate estimate of model performance for the feasibility of this study. Nevertheless, the MLP model consistently outperformed the transfer learning model in classification performance and real-time deployment. Additionally, patients with atypical anatomy, amputations, or complete visual occlusion present edge cases requiring further development. The single-institution training dataset also limits generalizability claims. However, these limitations are addressable through expanded data collection and multi-input hybrid architectures [14] combining pose and image features. Although the dataset included variability in camera angles and lighting both through natural acquisition differences and augmentation, the present study did not include a systematic robustness evaluation under controlled variations in illumination, viewpoint, or occlusion. It is important to note that this limitation primarily affects offline, frame-based dataset construction rather than the intended real-time deployment scenario. In practical use, the JASE system operates on continuous live video streams, where orientation predictions are generated across successive frames rather than relying on isolated still images. In this setting, transient landmark extraction failures in individual frames are mitigated by temporal redundancy, as subsequent frames typically yield valid pose landmarks. Orientation decisions can therefore be stabilized across multiple frames, reducing the impact of occasional landmark incompleteness on real-time system performance. Consequently, while landmark-incomplete frames were excluded during offline evaluation, this exclusion is less problematic in the envisioned real-time clinical workflow. Furthermore, although all data were collected from healthy volunteers rather than clinical patients, the acquisition was performed inside an actual CT simulation room using the same equipment, couch setup, lighting, and physical environment encountered in routine practice. Volunteers were positioned following standard CT simulation procedures so that the captured videos closely mimicked real clinical presentations. Nevertheless, volunteer-based simulations cannot fully replicate the anatomical variability, immobilization devices, movement patterns, or workflow factors seen in real patient populations. As a result, formal validation using real patient data and prospective pilot testing with radiologic technologists will be required to establish true clinical deployment readiness. Beyond radiation oncology, other imaging environments also use standardized patient-orientation conventions (HFS, HFP, FFS, FFP), suggesting potential applicability of the JASE framework to modalities such as diagnostic CT, PET/CT, and MRI [15]. The core model relies on skeletal pose geometry extracted from RGB video rather than modality-specific image features, which conceptually enables adaptation if new training datasets are collected under modality-appropriate positioning conditions. For example, capturing MRI-specific positioning videos (with modality-specific couch designs and coil placements) would allow retraining of the pose-based MLP architecture. Recent advances in MRI safety protocols have focused on developing real time monitoring systems and improved patient positioning techniques. The introduction of advanced positioning aids and continuous monitoring technologies has significantly reduced the incidence of positioning related adverse events in MRI suites [16, 17]. These developments are particularly important given the increasing complexity of MRI examinations and the growing patient population with implanted devices or medical conditions that require special attention during positioning. Although such modalities were not evaluated in this study, the concept may generalize pending the acquisition of modality-specific data and prospective validation. The path to clinical implementation therefore requires expanding training data to include edge cases, validating performance across multiple institutions, and exploring direct integration with imaging consoles. A prospective trial measuring error-prevention rates and workflow impact would definitively establish clinical value. Given that orientation errors can have high clinical consequence, even modest reductions could yield meaningful safety improvements across imaging departments. V. CONCLUSION JASE demonstrates that real-time orientation verification can effectively prevent a critical error mode in clinical imaging. The system's 98.3% accuracy, low cost, and seamless integration demonstrate its potential as a practical quality assurance tool aligned with modern risk-based management principles. Deployment across imaging modalities could significantly reduce orientation-related errors throughout healthcare institutions. Declarations Conflict of interest : The authors declare no conflict of interest. Ethics approval : None. Informed consent : None. Funding: This study was supported by a grant from the National Foundation of Korea (NRF) funded by the Korean government (Grant No. RS-2024-00354252 & RS-2025-02303425). Author Contribution All authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by HAK, Elliott C, Estella C, and JJS. The first draft of the manuscript was written by HAK and JJS and all authors commented on previous versions of the initial manuscript internally. All authors read and approved the final submitted version of manuscript. Acknowledgement: None. Data Availability Data will be made available on request. References M. S. Huq, B. A. Fraass, P. B. Dunscombe, J. P. Gibbons Jr., G. S. Ibbott, A. J. Mundt, S. Mutic, J. R. Palta, F. Rath, B. R. Thomadsen, J. F. Williamson and E. D. Yorke, Med. Phys. 43, 4209 (2016). S. H. Han, K. W. Kim, S. Y. Kim and Y. H. Youn, Dement. Neurocogn. Disord. 