Machine learning-based radiomics model for accurately predicting subtypes of neuroblastoma in children: A retrospective analysis

Machine learning-based radiomics model for accurately predicting subtypes of neuroblastoma in children: A retrospective analysis | 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 Machine learning-based radiomics model for accurately predicting subtypes of neuroblastoma in children: A retrospective analysis Yunyu He, Fan Sai Hou Alexandre, Weiqiang Xiao, Qinglin Yang, and 9 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7680804/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Objectives： This research seeks to create and validate radiomic models designed to enhance the diagnostic precision of neuroblastoma (NB), focusing specifically on distinguishing high-risk subtypes within its pathological classification. Methods: This retrospective study included 96 cases of NB, confirmed by histopathological evaluation. The cases were categorized according to the WHO classification into an aggressive group (n=55; neuroblastoma, Schwannian stroma-poor and ganglioneuroblastoma, nodular) and an indolent group (n=41; ganglioneuroma and ganglioneuroblastoma, intermixed). Radiomics features were extracted from CT images prior to biopsy or surgical resection. A radiomics model was constructed to predict NB classification using a random forest classifier. ROC curves were used to validate the capability of the models in the training and testing cohorts. Results: The final radiomics model incorporated 9 discriminative features. The model demonstrated strong diagnostic performance with an AUC of 0.874 in testing set, achieving a sensitivity of 77.5% and specificity of 88.6% in pathological classification. Decision curve analysis confirmed clinical utility across probability thresholds of 0.01-0.98 (training) and 0.01-0.81 (testing), indicating broad applicability for risk stratification. Conclusions： The developed radiomics model significantly improves diagnostic accuracy for neuroblastoma pathological classification, addressing a critical gap in current clinical practice. Unlike prior studies limited to binary NB diagnosis, this work provides granular discrimination between aggressive and indolent NB subtypes, enabling more precise risk stratification. While demonstrating robust performance, future multi-center validation is warranted to enhance generalizability and mitigate potential selection bias. This approach establishes a foundation for image-guided precision medicine in pediatric oncology. learning neuroblastoma radiomic model Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Neuroblastic tumors is a significant pediatric malignancy, accounting for approximately 8% of all childhood tumors [1]. This aggressive cancer primarily affects children under the age of five and is characterized by a notable mortality rate of around 15% [2], [3], which underscores the urgent need for improved diagnostic and treatment strategies. Neuroblastic tumors are classified into four major histological subtypes based on the degree of Schwannian stromal differentiation and biological behavior: neuroblastoma (NB, Schwannian stroma-poor), ganglioneuroblastoma intermixed (GNBi, Schwannian stroma-rich), ganglioneuroma (GN, Schwannian stroma-dominant), and ganglioneuroblastoma nodular (GNBn, a composite of stroma-rich/stroma-dominant and stroma-poor regions) [4]. These subtypes exhibit distinct clinical aggressiveness, with NB and GNBn demonstrating more aggressive behavior, including higher metastatic potential and poorer outcomes, whereas GN and GNBi typically follow an indolent course with favorable prognosis. Accurate subtyping is critical for risk stratification, treatment planning, and prognostic assessment, as it directly informs therapeutic decisions [5], [6]. Currently, the classification of neuroblastic tumor subtypes relies primarily on histopathological evaluation, which remains the gold standard for definitive diagnosis [7]. Nevertheless, this approach presents several inherent drawbacks, such as its invasive nature, the variability in sampling due to tumor heterogeneity, and the dependence on biopsy availability. These factors contribute to subjective diagnostic biases, reduced analytical efficiency, and, ultimately, delays in diagnosis or incomplete disease evaluation. Additionally, pathological analysis provides only a static snapshot of the tumor's morphology, lacking the ability to capture dynamic changes or spatial heterogeneity across the entire lesion. To address these challenges, recent advances in medical imaging offer a promising avenue for non-invasive classification [8] [9]. Radiomics entails the extraction of quantitative indicators, known as radiomic features, from various medical imaging modalities [10], enabling the capture of subtle patterns that may not be discernible through visual assessment alone. Multiple studies have illustrated the effectiveness of radiomics in the context of NB. For example, Wang et al. demonstrated that radiomic features could reliably predict progression-free survival in NB patients [11]. Additionally, Wu et al. employed radiomic features to identify key oncogene amplifications, such as MYCN, which are crucial for assessing tumor aggressiveness [12]. Furthermore, Koska et al. elucidated the ability of radiomics to predict tumor location and assist in the differential diagnosis of NB [13]. Hence, these findings underscore the transformative potential of radiomics in enhancing the precision of neuroblastoma management and addressing the limitations of traditional imaging modalities. This study aims to develop and validate a CT -based radiomics model for precise stratification of clinical aggressiveness in pediatric neuroblastic tumors. This investigation addresses the unmet need for robust quantitative biomarkers to guide risk-adapted treatment strategies in this clinically heterogeneous pediatric malignancy, and to provide a reference for choosing between surgical biopsy and primary resection. Methods Patient cohort This retrospective study was conducted at the Department of Thoracic Surgery, Guangzhou Women and Children's Medical Center, Guangzhou Medical University from January 2018 to August 2024. The inclusion criteria were as follows: (1) Complete and continuous radiomics data with adequate resolution and contrast; 2) Absence of significant motion artifacts or other interferences in the radiomic data; (3) Diagnosis of NB in the patient; (4) Tumor size greater than or equal to 5 mm with a well-defined lesion area. The exclusion criteria were: (1) CT image with low resolution, blurry images, or severe artifacts; (b) Incomplete CT image, missing relevant regions, or lack of continuous data across sections; (c) Irrelevant CT image to the study or the presence of incompatible image formats; (d) Tumor size smaller than 5 mm or poorly located lesion areas. Based on the criteria outlined above, a total of 96 patients diagnosed with NB were enrolled. The patient recruitment process is illustrated in Figure 1. The study protocol received ethical approval from the Institutional Review Board of Guangzhou Women and Children's Medical Center, with waiver of informed consent granted for this retrospective analysis. CT Imaging Acquisition This study utilized enhanced CT imaging to evaluate NB in the mediastinal region, with images obtained in DICOM format. Scanning was performed on a Philips CT Brilliance 64 scanner (Philips, Netherlands) using the following settings: a tube voltage of 120 kV, slice thickness and interval of 5 mm, rotation time of 0.5s, matrix size of 512 × 512, and a pitch of 0.9. A 300 mg/mL iodine contrast agent was administered intravenously, with the contrast scan completed within 17 seconds, followed by arterial and venous phase scans at 20-25 seconds and 55-65 seconds, respectively. Reconstructed images were transferred from the hospital's picture archiving and communication system to 3D Slicer (http://www.slicer.org) for radiomics analysis. Patients were positioned supine, and several preparatory measures were taken to ensure optimal image quality: fasting before the scan, sufficient rest to minimize motion artifacts, and breath-holding during the scan. Imaging was conducted during different phases of respiration to further reduce motion-related artifacts. To ensure consistency and reliability of the imaging data, the respiratory instructions were carefully followed, and contrast agents were used when necessary to improve image quality. No normalization or preprocessing of images was performed. ROI Segmentation The boundaries of the regions of interest (ROIs) were defined using routine CT sequences, specifically targeting the mediastinal region for delineation. The scope typically includes structures located behind the sternum, anterior to the spine, and between the pleural cavities to fully encompass the tumor lesion and its adjacent tissues. The ROIs were automatically segmented by two radiologists with more than 10 years of experience using 3D Slicer software. To ensure segmentation accuracy, any discrepancies between the radiologists' results were resolved by having one radiologist manually delineate the ROI, while the other independently re-delineated it after a one-week interval. The final ROI was determined through a discussion, comparing the two delineations. Neither of the radiologists had knowledge of the NB diagnosis.Figure 2. Supine Mediastinum CT images