Multi-Channel Multiphase CT-Based Deep Learning and Radiomics Fusion Model for Noninvasive Pathological Grading of Clear Cell Renal Cell Carcinoma

Multi-Channel Multiphase CT-Based Deep Learning and Radiomics Fusion Model for Noninvasive Pathological Grading of Clear Cell Renal Cell Carcinoma | 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 Multi-Channel Multiphase CT-Based Deep Learning and Radiomics Fusion Model for Noninvasive Pathological Grading of Clear Cell Renal Cell Carcinoma chongyang sun, qi chen, ze zhang, meng gao, shiqi he, wei zhang, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7430046/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 Purposes: To develop a combined model integrating multi-channel deep learning features, radiomics features, and clinical variables for noninvasive pathological grading of clear cell renal cell carcinoma (ccRCC). Material and methods: A retrospective study was conducted on 496 patients with pathologically confirmed ccRCC who underwent preoperative triple-phase contrast-enhanced CT. Multi-channel deep learning features were extracted from three ROI settings (conventional, tumor-only, and 5-mm expansion) by stacking arterial, medullary, and excretory phases. These were fused with arterial-phase radiomics features and clinical data to construct and compare predictive models. Results: In the ResNet50 model, the expanded 5mm ROI slice model had an AUC of 0.791 in the training group and 0.780 in the test group, indicating that the model could effectively predict the pathological grading of ccRCC. By combining deep learning features with radiomics and clinical features, the integrated model achieved AUCs of 0.855 in the training group and 0.849 in the test group, significantly outperforming the individual radiomics and clinical models. Decision curve analysis (DCA) further showed that the clinical-imaging combined model provided a higher net benefit. Conclusion: Multi-channel, multiphase CT fusion, when integrated with radiomics and clinical features, can significantly enhance predictive accuracy for ccRCC grading, providing a promising and interpretable noninvasive tool to support individualized treatment planning. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Clear cell renal cell carcinoma (ccRCC) is the most common subtype of renal malignancy, accounting for approximately 70–80% of all renal cell carcinoma cases[ 1 – 2 ]. Pathological grading of ccRCC plays a pivotal role in assessing tumor aggressiveness, metastatic potential, and patient prognosis[ 3 – 5 ]. Conventional grading methods rely on histopathological biopsy; however, due to the pronounced spatial and temporal heterogeneity of ccRCC, a single biopsy specimen often fails to fully capture the tumor’s biological characteristics, thereby limiting the accuracy of treatment planning[ 6 – 7 ]. This limitation underscores the urgent need for novel diagnostic strategies that can improve the accuracy and reliability of pathological grading[ 8 ]. In recent years, imaging-based, noninvasive predictive models have emerged as promising tools for tumor grading, risk stratification, and treatment guidance[ 9 – 11 ]. Among these, radiomics has attracted considerable attention by enabling the extraction of high-throughput quantitative features from standard-of-care medical images, revealing patterns beyond visual perception[ 12 ]. These handcrafted features—covering intensity, texture, and shape—can be used to train machine learning algorithms for disease classification and prognosis prediction[ 13 ]. However, the discriminative power of radiomics is often constrained by the predefinition of feature sets and susceptibility to variations in acquisition protocols[ 14 – 15 ]. Deep learning, which simulates the cognitive processes of the human brain by constructing multi-layer neural networks, optimizes a vast number of parameters via the backpropagation algorithm and progressively extracts features through nonlinear activation functions, has achieved remarkable success in medical imaging[ 16 ]. Its powerful feature-learning capability enables end-to-end modeling without the need for handcrafted feature engineering, allowing for the automatic discovery of complex hierarchical representations from raw data[ 17 ]. Convolutional neural networks (CNNs), in particular, have been extensively applied in tasks such as tumor detection, segmentation, and grading across various organ systems, demonstrating robustness and generalizability when adequately trained on diverse datasets[ 18 – 20 ]. In the development of deep learning models for tumor grading, different regions of interest (ROIs) allow networks to capture tumor biological characteristics at multiple spatial scales[ 21 – 23 ]. Narrow ROIs focus on the lesion core, potentially highlighting intra-tumoral heterogeneity, whereas expanded ROIs may include peritumoral tissue, which can harbor biologically relevant cues such as edema, angiogenesis, or subtle infiltration. By comparing model performance across different ROI configurations, one can evaluate the impact of spatial context on pathological grading accuracy and determine the optimal imaging analysis strategy. Multi-channel deep learning represents a cutting-edge paradigm for cross-modal information integration, implemented through a parallel heterogeneous data-stream processing architecture[ 24 ]. Its core concept involves employing independent branch