17, 83 (2018). A. El-Shahat, in Introductory Chapter: Artificial Neural Networks (InTech, 2018). M. Sandler, A. Howard, M. Zhu, A. Zhmoginov and L. C. Chen, in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 4510 (2018). N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever and R. Salakhutdinov, J. Mach. Learn. Res. 15, 1929 (2014). C. Yang and X. Ma, in Proc. EMNLP , 4854 (2022). S. Klein, J. A. Raine, S. Pina-Otey, S. Voloshynovskiy and T. Golling, arXiv:2112.08069 (2021). A. Meyer-Baese and V. Schmid, Foundations of Neural Networks (Elsevier, p. 197). S. S. Sengar, A. Kumar and O. Singh, arXiv:2406.15649 (2024). K. Bell, N. Lickt, C. Rübe and Y. Dzierma, Radiat. Oncol. 13, 1 (2018). X. Ying, J. Phys. Conf. Ser. 1168, 022022 (2019). J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran and R. Hadsell, Proc. Natl. Acad. Sci. U.S.A. 114, 3521 (2017). P. Kenneweg, A. Schulz, S. Schröder and B. Hammer, Lect. Notes Comput. Sci. , 252 (2022). N. Menet, M. Hersche, G. Karunaratne, L. Benini, A. Sebastian and A. Rahimi, arXiv:2312.02829 (2023). H. K. Lim, Y. Choi, J. H. Jung, J. W. Jung, C. H. Oh, H. W. Park, et al., J. Korean Phys. Soc. 80, 640 (2022). V. Edelman, H. Chassidim and I. Rabaev, Electronics 14, 668 (2025). ] Y. J. Seol, J. N. Kim, A. R. Kim, J. H. Hwang, T. G. Oh, J. S. Shin, et al., J. Korean Phys. Sco. 73, 1012 (2018). Additional Declarations No competing interests reported. Supplementary Files JKPSSupplementarymaterialJASE260401.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 26 Apr, 2026 Reviews received at journal 26 Apr, 2026 Reviews received at journal 14 Apr, 2026 Reviewers agreed at journal 07 Apr, 2026 Reviewers agreed at journal 06 Apr, 2026 Reviewers invited by journal 06 Apr, 2026 Editor assigned by journal 03 Apr, 2026 Submission checks completed at journal 02 Apr, 2026 First submitted to journal 31 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9284874","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":619207807,"identity":"3127feb1-3d61-4302-a6f5-c67813a82b5c","order_by":0,"name":"Hyunji Amy Kim","email":"","orcid":"","institution":"University of Chicago","correspondingAuthor":false,"prefix":"","firstName":"Hyunji","middleName":"Amy","lastName":"Kim","suffix":""},{"id":619207809,"identity":"8362f810-1640-4ecd-bc97-b4d69e842b5a","order_by":1,"name":"Elliott Choi","email":"","orcid":"","institution":"University of Chicago","correspondingAuthor":false,"prefix":"","firstName":"Elliott","middleName":"","lastName":"Choi","suffix":""},{"id":619207810,"identity":"a50be20a-ed49-44de-8013-0f4f8ede344b","order_by":2,"name":"Estella Choi","email":"","orcid":"","institution":"University of Chicago","correspondingAuthor":false,"prefix":"","firstName":"Estella","middleName":"","lastName":"Choi","suffix":""},{"id":619207812,"identity":"8f2c685f-c81f-4f31-971f-d4b7ff4bab18","order_by":3,"name":"Youngjin Lee","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0klEQVRIiWNgGAWjYFCCwwcOfABSBlCuAT61UHAs8eEMErXwGBvzkKTF4OABM2mbX3by5uy9Bxh+1DAYmzcQ0nLgQJp0bl+y4c6ecwmMPccYzGQOENZyTDq3hznB4EaOAQNvA4ONBEGHHTjYJm3ZUw/WwviXOC2HmY0ZfhwGa2EG2mJGUIvkgWOMD3sbjhtuOHPG4LDMMQljglr4bpz/cODHn2p5g+M9hg/f1NgYziCkReHGAQYGxjYIB8gkaAcDg3x/A5D8Q1jhKBgFo2AUjGAAALOpRSg3PuhRAAAAAElFTkSuQmCC","orcid":"","institution":"Gachon University","correspondingAuthor":true,"prefix":"","firstName":"Youngjin","middleName":"","lastName":"Lee","suffix":""},{"id":619207813,"identity":"6cfc7879-3749-420e-aac7-8792d4472c66","order_by":4,"name":"Daegun Kim","email":"","orcid":"","institution":"Soonchunhyang University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Daegun","middleName":"","lastName":"Kim","suffix":""},{"id":619207814,"identity":"21ae0041-a96b-4efd-9e29-cd32adb32f40","order_by":5,"name":"Jae-Hong Jung","email":"","orcid":"","institution":"Soonchunhyang University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jae-Hong","middleName":"","lastName":"Jung","suffix":""},{"id":619207815,"identity":"49e4fc44-8bf9-4add-bba3-f35ee73d2795","order_by":6,"name":"Kyung-Bae Lee","email":"","orcid":"","institution":"Wonkwang University","correspondingAuthor":false,"prefix":"","firstName":"Kyung-Bae","middleName":"","lastName":"Lee","suffix":""},{"id":619207816,"identity":"3c7efd82-7904-42f8-9369-d0dfe63776f2","order_by":7,"name":"James J. Sohn","email":"","orcid":"","institution":"University of Chicago","correspondingAuthor":false,"prefix":"","firstName":"James","middleName":"J.","lastName":"Sohn","suffix":""}],"badges":[],"createdAt":"2026-04-01 00:08:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9284874/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9284874/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106960544,"identity":"89ccdd13-9479-4383-a164-9f8ffff08e3e","added_by":"auto","created_at":"2026-04-15 09:21:43","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":314329,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart depicting the creation and implementation pipeline of the JASE model.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-9284874/v1/aaf7288dfdd149cbcfdf2974.png"},{"id":106836394,"identity":"49d4aee2-fa04-45a0-9438-99ebb84dfb42","added_by":"auto","created_at":"2026-04-14 02:08:01","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":655743,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Webcam placement inside the CT control room, positioned approximately 1.5 m above the couch height and angled downward at ~35–40° to capture the full patient body during positioning. (b) Architecture of the JASE pose-based MLP classification model using MediaPipe landmark inputs. (c) Pin configuration of the Arduino UNO R3 connected to the TFT LCD display module. (d) Example of real-time patient footage and position labeling. (e) Schematic representation of the four standard CT simulation orientations: HFS, HFP, FFS, and FFP.