for NB patients, the ROIs is outlined with color green. Feature Engineering The radiomics features were extracted from each ROI. For feature extraction, Pyradiomics (http://pyradiomics.readthedocs.io/, Ver3.0.1) was used to obtain a total of 2417 features, which included 2382 texture-based features and 35 shape-based features. These were further categorized into different types: 468 First Order Features, 14 Shape Features, 572 Gray Level Co-occurrence Matrix (GLCM) Features, 416 Gray Level Run Length Matrix (GLRLM) Features, 416 Gray Level Size Zone Matrix (GLSZM) Features, 364 Gray Level Dependence Matrix (GLDM) Features, and 130 Neighboring Gray Tone Difference Matrix (NGTDM) Features. To select the most relevant features for analysis, LASSO regression was employed as the criterion for feature selection. In this study, a Random Forest (RF) algorithm was trained using data from 43 patients and tested on a separate set of 11 patients. The final model parameters included 10 estimators, no maximum depth, a minimum of 2 samples required to split a node, 1 sample required at a leaf node, automatic feature selection, bootstrapping enabled, and the Gini impurity criterion for node splitting. The model's performance was assessed using the ROC curve, with the Area Under the Curve (AUC) calculated to evaluate its ability to differentiate between outcomes, where a higher AUC reflects better performance. Statistical analysis In this study, Python (www.python.org/, Ver 3.11.9) was used for statistical analysis. Descriptive statistics were presented by expressing continuous variables as mean ± standard deviation, while categorical variables were summarized as frequencies and percentages. To compare groups, the ANOVA test was used for continuous variables, and the chi-squared test was applied for categorical variables to assess differences between groups. Statistical tests were conducted with p < 0.05 as an indicator of statistical significance. Results Clinical characteristics of the patients This study enrolled a total of 96 patients (mean age: 41.8 ± 39.9 months; male predominance: 55.2%, n = 53), stratified by pathological stage into two groups: the NB&GNBN group (n = 55, 57%) and the GNBI&GN group (n = 41, 43%)(Table 1). No significant sex-based disparity was observed between the groups (p = 0.235). Tumor characteristics revealed that 65.6% (n = 63) of lesions crossed the midline (p = 0.070), while calcification was present in 34.3% (n = 33; p = 0.124). Among the 73 patients tested for MYCN amplification, all results were negative. For model development, the cohort was partitioned into a training set (n = 76) and an independent testing set (n = 20). As detailed in Table 2, both sets demonstrated balanced distributions across all clinical variables (p > 0.05), ensuring unbiased analytical validity. Table 1. Basic characteristics of the study subjects. Table 2. Clinicopathological characteristics of patients in the training and testing cohorts Radiomics feature selection A total of 2,418 radiomics features were initially extracted from the imaging data using the PyRadiomics toolkit. These features comprised 468 first-order statistics, 14 shape-based descriptors, and 1,898 texture-based features, including Gray Level Co-occurrence Matrix (GLCM, n=572), Gray Level Run Length Matrix (GLRLM, n=416), Gray Level Size Zone Matrix (GLSZM, n=416), Gray Level Dependence Matrix (GLDM, n=364), and Neighbouring Gray Tone Difference Matrix (NGTDM, n=130) features. The features were normalized using Robust Scaler to minimize the influence of outliers as shown in Figure 3, 4. Feature selection was carried out in three distinct stages: Initially, features with zero variance were excluded, leaving a total of 2,417 features. This was followed by univariate analysis using a two-sample t-test, which identified 38 features with significant intergroup differences (P 0.95) were removed through redundancy reduction, retaining the feature with the lower P-value, resulting in 28 features. Finally, LASSO regression was applied to select the 9 most discriminative features (shown in Table 3), based on non-zero coefficients and model convergence. The radiomics score (Rad-score) was computed using a logistic regression model, where each feature’s value was weighted by its respective LASSO-derived coefficient. The Rad-score was defined as the linear combination: where wi denotes the coefficient of the i -th feature, xi is the normalized feature value, and b is the intercept. This approach optimizes predictive performance while minimizing overfitting (Figure 4). Figure 3. Original features and normalized features distribution histogram, the normalized method is Robust Scaler. Figure 4. PCA plot with four quadrants obtained before (A) and after (B) feature normalization and screening. Figure 5. The disparity in Rad-score distributions in the training (A) and testing (B) cohorts, and the results showed that there was a significant association between Rad-score and pathological classification of NB patients. Table 3. The representative features after filtering. Performance outcomes of the radiomics models The radiomics model demonstrated strong discriminatory performance in the training cohort, achieving an AUC of 0.888 (95% CI: [0.810–0.957]) in distinguishing NB&GNBN from GNBI&GN pathological classifications. This robustness was confirmed in the testing cohort, where the model attained an even higher AUC of 0.920 (95% CI: [0.758–0.100]), underscoring its generalizability (Figure 6). Subsequent ROC analysis revealed a sensitivity of 0.8857 and specificity of 0.7805, indicating a well-balanced diagnostic performance. The decision curve analysis (Figure 7) underscored the model's clinical relevance across a range of risk thresholds. In the training cohort, a net benefit was observed at thresholds between 0.01 and 0.98, supporting its applicability in diverse clinical settings. In the testing cohort, a net benefit was maintained within thresholds of 0.01 to 0.81, although its utility appeared somewhat more limited, likely due to cohort-specific differences. Figure 6. The ROC curve of the training and testing cohorts. Figure 7. The decision curve analysis for the model is presented. The grey line represents the scenario where the model predicts all patients without differentiating based on the pathological classification of NB. In contrast, the black curve reflects the assumption that the model predicts all patients while taking into account the pathological classification of NB. Discussion The dataset utilized in this study was divided into a training set and a testing set. The training set comprised 76 pediatric patients diagnosed with NB, while the testing set contained 20 patients. The final radiomics model was constructed by selecting 9 of the most relevant features out of an initial pool of 2,417 radiomics features. The logical basis for this process is the importance of features in predicting disease outcomes in pediatric patients with NB. The Random Forest algorithm was employed as the predictive modeling technique, chosen for its ability to handle high-dimensional data and its resilience to overfitting. The final radiomics model demonstrated excellent diagnostic performance in distinguishing between the four pathological classifications of NB, achieving an AUC of 0.920 in the independent testing cohort. The results of our study align with and expand upon previous research in the field of neuroblastoma (NB) diagnostics. For example, Feng L et al. [14] reported that PET/CT imaging provides more accurate identification of event-free survival (EFS) in NB, which supports our findings that imaging techniques are essential for identifying tumor characteristics. Although our model is based on radiomics features extracted from CT imaging modalities, its high AUC values in distinguishing between NB subtypes suggest that combining such radiomics approaches with advanced imaging technologies could further enhance diagnostic precision and prognostic assessment in pediatric NB patients. Furthermore, Wang H et al. [15] highlighted the potential of radiomics and deep learning advancements to enhance diagnostic capabilities. This aligns well with our study's use of a Random Forest-based radiomics model, reinforcing the potential of AI-driven approaches in the pediatric oncology domain. While traditional imaging techniques like CT play an important role in NB diagnosis, they can have limitations in certain cases. For instance, fibrosis and calcification can lead to reduced signal intensity on CT scans, potentially underestimating the severity of NB when CT is used in isolation [16]. This highlights a key advantage of our approach: by incorporating radiomics features, which capture subtle patterns in imaging data, we can potentially overcome some of the limitations inherent to individual imaging techniques, such as CT. Additionally, while there are several studies using imaging techniques for NB diagnosis, the use of radiomics models for identifying pathological types remains relatively underexplored. One such study examined CT-based radio-genomics for predicting MYCN amplification, a genetic alteration in NB patients, which is associated with reduced survival rates [17]. Our research goes a step further by focusing on radiomics modeling for predicting multiple pathological classifications in pediatric NB, thus extending the application of radiomics to a broader range of NB subtypes. Finally, our study differentiates itself by showing that radiomics is not limited to adult patients but can also be applied effectively to pediatric patients. This extension of radiomics applications in the pediatric NB context opens new avenues for future research and clinical practice, further emphasizing the potential of this approach to contribute to personalized treatment strategies for young patients with NB. Building upon the current findings, future research can focus on the following directions: (1) Future studies could explore the integration of genomics, transcriptomics, and proteomics data with radiomics features to develop a more comprehensive understanding of NB. Combining these data types may provide deeper insights into tumor biology, enabling better prediction models that account for genetic and molecular characteristics alongside imaging features. (2) With the rapid development of AI technologies, particularly deep learning methodologies, future research should focus on incorporating these advanced techniques into the radiomics model. Deep learning models, with their ability to automatically learn complex patterns from large datasets, could significantly improve the accuracy and robustness of predictions in the context of NB diagnosis, risk stratification, and treatment planning. In its current form, the model can aid in the early and accurate diagnosis of NB, allowing for timely intervention. It can also be used to assess the efficacy of clinical treatments by monitoring changes in the tumor’s radiomic features over time. Ultimately, this model can enhance clinical decision-making, leading to improved prognostic assessment and more targeted treatments for pediatric NB patients. The distinctiveness of this study lies in its focus on predicting both the pathological classification and risk stratification of pediatric NB patients using a radiomics model. By combining advanced imaging features with machine learning techniques, this model provides a more comprehensive and nuanced approach to understanding the disease. This innovative methodology could significantly improve clinical decision-making, allowing for more accurate and personalized treatment strategies for pediatric NB patients. This study has important limitations to acknowledge. The manual segmentation of ROIs, though performed by experienced radiologists, remains subjective and could affect reproducibility. Additionally, the single-center design and limited sample size may constrain the generalizability of our findings. Future research should explore AI-based automated segmentation to improve consistency, building on existing successes in medical imaging applications. Multicenter collaborations would also help validate these results across more diverse populations and reduce potential selection bias. The radiomics model developed in this study uses imaging data from NB, including tumor pathological characteristics, as input. By integrating a large volume of imaging data, the model identifies potential correlations between imaging features and pathological traits, thereby providing more accurate predictions compared to conventional diagnostic methods. This model offers a dependable approach for predicting the pathological classification of NB and enables more effective patient stratification, which is especially useful for doctors in making personalized treatment decisions. Moreover, the application of radiomics models in pediatric oncology holds the potential to further investigate the relationships between imaging features, genetic markers, and clinical outcomes, thus bridging the blank between radiology, genomics, and clinical practice. Declarations Author Contribution Dr Yunyu He and Dr Fan Sai Hou Alexandre conceptualized and designed the study, drafted the initial manuscript, and critically reviewed and revised the manuscript.Dr Weiqiang Xiao, Qinglin Yang, and Xinwen Zhang, designed the data collection instruments, collected data, carried out the initial analyses, and critically reviewed and revised the manuscript.Dr Jianhua Liang, Zefeng Lin, Dongmei Huang, Hongmei Wu, Jiachi Liao, and Haonan Wang designed the data collection instruments, collected data, carried out the initial analyses, and critically reviewed and revised the manuscript.Dr Le Li, Jiahang Zeng conceptualized and designed the study, coordinated and supervised data collection, and critically reviewed and revised the manuscript for important intellectual content.All authors approved the final manuscript as submitted and agree to be accountable for all aspects of the work. References Zafar A, Wang W, Liu G, Wang X, Xian W, McKeon F et al (2021) Molecular targeting therapies for neuroblastoma: Progress and challenges. Med Res Rev 41(2):961–1021 Valter K, Zhivotovsky B, Gogvadze V (2018) Cell death-based treatment of neuroblastoma. Cell Death Dis 9(2):113 Ward E, DeSantis C, Robbins A, Kohler B, Jemal A (2014) Childhood and adolescent cancer statistics, 2014. CA Cancer J Clin 64(2):83–103 He WG, Yan Y, Tang W, Cai R, Ren G (2017) Clinical and biological features of neuroblastic tumors: A comparison of neuroblastoma and ganglioneuroblastoma. Oncotarget 8(23):37730–37739 Yoneda A (2023) Role of surgery in neuroblastoma. Pediatr Surg Int 39(1):177 Qiu B, Matthay KK (2022) Advancing therapy for neuroblastoma. Nat Rev Clin Oncol 19(8):515–533 Shimada H (2003) The International Neuroblastoma Pathology Classification. Pathologica 95(5):240–241 Zhang X, Li C, Xu C, Hao X, Yu X, Li Q (2018) Correlation of CT signs with lymphatic metastasis and pathology of neuroblastoma in children. Oncol Lett 16(2):2439–2443 Littooij AS, de Keizer B (2023) Imaging in neuroblastoma. Pediatr Radiol 53(4):783–787 Lambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J et al (2017) Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 14(12):749–762 Wang H, Li T, Xie M, Si J, Qin J, Yang Y et al (2023) Association of Computed Tomography Radiomics Signature with Progression-free Survival in Neuroblastoma Patients. Clin Oncol (R Coll Radiol) 35(11):e639–e47 Wu H, Wu C, Zheng H, Wang L, Guan W, Duan S et al (2021) Radiogenomics of neuroblastoma in pediatric patients: CT-based radiomics signature in predicting MYCN amplification. Eur Radiol 31(5):3080–3089 Koska IO, Ozcan HN, Tan AA, Beydogan B, Ozer G, Oguz B et al (2024) Radiomics in differential diagnosis of Wilms tumor and neuroblastoma with adrenal location in children. Eur Radiol 34(8):5016–5027 Sepehri A, Stockton DJ, Roffey DM, Lefaivre KA, Potter JM, Guy P (2024) Effect of humeral rotation on the reliability of radiographic measurements for proximal humerus fractures. J Orthop Sci 29(4):1078–1084 Wang H, Chen X, He L (2023) A narrative review of radiomics and deep learning advances in neuroblastoma: updates and challenges. Pediatr Radiol 53(13):2742–2755 Ladenstein R, Lambert B, Pötschger U, Castellani MR, Lewington V, Bar-Sever Z et al (2018) Validation of the mIBG skeletal SIOPEN scoring method in two independent high-risk neuroblastoma populations: the SIOPEN/HR-NBL1 and COG-A3973 trials. Eur J Nucl Med Mol Imaging 45(2):292–305 García N, Bermúdez A, Martín M, Carmona C, Jaén C, Daroca T (2020) [Analysis of the evolution of those operated on with minimal access surgery in our hospital. Does it present better results than the conventional one?]. Arch Cardiol Mex 91(3):321–326 Tables Tables 1 to 3 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1.pptx Table 1. Basic characteristics of the study subjects. *Chi-square test; p<0.05 indicates statistical significance. Table2.pptx Table 2. Clinicopathological characteristics of patients in the training and testing cohorts Table3.docx Table 3. The representative features after filtering. Cite Share Download PDF Status: Posted Version 1 posted 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-7680804","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":538525195,"identity":"5b71d60f-c07b-4978-8842-e60eb7d18d6b","order_by":0,"name":"Yunyu He","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yunyu","middleName":"","lastName":"He","suffix":""},{"id":538525196,"identity":"c56df970-8628-46c8-add8-569c524099d3","order_by":1,"name":"Fan Sai Hou Alexandre","email":"","orcid":"","institution":"Centro Hospitalar Conde de São Januário, Macau SAR of the people's","correspondingAuthor":false,"prefix":"","firstName":"Fan","middleName":"Sai Hou","lastName":"Alexandre","suffix":""},{"id":538525197,"identity":"7fa0209b-9f18-4a9d-a097-9aba8b1cd633","order_by":2,"name":"Weiqiang Xiao","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Weiqiang","middleName":"","lastName":"Xiao","suffix":""},{"id":538525198,"identity":"d303cc0c-5189-4337-9c79-c4d27453b0cb","order_by":3,"name":"Qinglin Yang","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Qinglin","middleName":"","lastName":"Yang","suffix":""},{"id":538525199,"identity":"aa96d9fc-591a-415f-a815-f222acc288a1","order_by":4,"name":"Xinwen Zhang","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Xinwen","middleName":"","lastName":"Zhang","suffix":""},{"id":538525200,"identity":"13236727-9925-4a60-895c-a7c6f2c86a65","order_by":5,"name":"Jianhua Liang","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Jianhua","middleName":"","lastName":"Liang","suffix":""},{"id":538525201,"identity":"4342635c-045d-47c7-b88a-66b27c537e0b","order_by":6,"name":"Zefeng Lin","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zefeng","middleName":"","lastName":"Lin","suffix":""},{"id":538525202,"identity":"75b53f5c-a73a-4a1b-892f-38c54be020a0","order_by":7,"name":"Dongmei