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9284874/v1/4b5b7b881e71ad95612fc01e.png"},{"id":106836396,"identity":"83606c28-0043-428b-a708-803413524d22","added_by":"auto","created_at":"2026-04-14 02:08:01","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":192348,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance metrics of the two orientation-classification models. The Image Data Model was evaluated on the 80-image test set (20 images per orientation class: HFS, HFP, FFS, FFP), while the MLP model, relying on pose landmarks, was evaluated on a subset of 49 images where landmark extraction was successful. Performance metrics include balanced accuracy, precision, recall, and F1-score. The red bars represent the fine-tuned MobileNetV2 image-based model, and the blue bars represent the pose-based MLP model. The x-axis indicates percentage values.\u003c/p\u003e","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9284874/v1/ae1db460cea410fc93619908.png"},{"id":106836395,"identity":"7c36afd0-2917-44b4-9628-f1701d40d55f","added_by":"auto","created_at":"2026-04-14 02:08:01","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":98912,"visible":true,"origin":"","legend":"\u003cp\u003eConfusion matrices for the fine-tuned MobileNetV2 model (left) and the pose-based MLP model (right), evaluated on the same 80-image test set (20 images per orientation class). Each matrix shows the distribution of predicted labels versus true labels across the four classes: FFP, FFS, HFP, and HFS. Color intensity corresponds to the number of samples per cell.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9284874/v1/4fd0994789f76f394b1556b3.png"},{"id":107480715,"identity":"7c898a6b-d552-48d2-80a2-9fbc23a7a75a","added_by":"auto","created_at":"2026-04-22 02:13:16","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1464658,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9284874/v1/49b4b7ad-b7e2-4931-99a4-60ccb027c885.pdf"},{"id":106836392,"identity":"8e70c656-6e2d-411d-945a-f0d928db72fb","added_by":"auto","created_at":"2026-04-14 02:08:01","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":276545,"visible":true,"origin":"","legend":"","description":"","filename":"JKPSSupplementarymaterialJASE260401.docx","url":"https://assets-eu.researchsquare.com/files/rs-9284874/v1/bd4e7fcc8f5b14a1f560132e.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Real-Time Patient Orientation Verification System using Pose Estimation in CT Simulation","fulltext":[{"header":"I. INTRODUCTION","content":"\u003cp\u003eModern radiation therapy relies on precise geometric accuracy throughout the treatment workflow. The foundation of this precision begins with a computed tomography (CT) simulation, where patient spatial information is first established. However, a critical yet often overlooked failure mode occurs when therapists manually enter incorrect patient orientation at the CT console. This human error \u0026ndash; recording head-first-supine (HFS) as head-first-prone (HFP), for instance \u0026ndash; fundamentally corrupts the digital imaging and communications in medicine (DICOM) coordinate system by incorrectly defining spatial tags.\u003c/p\u003e\n\u003cp\u003eThe consequences of orientation mis-entry are severe. In Eclipse treatment planning system (TPS) (Varian Medical System, Palo Alto, CA), this error causes complete geometric misinterpretation: anterior-posterior axes invert, left-right orientations reverse, and the entire 3D coordinate system becomes corrupted. Treatment beams intended for tumors may instead target healthy tissue, rendering plans clinically useless and potentially catastrophic. Such errors have been documented to cause wrong-site treatments and near-miss events. These events, along with others such as incorrect protocol selection or scanning unintended anatomy, pose risks to patient safety, increase healthcare costs, and have potential legal and regulatory implications.\u003c/p\u003e\n\u003cp\u003eCurrent remediation strategies are reactive and resource intensive. Most institutions resort to patient recall for complete re-scanning, exposing patients to additional radiation, delaying treatment starts, and consuming valuable scanner time and resources. The alternative \u0026ndash; manual DICOM header correction \u0026ndash; requires specialized expertise and carries its own risks of introducing further errors. Our institution\u0026rsquo;s correction protocol (see supplementary material) illustrates the complexity: correcting HFS to HFP requires sequential patient position changes, image reorientation with 180\u0026deg; rotation, and new 3D dataset generation. For HFS to FFS errors, the process involves position correction, slice order inversion, and beta-axis rotation \u0026ndash; each step introducing potential for additional errors.\u003c/p\u003e\n\u003cp\u003eThe American Association of Physicists in Medicine (AAPM) Task Group 100 (TG\u0026ndash;100) [1] advocates for proactive, risk-based quality management that identifies and mitigates potential failures before they occur. From this perspective, orientation mis-entry represents a high-risk failure mode due to its catastrophic potential combined with low detectability using current procedures. The AAPM has a proven track record of proactively addressing patient safety concerns, such as the collaboration with Medical Imaging and Technology Alliance (MITA) that resulted in the dose alert feature now standard on all CT scanners. Despite clear risks, no proactive verification systems exist that utilize standard clinical hardware.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWe developed a tool named just another secondary evaluation (JASE) to fill this critical gap. By providing real-time, automated orientation verification before DICOM data finalization, JASE transforms a low-detectability/high-impact failure into a preventable error, aligning with modern patient safety principles.