Huang","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Dongmei","middleName":"","lastName":"Huang","suffix":""},{"id":538525203,"identity":"6d7c5fb3-8d00-4074-9162-ff1d214652be","order_by":8,"name":"Hongmei Wu","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Hongmei","middleName":"","lastName":"Wu","suffix":""},{"id":538525204,"identity":"d92b037d-d8b4-4b09-afbe-8acd3452258a","order_by":9,"name":"Jiachi Liao","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Jiachi","middleName":"","lastName":"Liao","suffix":""},{"id":538525205,"identity":"79d5d798-3a58-474d-8e7c-fbec7b1ba338","order_by":10,"name":"Haonan Wang","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Haonan","middleName":"","lastName":"Wang","suffix":""},{"id":538525206,"identity":"2a6196b5-3ddc-453f-ba39-384be2c522b6","order_by":11,"name":"Le Li","email":"","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Le","middleName":"","lastName":"Li","suffix":""},{"id":538525207,"identity":"b35dfc4e-e320-4188-962c-beedd5638a2a","order_by":12,"name":"Jiahang Zeng","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA90lEQVRIiWNgGAWjYBACxmYGNjj7MZhiZm4gWguzMQODAZBixK8FCOBa2KTBWhgIaGFuZ3724OeOWnt+6fZr1QUVf6L524FaflRsw+MwNnPD3jPHmSXnnCm7PeOMQe6Mw4wNjD1nbuPzi5kEb9sxNoMbOWm3edsMchuAWpgZ2/BpYf8m+bftGI89UEsxSMt8wlp4zKR522okDCTSjzGDtGwgQkuZtGzbAQOJGznM0jxnjHM3ArUcxOcXw/7j2yTfttXZ889If/iZp0Iud975wwcf/KjAo6UBTB0GYh4DuOgBnOqBQB5C1QEx+wN8CkfBKBgFo2AEAwD9qVWYC+3KfwAAAABJRU5ErkJggg==","orcid":"","institution":"Guangzhou Women and Children’s Medical Center, Guangzhou Medical University","correspondingAuthor":true,"prefix":"","firstName":"Jiahang","middleName":"","lastName":"Zeng","suffix":""}],"badges":[],"createdAt":"2025-09-22 15:40:00","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7680804/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7680804/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":95524695,"identity":"80be2d44-90a6-4998-bf97-a744b91794dd","added_by":"auto","created_at":"2025-11-10 10:03:14","extension":"tif","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4292484,"visible":true,"origin":"","legend":"","description":"","filename":"Fig1.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/f707dba9669359b9785ffe47.tif"},{"id":95357900,"identity":"b3d44874-55cd-4219-a4d2-83b330ecd432","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":80860,"visible":true,"origin":"","legend":"","description":"","filename":"MainDocument.docx","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/c4ed18076b27aff82aa7de5b.docx"},{"id":95525558,"identity":"452d41c0-ea9a-4df9-a96f-8fe3f04a4b54","added_by":"auto","created_at":"2025-11-10 10:05:17","extension":"tif","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4640744,"visible":true,"origin":"","legend":"","description":"","filename":"Fig2.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/2d54d973c054993068d891cc.tif"},{"id":95357910,"identity":"a4628979-7e64-49c0-b519-9ca7173f3187","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"tif","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6187924,"visible":true,"origin":"","legend":"","description":"","filename":"Fig3.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/e8a556fa3b5a7ac7787eb52e.tif"},{"id":95524487,"identity":"625ce75a-be7c-401d-b65b-98ded52cc680","added_by":"auto","created_at":"2025-11-10 10:02:49","extension":"tif","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2147684,"visible":true,"origin":"","legend":"","description":"","filename":"Fig4.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/aa83b8bbbd05ee4f165e1fc5.tif"},{"id":95357908,"identity":"be59c18f-c580-4d05-8935-0f3da8d48d29","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"tif","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1950732,"visible":true,"origin":"","legend":"","description":"","filename":"Fig5.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/79a6510ffdca1f7336d843c0.tif"},{"id":95525839,"identity":"33b0fa54-0b5b-404c-b79c-8c7905e19c98","added_by":"auto","created_at":"2025-11-10 10:05:43","extension":"docx","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":17004,"visible":true,"origin":"","legend":"","description":"","filename":"Table3.docx","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/d9bdfeb47ec47468e623e019.docx"},{"id":95525432,"identity":"3e4f7a87-bda8-4629-8ea4-2c11e847a636","added_by":"auto","created_at":"2025-11-10 10:05:00","extension":"tif","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2599980,"visible":true,"origin":"","legend":"","description":"","filename":"Fig6.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/af5e7d44516415f7dbfd8378.tif"},{"id":95357917,"identity":"31c1474d-ac20-4ee7-88f3-e407f6bb91b8","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"tif","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1918320,"visible":true,"origin":"","legend":"","description":"","filename":"Fig7.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/a7fb593acd48eddf345d193e.tif"},{"id":95524460,"identity":"6be6fc7b-291c-4c45-8757-3db9175110a5","added_by":"auto","created_at":"2025-11-10 10:02:47","extension":"json","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":12353,"visible":true,"origin":"","legend":"","description":"","filename":"a488a69cf224482b91f29c664879e9ce.json","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/786606fd0bf18c08f629763d.json"},{"id":95357918,"identity":"1fb56d2c-e0d6-4543-a5f1-dae2d51f5b11","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"xml","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":71243,"visible":true,"origin":"","legend":"","description":"","filename":"a488a69cf224482b91f29c664879e9ce1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/de17f1c911b3daddf631458f.xml"},{"id":95524700,"identity":"7da58742-b4b1-4735-b592-683f55164adb","added_by":"auto","created_at":"2025-11-10 10:03:14","extension":"tif","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4292484,"visible":true,"origin":"","legend":"","description":"","filename":"Fig1.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/187b0725ad4274f8ca3c60fc.tif"},{"id":95357928,"identity":"5056b784-e931-4cdc-a98c-307cf0427c73","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"tif","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":4640744,"visible":true,"origin":"","legend":"","description":"","filename":"Fig2.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/5ca3689e780102be97323efe.tif"},{"id":95524872,"identity":"5b95a06c-6f2d-46ff-a3db-5ac7f857ecfc","added_by":"auto","created_at":"2025-11-10 10:03:42","extension":"tif","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6187924,"visible":true,"origin":"","legend":"","description":"","filename":"Fig3.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/8e9545cf4be3d64be7319711.tif"},{"id":95526042,"identity":"4555721a-7915-4806-a014-4d2c1d54f5ad","added_by":"auto","created_at":"2025-11-10 10:06:08","extension":"tif","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2147684,"visible":true,"origin":"","legend":"","description":"","filename":"Fig4.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/32fc6821b7407878bcfff734.tif"},{"id":95357930,"identity":"97d3e101-e075-4663-96b2-0dd69996d618","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"tif","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1950732,"visible":true,"origin":"","legend":"","description":"","filename":"Fig5.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/1e2bbc75ae8aeb3864ea04d3.tif"},{"id":95357925,"identity":"71df4564-72f7-435d-8680-d9ec5155cfd4","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"tif","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2599980,"visible":true,"origin":"","legend":"","description":"","filename":"Fig6.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/7c67b0e9947c7d846a712171.tif"},{"id":95525555,"identity":"7e3ec832-3391-43a4-bf74-27f4d0d07fd5","added_by":"auto","created_at":"2025-11-10 10:05:17","extension":"tif","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1918320,"visible":true,"origin":"","legend":"","description":"","filename":"Fig7.tif","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/83dbcb5f9f0a8c0746505b7f.tif"},{"id":95357912,"identity":"6a56371a-91cb-4666-845e-04ea506e8c36","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6129,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/b3a5b8fc06c9371cf66c6c7a.png"},{"id":95524848,"identity":"45a030bb-6caa-4c05-8c41-8a248e5a68ae","added_by":"auto","created_at":"2025-11-10 10:03:39","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":501506,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig1.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/245c796b9a7657d482fa2eea.png"},{"id":95357915,"identity":"92c40606-509b-496b-a737-6c475b2f3cae","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2132418,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig2.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/d7e8657a24ccf77ebc985024.png"},{"id":95357916,"identity":"f5f4b4ed-c980-486e-a586-812c50526d5b","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1275921,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig3.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/2fcca7c1ea52a08fff437442.png"},{"id":95525747,"identity":"edc36b71-83b1-46a3-a3ae-49f529cfa64c","added_by":"auto","created_at":"2025-11-10 10:05:38","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":382375,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig4.