\u003c/p\u003e"},{"header":"II. MATERIALS AND METHODS","content":"\u003cp\u003e\u003cstrong\u003e1.\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eSystem architecture and dataset\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJASE consists of a webcam positioned to view the CT simulation couch, classification algorithm running on a standard computer, and an LCD display mounted adjacent to the CT console. The webcam captures continuous video of patients during positioning, which is processed in real-time to determine orientation before the therapist finalizes the DICOM header information. Figure 1 demonstrates the system architecture, showing the data flow from webcam capture through pose extraction and classification to the final LCD display, along with examples of the four standard patient orientations detected by the system.\u003c/p\u003e\n\u003cp\u003eTo develop and validate the classification algorithm, we created the JASE dataset by filming volunteers positioned on a CT simulation couch in each of the four standard orientations: HFS, HFP, FFS, and FFP. A total of eight healthy volunteers (4 female, 4 males; age range 16–35 years) were recruited for dataset creation. All participants were non-patients, and recordings were performed with volunteers fully clothed. The webcam was mounted inside the CT control room along the x-ray glass window, positioned approximately 1.5 meters above couch height and angled downward at roughly 35–40 degrees toward the patient to ensure full-body visibility during positioning. The videos were then converted to MP4 format and processed using open-source computer vision library (OpenCV) to extract individual frames. Every 10th frame was extracted to reduce redundancy while maintaining sufficient variation. This process yielded 1,539 original images distributed across the four orientation classes. The dataset was expanded through manual augmentation including vertical flipping, horizontal mirroring, and rotation to introduce greater positional variation. Additional fine-tuning data was collected using the same protocol and integrated into the dataset. For testing, we reserved 20 images per orientation class (10 from original data, 10 from fine-tuning data) creating an 80-image test set separate from training data. The final dataset contained 6,442 images organized into folders by orientation class, processed using Keras ImageDataGenerator with automatic assignment of class labels (FFP=0, FFS=1, HFP=2, HFS=3).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2.\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eModel development\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBoth models used in this project—MobileNetV2 and multilayer perceptron (MLP)—are forms of artificial neural networks (ANNs), computational systems inspired by the human brain [2]. ANNs consist of interconnected layers: input, hidden, and output layers [2]. For JASE, inputs are either raw image pixels (convolutional neural network (CNN) model) or body joint coordinates from pose estimation (MLP model). The output layer contains four neurons corresponding to the orientation classes (FFP, FFS, HFP, HFS). Models learn by adjusting weights and biases through backpropagation to minimize loss functions, enabling them to learn complex patterns required for accurate classification [3].\u003c/p\u003e\n\u003cp\u003eWe implemented two distinct approaches. The first utilized transfer learning with MobileNetV2, a lightweight CNN optimized for embedded applications. MobileNetV2 employs inverted residual blocks and linear bottlenecks that maintain spatial information while reducing computation through depthwise separable convolutions. Unlike traditional CNNs, its inverted residual blocks expand input with 1×1 convolution, apply depthwise convolution, then compress back to low dimensions—ideal for real-time deployment [4].\u003c/p\u003e\n\u003cp\u003eFor transfer learning, we froze the pretrained base layers and added custom classification layers consisting of a dropout layer with 30% rate, followed by a dense layer with 128 neurons using a Rectified Linear Unit (ReLU) activation, another dropout layer at 25%, and a final dense output layer with 4 neurons using softmax activation. The 128-unit dense layer was selected for sufficient representational capacity for the four-class orientation task without over-parameterizing the model relative to the size of our dataset. Freezing basic layers was necessary since the architecture for a basic MobileNetV2 is very extensive with millions of trainable parameters. If these layers were not frozen, training the model would increase the risk of overfitting, and require substantially more computational resources. Moreover, dropout layers at 25% randomly drop 25% of the neurons during training, which prevents overfitting and significantly increases the performance of a neural network by contributing to variance in the fine-tuned model [5]. To further inhibit overfitting, class weights were calculated so that the model can prioritize classes that are underrepresented in the dataset. \u0026nbsp;Subsequently, during fine-tuning we unfroze top layers and trained with learning rate 0.0001 to prevent catastrophic forgetting, using early stopping to monitor validation loss [6].\u003c/p\u003e\n\u003cp\u003eThe second approach used MediaPipe to extract 33 anatomical landmarks from each image, generating normalized x, y, z coordinates. These coordinates represent relative positions—for instance, (0.388, 0.303, -0.463) indicates 38.8% from left, 30.3% from top, with negative z showing camera proximity. This created 99–dimensional feature vectors capturing body configuration independent of visual appearance.