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/7f76a6f2740a95fef9e47755.png"},{"id":95357920,"identity":"0a4df775-e408-4600-9734-d570cc789c7e","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":220245,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig5.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/625dfc0d9e04bfc90b443512.png"},{"id":95357931,"identity":"41a10f1d-3928-4189-8572-e2bf167dd901","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":543554,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig6.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/532be85b4b2f1d2e8c6b0bfc.png"},{"id":95357923,"identity":"170fa582-a7c7-4729-bc5c-a2ca648b9cbb","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":352568,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFig7.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/a0243e7abef96062d7f99690.png"},{"id":95357933,"identity":"ab92d860-15fc-4101-aeb6-7f1601724484","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3011,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/2f505b2a320e3386c324fc4e.png"},{"id":95525898,"identity":"35df11b0-2369-44fb-a1de-10a67e975a98","added_by":"auto","created_at":"2025-11-10 10:05:49","extension":"xml","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":67911,"visible":true,"origin":"","legend":"","description":"","filename":"a488a69cf224482b91f29c664879e9ce1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/2639b70925f33067fcb07e23.xml"},{"id":95357921,"identity":"a3aaa834-6454-403c-8b4e-e8b919335a6b","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"html","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":79953,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/a87c433b6c6ea3aa42435744.html"},{"id":95525401,"identity":"8c2507ee-3480-42f3-9d71-c0eb9aaa445f","added_by":"auto","created_at":"2025-11-10 10:04:58","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":3755835,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of patient selection.\u003c/p\u003e","description":"","filename":"Fig1.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/7180e7e6796c06c21455b198.jpg"},{"id":95357901,"identity":"ec27e17b-1a86-4bed-ab75-909fe418c040","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1132226,"visible":true,"origin":"","legend":"\u003cp\u003eSupine Mediastinum CT images for NB patients, the ROIs is outlined with color green. (Group 1: NB; Group 2: GNBI; Group 3: GN; Group 4: GNBN.)\u003c/p\u003e\n\u003cp\u003eA. Mask Group 1 CT images; B. Mask Group 2 CT images; C. Mask Group 3 CT images; D. Mask Group 4 CT images;\u003c/p\u003e\n\u003cp\u003eE. Unmask Group 1 CT images; F. Unmask Group 2 CT images; G. Unmask Group 3 CT images; H. Unmask Group 4 CT images.\u003c/p\u003e","description":"","filename":"Fig2.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/c03279296f1fe0c854d4deae.jpg"},{"id":95357914,"identity":"13cbea22-7b07-44e5-8e7d-d9bc6bf1a499","added_by":"auto","created_at":"2025-11-07 07:03:29","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2551265,"visible":true,"origin":"","legend":"\u003cp\u003eOriginal features and normalized features distribution histogram, the normalized method is Robust Scaler.\u003c/p\u003e","description":"","filename":"Fig3.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/ce519fa21a35ed4def93c0a5.jpg"},{"id":95357906,"identity":"66f3b85a-2c6a-4e80-bb5d-49a0f1f7589c","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":924314,"visible":true,"origin":"","legend":"\u003cp\u003ePCA plot with four quadrants obtained before (A) and after (B) feature normalization and screening.\u003c/p\u003e","description":"","filename":"Fig4.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/174b3b79513370ee40b0cc54.jpg"},{"id":95357903,"identity":"68eb0b8a-d6db-40fa-b987-81957a5b1881","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":890077,"visible":true,"origin":"","legend":"\u003cp\u003eThe disparity in Rad-score distributions in the training (A) and testing (B) cohorts, and the results showed that there was a significant association between Rad-score and pathological classification of NB patients.\u003c/p\u003e","description":"","filename":"Fig5.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/b802ed446f495df7117a382b.jpg"},{"id":95525670,"identity":"e08a88be-2a5c-4710-89ab-07ac87a04b49","added_by":"auto","created_at":"2025-11-10 10:05:32","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1078131,"visible":true,"origin":"","legend":"\u003cp\u003eThe ROC curve of the training and testing cohorts.\u003c/p\u003e","description":"","filename":"Fig6.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/3731ec18603e4e6b6841bd0c.jpg"},{"id":95526051,"identity":"e185f40f-9585-4194-bb1e-e58635832b01","added_by":"auto","created_at":"2025-11-10 10:06:09","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":816865,"visible":true,"origin":"","legend":"\u003cp\u003eThe decision curve analysis for the model is presented. The grey line represents the scenario where the model predicts all patients without differentiating based on the pathological classification of NB. In contrast, the black curve reflects the assumption that the model predicts all patients while taking into account the pathological classification of NB.\u003c/p\u003e","description":"","filename":"Fig7.tif.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/5201971dfde7eeb660a6ff76.jpg"},{"id":96315211,"identity":"a595c63c-64a9-4e28-b241-e07cfd1fad4b","added_by":"auto","created_at":"2025-11-19 17:23:39","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":11648820,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/73158b78-3740-4def-b2b7-48984480fd4d.pdf"},{"id":95357895,"identity":"4cd7cc17-8301-4182-a33d-69bafd4b1f1f","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"pptx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":42923,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTable 1. \u003c/strong\u003eBasic characteristics of the study subjects. *Chi-square test; p\u0026lt;0.05 indicates statistical significance.\u003c/p\u003e","description":"","filename":"Table1.pptx","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/2cb8242a5e11e9d9837c5619.pptx"},{"id":95357897,"identity":"c3c97ea9-c40d-414a-8d87-11252db8bfc3","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"pptx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":42647,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTable 2. \u003c/strong\u003eClinicopathological characteristics of patients in the training and testing cohorts\u003c/p\u003e","description":"","filename":"Table2.pptx","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/6bef137168a4db88e92faeef.pptx"},{"id":95357898,"identity":"ad9a4746-7df5-4a9e-9b7c-0e8aab5a58a1","added_by":"auto","created_at":"2025-11-07 07:03:28","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":17004,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTable 3.\u003c/strong\u003e The representative features after filtering.\u003c/p\u003e","description":"","filename":"Table3.docx","url":"https://assets-eu.researchsquare.com/files/rs-7680804/v1/2e2322020e95f6426bc66b66.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Machine learning-based radiomics model for accurately predicting subtypes of neuroblastoma in children: A retrospective analysis","fulltext":[{"header":"Introduction ","content":"\u003cp\u003eNeuroblastic tumors is a significant pediatric malignancy, accounting for approximately 8% of all childhood tumors [1]. This aggressive cancer primarily affects children under the age of five and is characterized by a notable mortality rate of around 15% [2], [3], which underscores the urgent need for improved diagnostic and treatment strategies. Neuroblastic tumors are classified into four major histological subtypes based on the degree of Schwannian stromal differentiation and biological behavior: neuroblastoma (NB, Schwannian stroma-poor), ganglioneuroblastoma intermixed (GNBi, Schwannian stroma-rich), ganglioneuroma (GN, Schwannian stroma-dominant), and ganglioneuroblastoma nodular (GNBn, a composite of stroma-rich/stroma-dominant and stroma-poor regions) [4]. These subtypes exhibit distinct clinical aggressiveness, with NB and GNBn demonstrating more aggressive behavior, including higher metastatic potential and poorer outcomes, whereas GN and GNBi typically follow an indolent course with favorable prognosis. Accurate subtyping is critical for risk stratification, treatment planning, and prognostic assessment, as it directly informs therapeutic decisions [5], [6].\u003c/p\u003e\n\u003cp\u003eCurrently, the classification of neuroblastic tumor subtypes relies primarily on histopathological evaluation, which remains the gold standard for definitive diagnosis [7]. Nevertheless, this approach presents several inherent drawbacks, such as its invasive nature, the variability in sampling due to tumor heterogeneity, and the dependence on biopsy availability. These factors contribute to subjective diagnostic biases, reduced analytical efficiency, and, ultimately, delays in diagnosis or incomplete disease evaluation. Additionally, pathological analysis provides only a static snapshot of the tumor\u0026apos;s morphology, lacking the ability to capture dynamic changes or spatial heterogeneity across the entire lesion. To address these challenges, recent advances in medical imaging offer a promising avenue for non-invasive classification [8] [9].