\u003c/p\u003e\n\u003cp\u003eThe MLP followed a funnel architecture [7] with dense layers progressively decreasing from 512 to 256 to 128 to 64 units, incorporating batch normalization and a 25% dropout between layers. This configuration balanced model capacity and regularization relative to the low-dimensional 99–point pose input. Larger or wider architectures resulted in overfitting during development, whereas smaller networks underperformed due to insufficient representational capacity [8]. The final classification layer used softmax activation across four outputs. Both models employed class weight balancing to handle minor dataset imbalances and prevent overfitting.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3.\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eReal-time implementation and evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eReal-time deployment integrated video capture, model inference, and result display using OpenCV for image acquisition and preprocessing. The system captured continuous video from the USB webcam, extracting every 10th frame to balance processing load with responsiveness. Each frame underwent resizing and normalization before classification. Hardware display utilized an Arduino UNO R3 microcontroller connected to a 2-inch TFT LCD module (ST7789-based) via serial peripheral interface (SPI) interface. The Python classification script transmitted predicted orientation labels to the Arduino through USB serial communication at 9600 baud using PySerial. The LCD initially displayed \"READY\" and updated dynamically as predictions arrived, with text color indicating confidence levels: green for high confidence, yellow for moderate, and red for low confidence predictions.\u003c/p\u003e\n\u003cp\u003eModel evaluation employed standard classification metrics including accuracy, precision, recall, and F1-score calculated using scikit-learn. Confusion matrices visualized error patterns between orientation classes. Real-time performance assessment measured inference latency, end-to-end system delay, and operational stability over extended periods. Figure 2 illustrates the complete hardware setup including Arduino wiring configuration and example real-time orientation display.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4\u003c/strong\u003e\u003cstrong\u003e.\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eEvaluation metrics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eModel performance was evaluated using standard classification metrics: balanced accuracy, precision, recall, and F1-score, calculated using scikit-learn. These metrics provide a comprehensive assessment of classification performance, with F1-score and balanced accuracy offering a balanced measure particularly important for datasets with minor class imbalances.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e5\u003c/strong\u003e\u003cstrong\u003e.\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eEthical Considerations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll videos used in this study were captured on an iPhone as MP4 recordings during volunteer simulations conducted by members of the research team and colleagues from the same laboratory. The data did not originate from clinical scanners, hospital information systems, or DICOM-based workflows and contained no PHI beyond visible facial features. According to institutional policy, such non-clinical simulation activities are classified as not human subjects research and are exempt from IRB review and written consent requirements. All participants were informed of the project purpose and verbally agreed to the use of their images for research and publication.\u003c/p\u003e\n\u003cp\u003eData collection was limited to volunteer simulations rather than the clinical patients because RGB video capture may reveal identifiable facial features. This constrain influenced the scope of the present study, which focuses on proof-of-concept validation rather than clinical generalization.\u003c/p\u003e"},{"header":"III. RESULTS","content":"\u003cp\u003eThe fine-tuned transfer learning model achieved a balanced accuracy of 96.2%, a precision of 96.7%, a recall of 96.3%, and an F1-score of 96.3%. The MLP model had a better performance, as it achieved a balanced accuracy of 98.3%, a precision of 98.1%, a recall of 98.0%, and an F1-score of 98.0%. Based on the confusion matrix below, the fine-tuned transfer learning model misclassified the HFP class once and the FFP class twice. The MLP model only misclassified the FFP class as the HFP class once. These performances indicate that the MLP model outperforms the fine-tuned transfer learning model in all evaluated metrics.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eGiven the limited size of the hold-out test set (80 images for the image-based model and 49 landmark-valid images for the MLP model), these metrics should be interpreted as indicative of technical feasibility rather than statistically powered performance estimates. Results are visualized in Figure 3 (performance metrics comparison) and Figure 4 (confusion matrices for both models), where diagonal entries represent correct predictions and off-diagonal values indicate specific misclassification patterns between orientation classes.\u003c/p\u003e"},{"header":"IV. DISCUSSION","content":"\u003cp\u003eThe superiority of pose-based classification over raw image analysis (MLP 98.3% vs transfer learning 96.2% accuracy) reveals a fundamental insight: skeletal geometry provides more robust orientation markers than pixel-level features. While image-based models must contend with clinically irrelevant variables\u0026mdash;ambient lighting, patient clothing, blankets\u0026mdash;pose landmarks offer an invariant representation of patient configuration [9]. This robustness proves essential for deployment across diverse clinical environments where visual conditions vary unpredictably. JASE exemplifies the proactive quality management philosophy advocated by AAPM TG\u0026ndash;100 [1]. Traditional approaches detect orientation errors only after they propagate downstream, often during plan review or, worse, at treatment delivery [10]. By implementing targeted process control at the point of error origin\u0026mdash;the CT console\u0026mdash;JASE transforms a low-detectability/high-impact failure into a preventable event. The system enhances what TG\u0026ndash;100 terms \u0026quot;error detectability\u0026quot; without adding workflow complexity, directly reducing the risk priority number for this failure mode. It is important to note that the application of TG\u0026ndash;100 principles in this study is conceptual, as we did not measure quantitative changes in event frequency or workflow time. Such risk-reduction analysis requires prospective deployment in clinical CT simulation workflows, which is beyond the scope of this technical feasibility study but will be a focus of future works.\u003c/p\u003e\n\u003cp\u003eDevelopment challenges illuminated important technical considerations. The transfer learning model\u0026apos;s susceptibility to overfitting [11] and catastrophic forgetting [12] required careful regularization through early stopping, data augmentation, and learning rate optimization [13]. These challenges, while successfully addressed, underscore why the simpler pose-based approach ultimately proved superior\u0026mdash;it inherently avoids many pitfalls of image-based learning by operating on abstracted anatomical features rather than raw pixels.\u003c/p\u003e\n\u003cp\u003eThe present study did not include a systematic robustness evaluation across varying patient body habitus, clothing, or immobilization devices. Model evaluation was intentionally limited to common positioning conditions encountered within our acquisition environment, consistent with the feasibility-oriented scope of this work.\u003c/p\u003e\n\u003cp\u003eCurrent limitations center on MediaPipe\u0026apos;s landmark detection reliability. Some images lacked pose landmark visibility, causing MediaPipe to fail in extracting key points. This limitation is a result of the setup of the webcam. In our setup, the slightly elevated and oblique angle sometimes obscured extremities such as feet or hands, which led to incomplete landmark extraction and contributed to MediaPipe\u0026rsquo;s occasional detection failures. This finding highlights the significance of the camera positioning; it is crucial that major body landmarks are visible to the camera, or the MLP will have many inaccuracies and impair its performance. As a result, a number of images were excluded from both the training set and test set, with 2282 images excluded from the training set and 31 images excluded in the testing set. Because only landmark-complete samples could be used by the MLP architecture, the effective dataset represents only a subset of the full image collection. This introduces a selection bias, as cases with occluded extremities or partial body visibility are systematically underrepresented in the landmark-valid subset. As a result, performance comparisons between the MLP and MobileNetV2 models should be interpreted within the context of this evaluable subset rather than as a fully matched dataset-wide comparison. The final hold-out test set contained only 80 images (20 per class), which limits the statistical robustness of the reported metrics. The results should therefore be interpreted as feasibility-level findings rather than fully powered performance estimates. Cross-validation was considered; however, it was not appropriate for this dataset. First, as the MLP model could only be evaluated on images in which all 99 pose landmarks were successfully extracted, incomplete landmark sets prevented a consistent division of data across folds. In addition, ensuring evaluation consistency between the MobileNetV2 model\u0026mdash;which uses all images\u0026mdash;and the MLP model\u0026mdash;which uses only landmark-valid images\u0026mdash;would not be feasible under a cross-validation scheme. Inference-time measurements, confidence distribution analysis, or statistical comparison tests (e.g., McNemar\u0026rsquo;s test) were not included in the present analysis, as the limited and imbalanced landmark-valid subset prevents sufficiently powered paired statistical evaluation. These analyses will require a larger and more diverse dataset and are identified as important directions for future work. For these reasons, we adopted a strict hold-out test set, which provides a more conservative and methodologically appropriate estimate of model performance for the feasibility of this study. Nevertheless, the MLP model consistently outperformed the transfer learning model in classification performance and real-time deployment. Additionally, patients with atypical anatomy, amputations, or complete visual occlusion present edge cases requiring further development. The single-institution training dataset also limits generalizability claims. However, these limitations are addressable through expanded data collection and multi-input hybrid architectures [14] combining pose and image features. \u0026nbsp;Although the dataset included variability in camera angles and lighting both through natural acquisition differences and augmentation, the present study did not include a systematic robustness evaluation under controlled variations in illumination, viewpoint, or occlusion. It is important to note that this limitation primarily affects offline, frame-based dataset construction rather than the intended real-time deployment scenario. In practical use, the JASE system operates on continuous live video streams, where orientation predictions are generated across successive frames rather than relying on isolated still images. In this setting, transient landmark extraction failures in individual frames are mitigated by temporal redundancy, as subsequent frames typically yield valid pose landmarks. Orientation decisions can therefore be stabilized across multiple frames, reducing the impact of occasional landmark incompleteness on real-time system performance. Consequently, while landmark-incomplete frames were excluded during offline evaluation, this exclusion is less problematic in the envisioned real-time clinical workflow.