\u003c/p\u003e\n\u003cp\u003eRadiomics entails the extraction of quantitative indicators, known as radiomic features, from various medical imaging modalities [10], enabling the capture of subtle patterns that may not be discernible through visual assessment alone. Multiple studies have illustrated the effectiveness of radiomics in the context of NB. For example, Wang et al. demonstrated that radiomic features could reliably predict progression-free survival in NB patients [11]. Additionally, Wu et al. employed radiomic features to identify key oncogene amplifications, such as MYCN, which are crucial for assessing tumor aggressiveness [12]. Furthermore, Koska et al. elucidated the ability of radiomics to predict tumor location and assist in the differential diagnosis of NB [13]. Hence, these findings underscore the transformative potential of radiomics in enhancing the precision of neuroblastoma management and addressing the limitations of traditional imaging modalities.\u003c/p\u003e\n\u003cp\u003eThis study aims to develop and validate a CT -based radiomics model for precise stratification of clinical aggressiveness in pediatric neuroblastic tumors. This investigation addresses the unmet need for robust quantitative biomarkers to guide risk-adapted treatment strategies in this clinically heterogeneous pediatric malignancy, and to provide a reference for choosing between surgical biopsy and primary resection.\u003c/p\u003e"},{"header":"Methods ","content":"\u003cp\u003e\u003cstrong\u003ePatient cohort\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis retrospective study was conducted at the Department of Thoracic Surgery, Guangzhou Women and Children\u0026apos;s Medical Center, Guangzhou Medical University from January 2018 to August 2024.\u0026nbsp;The inclusion criteria were as follows: (1) Complete and continuous radiomics data with adequate resolution and contrast; 2) Absence of significant motion artifacts or other interferences in the radiomic data; (3) Diagnosis of NB in the patient; (4) Tumor size greater than or equal to 5 mm with a well-defined lesion area. The exclusion criteria were: (1) CT image with low resolution, blurry images, or severe artifacts; (b) Incomplete CT image, missing relevant regions, or lack of continuous data across sections; (c) Irrelevant CT image to the study or the presence of incompatible image formats; (d) Tumor size smaller than 5 mm or poorly located lesion areas.\u0026nbsp;Based on the criteria outlined above, a total of 96 patients diagnosed with NB were enrolled. The patient recruitment process is illustrated in Figure 1.\u003c/p\u003e\n\u003cp\u003eThe study protocol received ethical approval from the Institutional Review Board of Guangzhou Women and Children\u0026apos;s Medical Center, with waiver of informed consent granted for this retrospective analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCT Imaging Acquisition\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study utilized enhanced CT imaging to evaluate NB in the mediastinal region, with images obtained in DICOM format. Scanning was performed on a Philips CT Brilliance 64 scanner (Philips, Netherlands) using the following settings: a tube voltage of 120 kV, slice thickness and interval of 5 mm, rotation time of 0.5s, matrix size of 512 \u0026times; 512, and a pitch of 0.9. A 300 mg/mL iodine contrast agent was administered intravenously, with the contrast scan completed within 17 seconds, followed by arterial and venous phase scans at 20-25 seconds and 55-65 seconds, respectively. Reconstructed images were transferred from the hospital\u0026apos;s picture archiving and communication system to 3D Slicer (http://www.slicer.org) for radiomics analysis.\u003c/p\u003e\n\u003cp\u003ePatients were positioned supine, and several preparatory measures were taken to ensure optimal image quality: fasting before the scan, sufficient rest to minimize motion artifacts, and breath-holding during the scan. Imaging was conducted during different phases of respiration to further reduce motion-related artifacts. To ensure consistency and reliability of the imaging data, the respiratory instructions were carefully followed, and contrast agents were used when necessary to improve image quality. No normalization or preprocessing of images was performed.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eROI Segmentation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe boundaries of the regions of interest (ROIs) were defined using routine CT sequences, specifically targeting the mediastinal region for delineation. The scope typically includes structures located behind the sternum, anterior to the spine, and between the pleural cavities to fully encompass the tumor lesion and its adjacent tissues. The ROIs were automatically segmented by two radiologists with more than 10 years of experience using 3D Slicer software. To ensure segmentation accuracy, any discrepancies between the radiologists\u0026apos; results were resolved by having one radiologist manually delineate the ROI, while the other independently re-delineated it after a one-week interval. The final ROI was determined through a discussion, comparing the two delineations. Neither of the radiologists had knowledge of the NB diagnosis.Figure 2. Supine Mediastinum CT images for NB patients, the ROIs is outlined with color green.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFeature Engineering\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe radiomics features were extracted from each ROI. For feature extraction, Pyradiomics (http://pyradiomics.readthedocs.io/, Ver3.0.1) was used to obtain a total of 2417 features, which included 2382 texture-based features and 35 shape-based features. These were further categorized into different types: 468 First Order Features, 14 Shape Features, 572 Gray Level Co-occurrence Matrix (GLCM) Features, 416 Gray Level Run Length Matrix (GLRLM) Features, 416 Gray Level Size Zone Matrix (GLSZM) Features, 364 Gray Level Dependence Matrix (GLDM) Features, and 130 Neighboring Gray Tone Difference Matrix (NGTDM) Features. To select the most relevant features for analysis, LASSO regression was employed as the criterion for feature selection.\u003c/p\u003e\n\u003cp\u003eIn this study, a Random Forest (RF) algorithm was trained using data from 43 patients and tested on a separate set of 11 patients. The final model parameters included 10 estimators, no maximum depth, a minimum of 2 samples required to split a node, 1 sample required at a leaf node, automatic feature selection, bootstrapping enabled, and the Gini impurity criterion for node splitting. The model\u0026apos;s performance was assessed using the ROC curve, with the Area Under the Curve (AUC) calculated to evaluate its ability to differentiate between outcomes, where a higher AUC reflects better performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn this study, Python (www.python.org/, Ver 3.11.9) was used for statistical analysis. Descriptive statistics were presented by expressing continuous variables as mean \u0026plusmn; standard deviation, while categorical variables were summarized as frequencies and percentages. To compare groups, the ANOVA test was used for continuous variables, and the chi-squared test was applied for categorical variables to assess differences between groups. Statistical tests were conducted with p \u0026lt; 0.05 as an indicator of statistical significance.\u0026nbsp;\u003c/p\u003e"},{"header":"Results ","content":"\u003cp\u003e\u003cstrong\u003eClinical characteristics of the patients\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study enrolled a total of 96 patients (mean age: 41.8 \u0026plusmn; 39.9 months; male predominance: 55.2%,\u0026nbsp;n\u0026nbsp;= 53), stratified by pathological stage into two groups: the NB\u0026amp;GNBN group (n\u0026nbsp;= 55, 57%) and the GNBI\u0026amp;GN group (n\u0026nbsp;= 41, 43%)(Table 1). No significant sex-based disparity was observed between the groups (p\u0026nbsp;= 0.235). Tumor characteristics revealed that 65.6% (n\u0026nbsp;= 63) of lesions crossed the midline (p\u0026nbsp;= 0.070), while calcification was present in 34.3% (n\u0026nbsp;= 33;\u0026nbsp;p\u0026nbsp;= 0.124). Among the 73 patients tested for MYCN amplification, all results were negative. For model development, the cohort was partitioned into a training set (n\u0026nbsp;= 76) and an independent testing set (n\u0026nbsp;= 20). As detailed in Table 2, both sets demonstrated balanced distributions across all clinical variables (p\u0026nbsp;\u0026gt; 0.05), ensuring unbiased analytical validity.\u003c/p\u003e\n\u003cp\u003eTable 1. Basic characteristics of the study subjects.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTable 2. Clinicopathological characteristics of patients in the training and testing cohorts\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRadiomics feature selection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 2,418 radiomics features were initially extracted from the imaging data using the PyRadiomics toolkit. These features comprised 468 first-order statistics, 14 shape-based descriptors, and 1,898 texture-based features, including Gray Level Co-occurrence Matrix (GLCM, n=572), Gray Level Run Length Matrix (GLRLM, n=416), Gray Level Size Zone Matrix (GLSZM, n=416), Gray Level Dependence Matrix (GLDM, n=364), and Neighbouring Gray Tone Difference Matrix (NGTDM, n=130) features.