\u003c/p\u003e\n\u003cp\u003eFurthermore, although all data were collected from healthy volunteers rather than clinical patients, the acquisition was performed inside an actual CT simulation room using the same equipment, couch setup, lighting, and physical environment encountered in routine practice. Volunteers were positioned following standard CT simulation procedures so that the captured videos closely mimicked real clinical presentations. Nevertheless, volunteer-based simulations cannot fully replicate the anatomical variability, immobilization devices, movement patterns, or workflow factors seen in real patient populations. As a result, formal validation using real patient data and prospective pilot testing with radiologic technologists will be required to establish true clinical deployment readiness.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eBeyond radiation oncology, other imaging environments also use standardized patient-orientation conventions (HFS, HFP, FFS, FFP), suggesting potential applicability of the JASE framework to modalities such as diagnostic CT, PET/CT, and MRI [15]. The core model relies on skeletal pose geometry extracted from RGB video rather than modality-specific image features, which conceptually enables adaptation if new training datasets are collected under modality-appropriate positioning conditions. For example, capturing MRI-specific positioning videos (with modality-specific couch designs and coil placements) would allow retraining of the pose-based MLP architecture.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eRecent advances in MRI safety protocols have focused on developing real time monitoring systems and improved patient positioning techniques. The introduction of advanced positioning aids and continuous monitoring technologies has significantly reduced the incidence of positioning related adverse events in MRI suites [16, 17]. These developments are particularly important given the increasing complexity of MRI examinations and the growing patient population with implanted devices or medical conditions that require special attention during positioning.\u003c/p\u003e\n\u003cp\u003eAlthough such modalities were not evaluated in this study, the concept may generalize pending the acquisition of modality-specific data and prospective validation. The path to clinical implementation therefore requires expanding training data to include edge cases, validating performance across multiple institutions, and exploring direct integration with imaging consoles. A prospective trial measuring error-prevention rates and workflow impact would definitively establish clinical value. Given that orientation errors can have high clinical consequence, even modest reductions could yield meaningful safety improvements across imaging departments.\u003c/p\u003e"},{"header":"V. CONCLUSION","content":"\u003cp\u003eJASE demonstrates that real-time orientation verification can effectively prevent a critical error mode in clinical imaging. The system\u0026apos;s 98.3% accuracy, low cost, and seamless integration demonstrate its potential as a practical quality assurance tool aligned with modern risk-based management principles. Deployment across imaging modalities could significantly reduce orientation-related errors throughout healthcare institutions.\u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003e \u003cb\u003eConflict of interest\u003c/b\u003e:\u003c/h2\u003e \u003cp\u003eThe authors declare no conflict of interest.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003e \u003cb\u003eEthics approval\u003c/b\u003e:\u003c/strong\u003e \u003cp\u003eNone.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003e \u003cb\u003eInformed consent\u003c/b\u003e:\u003c/strong\u003e \u003cp\u003eNone.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis study was supported by a grant from the National Foundation of Korea (NRF) funded by the Korean government (Grant No. RS-2024-00354252 \u0026amp; RS-2025-02303425).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAll authors contributed to the study conception and design. Material preparation, data collection, and analysis were performed by HAK, Elliott C, Estella C, and JJS. The first draft of the manuscript was written by HAK and JJS and all authors commented on previous versions of the initial manuscript internally. All authors read and approved the final submitted version of manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgement:\u003c/h2\u003e \u003cp\u003eNone.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eData will be made available on request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eM. S. Huq, B. A. Fraass, P. B. Dunscombe, J. P. Gibbons Jr., G. S. Ibbott, A. J. Mundt, S. Mutic, J. R. Palta, F. Rath, B. R. Thomadsen, J. F. Williamson and E. D. Yorke, Med. Phys. 43, 4209 (2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. H. Han, K. W. Kim, S. Y. Kim and Y. H. Youn, \u003cem\u003eDement. Neurocogn. Disord.\u003c/em\u003e 17, 83 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. El-Shahat, in \u003cem\u003eIntroductory Chapter: Artificial Neural Networks\u003c/em\u003e (InTech, 2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM. Sandler, A. Howard, M. Zhu, A. Zhmoginov and L. C. Chen, in \u003cem\u003eProc. IEEE Conf. Comput. Vis. Pattern Recognit.\u003c/em\u003e, 4510 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eN. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever and R. Salakhutdinov, \u003cem\u003eJ. Mach. Learn. Res.\u003c/em\u003e 15, 1929 (2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eC. Yang and X. Ma, in \u003cem\u003eProc. EMNLP\u003c/em\u003e, 4854 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. Klein, J. A. Raine, S. Pina-Otey, S. Voloshynovskiy and T. Golling, arXiv:2112.08069 (2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. Meyer-Baese and V. Schmid, \u003cem\u003eFoundations of Neural Networks\u003c/em\u003e (Elsevier, p. 197).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eS. S. Sengar, A. Kumar and O. Singh, arXiv:2406.15649 (2024).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eK. Bell, N. Lickt, C. R\u0026uuml;be and Y. Dzierma, \u003cem\u003eRadiat. Oncol.\u003c/em\u003e 13, 1 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eX. Ying, \u003cem\u003eJ. Phys. Conf. Ser.\u003c/em\u003e 1168, 022022 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska, D. Hassabis, C. Clopath, D. Kumaran and R. Hadsell, \u003cem\u003eProc. Natl. Acad. Sci. U.S.A.\u003c/em\u003e 114, 3521 (2017).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eP. Kenneweg, A. Schulz, S. Schr\u0026ouml;der and B. Hammer, \u003cem\u003eLect. Notes Comput. Sci.\u003c/em\u003e, 252 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eN. Menet, M. Hersche, G. Karunaratne, L. Benini, A. Sebastian and A. Rahimi, arXiv:2312.02829 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eH. K. Lim, Y. Choi, J. H. Jung, J. W. Jung, C. H. Oh, H. W. Park, et al., J. Korean Phys. Soc. 80, 640 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eV. Edelman, H. Chassidim and I. Rabaev, \u003cem\u003eElectronics\u003c/em\u003e 14, 668 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e] Y. J. Seol, J. N. Kim, A. R. Kim, J. H. Hwang, T. G. Oh, J. S. Shin, et al., J. Korean Phys. Sco. 73, 1012 (2018).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"journal-of-the-korean-physical-society","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Journal of the Korean Physical Society](https://link.springer.com/journal/40042)","snPcode":"40042","submissionUrl":"https://submission.springernature.com/new-submission/40042/3","title":"Journal of the Korean Physical Society","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Patient safety, CT simulation, Orientation verification, Pose estimation, Quality assurance","lastPublishedDoi":"10.21203/rs.3.rs-9284874/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9284874/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePatient orientation mis-entry during computed tomography (CT) simulation represents a critical failure mode that corrupts the digital imaging and communications in medicine (DICOM) coordinate system, necessitating patient recall for re-scanning or complex manual corrections. We developed a tool called just another secondary evaluation (JASE), an automated real-time verification system to prevent these errors. JASE uses a standard webcam to classify patient orientation into four standard categories: head-first-supine (HFS), head-first-prone (HFP), feet-first-supine (FFS), and feet-first-prone (FFP). We compared two approaches: transfer learning with MobileNetV2 and a multilayer perceptron (MLP) using MediaPipe pose landmarks. The system displays real-time orientation on an LCD screen adjacent to the CT console. The MLP model achieved 98.3% accuracy (1/49 misclassification) versus 96.2% for MobileNetV2 (3/80 misclassifications) on an independent test set. Real-time implementation demonstrated stable performance, which indicates the potential for future clinical deployment pending validation in actual CT simulation workflows. JASE provides a low-cost, effective solution for preventing orientation errors at the point of entry. The system\u0026rsquo;s applicability extends beyond radiation oncology to radiology and nuclear medicine departments using similar imaging protocols.\u003c/p\u003e","manuscriptTitle":"A Real-Time Patient Orientation Verification System using Pose Estimation in CT Simulation","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-14 02:07:57","doi":"10.21203/rs.3.rs-9284874/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-04-27T01:48:10+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-27T00:21:05+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-14T05:28:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"75176243010636262696252939272434495368","date":"2026-04-07T21:41:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"164733945243156627758702632175598413960","date":"2026-04-07T00:22:44+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-07T00:11:30+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-03T05:14:58+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-02T09:00:02+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of the Korean Physical Society","date":"2026-03-31T23:58:26+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"journal-of-the-korean-physical-society","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Journal of the Korean Physical Society](https://link.springer.com/journal/40042)","snPcode":"40042","submissionUrl":"https://submission.springernature.com/new-submission/40042/3","title":"Journal of the Korean Physical Society","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"e9b16658-705e-4239-aeb9-ecb2853d1759","owner":[],"postedDate":"April 14th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-13T04:41:18+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-14 02:07:57","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9284874","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9284874","identity":"rs-9284874","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}