\u003c/p\u003e\n\u003cp\u003eThe features were normalized using Robust Scaler to minimize the influence of outliers as shown in Figure 3, 4. Feature selection was carried out in three distinct stages: Initially, features with zero variance were excluded, leaving a total of 2,417 features. This was followed by univariate analysis using a two-sample t-test, which identified 38 features with significant intergroup differences (P \u0026lt; 0.05). In the next step, highly correlated features (Pearson r \u0026gt; 0.95) were removed through redundancy reduction, retaining the feature with the lower P-value, resulting in 28 features. Finally, LASSO regression was applied to select the 9 most discriminative features (shown in Table 3), based on non-zero coefficients and model convergence.\u003c/p\u003e\n\u003cp\u003eThe\u0026nbsp;radiomics score (Rad-score)\u0026nbsp;was computed using a\u0026nbsp;logistic regression model, where each feature\u0026rsquo;s value was weighted by its respective LASSO-derived coefficient. The Rad-score was defined as the linear combination:\u003c/p\u003e\n\u003cp\u003e\u003cimg width=\"326\" height=\"81\" src=\"data:image/png;base64,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\" alt=\"image\"\u003e\u003c/p\u003e\n\u003cp\u003ewhere \u003cem\u003ewi\u003c/em\u003e denotes the coefficient of the \u003cem\u003ei\u003c/em\u003e-th feature, \u003cem\u003exi\u003c/em\u003e is the normalized feature value, and \u003cem\u003eb\u003c/em\u003e is the intercept. This approach optimizes predictive performance while minimizing overfitting (Figure 4).\u003c/p\u003e\n\u003cp\u003eFigure 3. Original features and normalized features distribution histogram, the normalized method is Robust Scaler.\u003c/p\u003e\n\u003cp\u003eFigure 4. PCA plot with four quadrants obtained before (A) and after (B) feature normalization and screening.\u003c/p\u003e\n\u003cp\u003eFigure 5. The disparity in Rad-score distributions in the training (A) and testing (B) cohorts, and the results showed that there was a significant association between Rad-score and pathological classification of NB patients.\u003c/p\u003e\n\u003cp\u003eTable 3. The representative features after filtering.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePerformance outcomes of the radiomics models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe radiomics model demonstrated strong discriminatory performance in the training cohort, achieving an AUC of 0.888 (95% CI: [0.810\u0026ndash;0.957]) in distinguishing NB\u0026amp;GNBN from GNBI\u0026amp;GN pathological classifications. This robustness was confirmed in the testing cohort, where the model attained an even higher AUC of 0.920 (95% CI: [0.758\u0026ndash;0.100]), underscoring its generalizability (Figure 6).\u003c/p\u003e\n\u003cp\u003eSubsequent ROC analysis revealed a sensitivity of 0.8857 and specificity of 0.7805, indicating a well-balanced diagnostic performance. The decision curve analysis (Figure 7) underscored the model\u0026apos;s clinical relevance across a range of risk thresholds. In the training cohort, a net benefit was observed at thresholds between 0.01 and 0.98, supporting its applicability in diverse clinical settings. In the testing cohort, a net benefit was maintained within thresholds of 0.01 to 0.81, although its utility appeared somewhat more limited, likely due to cohort-specific differences.\u003c/p\u003e\n\u003cp\u003eFigure 6. The ROC curve of the training and testing cohorts.\u003c/p\u003e\n\u003cp\u003eFigure 7. The decision curve analysis for the model is presented. The grey line represents the scenario where the model predicts all patients without differentiating based on the pathological classification of NB. In contrast, the black curve reflects the assumption that the model predicts all patients while taking into account the pathological classification of NB.\u003c/p\u003e"},{"header":"Discussion ","content":"\u003cp\u003eThe dataset utilized in this study was divided into a training set and a testing set. The training set comprised 76 pediatric patients diagnosed with NB, while the testing set contained 20 patients. The final radiomics model was constructed by selecting 9 of the most relevant features out of an initial pool of 2,417 radiomics features. The logical basis for this process is the importance of features in predicting disease outcomes in pediatric patients with NB. The Random Forest algorithm was employed as the predictive modeling technique, chosen for its ability to handle high-dimensional data and its resilience to overfitting. The final radiomics model demonstrated excellent diagnostic performance in distinguishing between the four pathological classifications of NB, achieving an AUC of 0.920 in the independent testing cohort.\u003c/p\u003e\n\u003cp\u003eThe results of our study align with and expand upon previous research in the field of neuroblastoma (NB) diagnostics. For example, Feng L et al.\u0026nbsp;[14]\u0026nbsp;reported that PET/CT imaging provides more accurate identification of event-free survival (EFS) in NB, which supports our findings that imaging techniques are essential for identifying tumor characteristics. Although our model is based on radiomics features extracted from CT imaging modalities, its high AUC values in distinguishing between NB subtypes suggest that combining such radiomics approaches with advanced imaging technologies could further enhance diagnostic precision and prognostic assessment in pediatric NB patients. Furthermore, Wang H et al.\u0026nbsp;[15]\u0026nbsp;highlighted the potential of radiomics and deep learning advancements to enhance diagnostic capabilities. This aligns well with our study\u0026apos;s use of a Random Forest-based radiomics model, reinforcing the potential of AI-driven approaches in the pediatric oncology domain.\u003c/p\u003e\n\u003cp\u003eWhile traditional imaging techniques like CT play an important role in NB diagnosis, they can have limitations in certain cases. For instance, fibrosis and calcification can lead to reduced signal intensity on CT scans, potentially underestimating the severity of NB when CT is used in isolation\u0026nbsp;[16]. This highlights a key advantage of our approach: by incorporating radiomics features, which capture subtle patterns in imaging data, we can potentially overcome some of the limitations inherent to individual imaging techniques, such as CT. Additionally, while there are several studies using imaging techniques for NB diagnosis, the use of radiomics models for identifying pathological types remains relatively underexplored. One such study examined CT-based radio-genomics for predicting MYCN amplification, a genetic alteration in NB patients, which is associated with reduced survival rates\u0026nbsp;[17]. Our research goes a step further by focusing on radiomics modeling for predicting multiple pathological classifications in pediatric NB, thus extending the application of radiomics to a broader range of NB subtypes. Finally, our study differentiates itself by showing that radiomics is not limited to adult patients but can also be applied effectively to pediatric patients. This extension of radiomics applications in the pediatric NB context opens new avenues for future research and clinical practice, further emphasizing the potential of this approach to contribute to personalized treatment strategies for young patients with NB.\u003c/p\u003e\n\u003cp\u003eBuilding upon the current findings, future research can focus on the following directions: (1) Future studies could explore the integration of genomics, transcriptomics, and proteomics data with radiomics features to develop a more comprehensive understanding of NB. Combining these data types may provide deeper insights into tumor biology, enabling better prediction models that account for genetic and molecular characteristics alongside imaging features. (2) With the rapid development of AI technologies, particularly deep learning methodologies, future research should focus on incorporating these advanced techniques into the radiomics model. Deep learning models, with their ability to automatically learn complex patterns from large datasets, could significantly improve the accuracy and robustness of predictions in the context of NB diagnosis, risk stratification, and treatment planning. In its current form, the model can aid in the early and accurate diagnosis of NB, allowing for timely intervention. It can also be used to assess the efficacy of clinical treatments by monitoring changes in the tumor\u0026rsquo;s radiomic features over time. Ultimately, this model can enhance clinical decision-making, leading to improved prognostic assessment and more targeted treatments for pediatric NB patients.\u003c/p\u003e\n\u003cp\u003eThe distinctiveness of this study lies in its focus on predicting both the pathological classification and risk stratification of pediatric NB patients using a radiomics model. By combining advanced imaging features with machine learning techniques, this model provides a more comprehensive and nuanced approach to understanding the disease. This innovative methodology could significantly improve clinical decision-making, allowing for more accurate and personalized treatment strategies for pediatric NB patients.\u003c/p\u003e\n\u003cp\u003eThis study has important limitations to acknowledge. The manual segmentation of ROIs, though performed by experienced radiologists, remains subjective and could affect reproducibility. Additionally, the single-center design and limited sample size may constrain the generalizability of our findings. Future research should explore AI-based automated segmentation to improve consistency, building on existing successes in medical imaging applications. Multicenter collaborations would also help validate these results across more diverse populations and reduce potential selection bias.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;The radiomics model developed in this study uses imaging data from NB, including tumor pathological characteristics, as input. By integrating a large volume of imaging data, the model identifies potential correlations between imaging features and pathological traits, thereby providing more accurate predictions compared to conventional diagnostic methods. This model offers a dependable approach for predicting the pathological classification of NB and enables more effective patient stratification, which is especially useful for doctors in making personalized treatment decisions. \u0026nbsp; Moreover, the application of radiomics models in pediatric oncology holds the potential to further investigate the relationships between imaging features, genetic markers, and clinical outcomes, thus bridging the blank between radiology, genomics, and clinical practice.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eDr Yunyu He and Dr Fan Sai Hou Alexandre conceptualized and designed the study, drafted the initial manuscript, and critically reviewed and revised the manuscript.Dr Weiqiang Xiao, Qinglin Yang, and Xinwen Zhang, designed the data collection instruments, collected data, carried out the initial analyses, and critically reviewed and revised the manuscript.Dr Jianhua Liang, Zefeng Lin, Dongmei Huang, Hongmei Wu, Jiachi Liao, and Haonan Wang designed the data collection instruments, collected data, carried out the initial analyses, and critically reviewed and revised the manuscript.Dr Le Li, Jiahang Zeng conceptualized and designed the study, coordinated and supervised data collection, and critically reviewed and revised the manuscript for important intellectual content.All authors approved the final manuscript as submitted and agree to be accountable for all aspects of the work.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eZafar A, Wang W, Liu G, Wang X, Xian W, McKeon F et al (2021) Molecular targeting therapies for neuroblastoma: Progress and challenges. Med Res Rev 41(2):961\u0026ndash;1021\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eValter K, Zhivotovsky B, Gogvadze V (2018) Cell death-based treatment of neuroblastoma. Cell Death Dis 9(2):113\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWard E, DeSantis C, Robbins A, Kohler B, Jemal A (2014) Childhood and adolescent cancer statistics, 2014. CA Cancer J Clin 64(2):83\u0026ndash;103\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe WG, Yan Y, Tang W, Cai R, Ren G (2017) Clinical and biological features of neuroblastic tumors: A comparison of neuroblastoma and ganglioneuroblastoma. Oncotarget 8(23):37730\u0026ndash;37739\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYoneda A (2023) Role of surgery in neuroblastoma. Pediatr Surg Int 39(1):177\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQiu B, Matthay KK (2022) Advancing therapy for neuroblastoma. Nat Rev Clin Oncol 19(8):515\u0026ndash;533\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShimada H (2003) The International Neuroblastoma Pathology Classification. Pathologica 95(5):240\u0026ndash;241\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang X, Li C, Xu C, Hao X, Yu X, Li Q (2018) Correlation of CT signs with lymphatic metastasis and pathology of neuroblastoma in children. Oncol Lett 16(2):2439\u0026ndash;2443\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLittooij AS, de Keizer B (2023) Imaging in neuroblastoma. Pediatr Radiol 53(4):783\u0026ndash;787\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLambin P, Leijenaar RTH, Deist TM, Peerlings J, de Jong EEC, van Timmeren J et al (2017) Radiomics: the bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol 14(12):749\u0026ndash;762\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang H, Li T, Xie M, Si J, Qin J, Yang Y et al (2023) Association of Computed Tomography Radiomics Signature with Progression-free Survival in Neuroblastoma Patients. Clin Oncol (R Coll Radiol) 35(11):e639\u0026ndash;e47\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWu H, Wu C, Zheng H, Wang L, Guan W, Duan S et al (2021) Radiogenomics of neuroblastoma in pediatric patients: CT-based radiomics signature in predicting MYCN amplification. Eur Radiol 31(5):3080\u0026ndash;3089\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKoska IO, Ozcan HN, Tan AA, Beydogan B, Ozer G, Oguz B et al (2024) Radiomics in differential diagnosis of Wilms tumor and neuroblastoma with adrenal location in children. Eur Radiol 34(8):5016\u0026ndash;5027\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSepehri A, Stockton DJ, Roffey DM, Lefaivre KA, Potter JM, Guy P (2024) Effect of humeral rotation on the reliability of radiographic measurements for proximal humerus fractures. J Orthop Sci 29(4):1078\u0026ndash;1084\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang H, Chen X, He L (2023) A narrative review of radiomics and deep learning advances in neuroblastoma: updates and challenges. Pediatr Radiol 53(13):2742\u0026ndash;2755\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLadenstein R, Lambert B, P\u0026ouml;tschger U, Castellani MR, Lewington V, Bar-Sever Z et al (2018) Validation of the mIBG skeletal SIOPEN scoring method in two independent high-risk neuroblastoma populations: the SIOPEN/HR-NBL1 and COG-A3973 trials. Eur J Nucl Med Mol Imaging 45(2):292\u0026ndash;305\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGarc\u0026iacute;a N, Berm\u0026uacute;dez A, Mart\u0026iacute;n M, Carmona C, Ja\u0026eacute;n C, Daroca T (2020) [Analysis of the evolution of those operated on with minimal access surgery in our hospital. Does it present better results than the conventional one?]. Arch Cardiol Mex 91(3):321\u0026ndash;326\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTables 1 to 3 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"learning, neuroblastoma, radiomic model","lastPublishedDoi":"10.21203/rs.3.rs-7680804/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7680804/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eObjectives：\u003c/strong\u003eThis research seeks to create and validate radiomic models designed to enhance the diagnostic precision of neuroblastoma (NB), focusing specifically on distinguishing high-risk subtypes within its pathological classification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods: \u003c/strong\u003eThis retrospective study included 96 cases of NB, confirmed by histopathological evaluation. The cases were categorized according to the WHO classification into an aggressive group (n=55; neuroblastoma, Schwannian stroma-poor and ganglioneuroblastoma, nodular) and an indolent group (n=41; ganglioneuroma and ganglioneuroblastoma, intermixed). Radiomics features were extracted from CT images prior to biopsy or surgical resection. A radiomics model was constructed to predict NB classification using a random forest classifier. ROC curves were used to validate the capability of the models in the training and testing cohorts.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults: \u003c/strong\u003eThe final radiomics model incorporated 9 discriminative features. The model demonstrated strong diagnostic performance with an AUC of 0.874 in testing set, achieving a sensitivity of 77.5% and specificity of 88.6% in pathological classification. Decision curve analysis confirmed clinical utility across probability thresholds of 0.01-0.98 (training) and 0.01-0.81 (testing), indicating broad applicability for risk stratification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions：\u003c/strong\u003eThe developed radiomics model significantly improves diagnostic accuracy for neuroblastoma pathological classification, addressing a critical gap in current clinical practice. Unlike prior studies limited to binary NB diagnosis, this work provides granular discrimination between aggressive and indolent NB subtypes, enabling more precise risk stratification. While demonstrating robust performance, future multi-center validation is warranted to enhance generalizability and mitigate potential selection bias. This approach establishes a foundation for image-guided precision medicine in pediatric oncology.\u003c/p\u003e","manuscriptTitle":"Machine learning-based radiomics model for accurately predicting subtypes of neuroblastoma in children: A retrospective analysis","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-07 07:03:23","doi":"10.21203/rs.3.rs-7680804/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"3d55405d-7281-41b8-976e-e04b90af6556","owner":[],"postedDate":"November 7th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-11-19T17:23:18+00:00","versionOfRecord":[],"versionCreatedAt":"2025-11-07 07:03:23","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7680804","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7680804","identity":"rs-7680804","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}