networks to extract modality-specific features, followed by a hierarchical fusion strategy to construct a complementary and enriched joint representation[ 25 ]. In the context of multiphasic contrast-enhanced CT, stacking arterial, medullary, and excretory phases as multi-channel inputs can enrich the feature space by simultaneously capturing tumor vascularity, parenchymal enhancement patterns, and delayed washout characteristics[ 26 ]. Such fusion may enhance diagnostic accuracy, especially in conditions like ccRCC where vascular dynamics and tissue heterogeneity are key pathological hallmarks[ 27 – 28 ]. Previous studies, such as those by Ki Choon Sim et al. [ 29 ] and Enming Cui et al. [ 30 ], extracted radiomics features independently from MRI and CT sequences and performed feature-level fusion, without prior image-level fusion. While effective, these approaches may underutilize the synergistic potential of spatially aligned, multi-phase image data. In this study, we integrate multi-ROI images from three CT phases into a single three-channel representation and employ an end-to-end deep learning framework to directly learn high-level imaging features. These features are subsequently combined with arterial-phase radiomics features and clinical data to construct a more comprehensive and discriminative diagnostic model. We hypothesize that such integration will enable efficient, noninvasive pathological grading of ccRCC, ultimately facilitating more precise and individualized treatment strategies. Materials and Methods Study Population This retrospective study was approved by our institutional review board, and the requirement for informed consent was waived. All procedures were conducted in accordance with the Declaration of Helsinki. From June 2021 to December 2024, consecutive patients with pathologically confirmed clear cell renal cell carcinoma (ccRCC) who underwent preoperative contrast-enhanced CT scans at our institution were enrolled.Inclusion criteria were as follows: (1) Availability of preoperative CT images in both Medullary phase and excretory phase; (2) Solitary tumor with no prior history of radiotherapy or chemotherapy; (3) Histologically confirmed diagnosis postoperatively with complete clinical documentation. Exclusion criteria were: (1)Images with severe motion artifacts or significant noise compromising diagnostic quality.(2)Incomplete radiological or pathological documentation.(3)Non-clear cell renal carcinomas. A total of 496 patients met the eligibility criteria and were randomly assigned to a training cohort (n = 347) and a testing cohort (n = 149) in a 7:3 ratio (Fig. 1 ). Clinical and Imaging Data Clinical variables including sex, age, hematuria, and flank pain were collected from the hospital information system (HIS). Pathological grading was recorded according to standard criteria, with grades 1–2 defined as low grade and grades 3–4 as high grade. Two radiologists (with 6 and 8 years of CT experience, respectively) independently reviewed preoperative triple-phase CT scans to assess imaging features, including presence of calcification, intratumoral necrosis, and tumor interlobular arteries during the arterial phase. Both readers were blinded to the pathological diagnosis. Each tumor was assessed once by one radiologist; discrepancies were resolved by consensus with a third senior radiologist. CT Acquisition Protocol All CT examinations were performed using high-end multi-detector CT scanners (Philips Brilliance iCT 256, GE Revolution 128). Triple-phase contrast-enhanced scans included the arterial, medullary, and excretory phases. Scanning parameters were: tube voltage, 120 kV; tube current, 250 mA; slice thickness, 1 mm; and interslice gap, 1 mm. A nonionic iodinated contrast agent (70 mL) was injected at 4 mL/s, followed by a saline flush. The scan range covered the entire kidneys and tumors to ensure complete lesion capture. Image Segmentation All CT data were resampled to an isotropic voxel size of 1 mm×1 mm×1 mm to reduce inter-scanner variability. Window level and width were standardized to 50 HU and 350 HU, respectively, and intensity normalization was performed using Z-score standardization. Two experienced radiologists manually delineated three-dimensional regions of interest (ROIs) for each tumor in the arterial, medullary, and excretory phases using ITK-SNAP v3.8.0 software. Discrepancies were resolved by a third senior radiologist. To assess interobserver reproducibility, 50 randomly selected cases were used to calculate the intraclass correlation coefficient (ICC), with ICC > 0.75 indicating good agreement. 2D ROI Construction and Multi-channel Deep Learning Model Development Tumor ROIs were extracted from the original images using OneKey software. Three ROI configurations were generated: Originial ROI slice, encompassing the entire kidney region, providing a global context for tumor analysis; Tumor-only ROI slice, focusing solely on the tumor core to reduce background noise; Expanded ROI slice (Enlarge 5mm peritumoral region), incorporating peritumoral tissues to capture tumor microenvironmental features. All images were resized to 224×224 pixels. For each case, ROIs from the three contrast phases were stacked along the channel dimension to form a 224×224×3 multi-channel image as input. A ResNet50-based transfer learning framework was employed. Model training involved backpropagation for parameter optimization, with real-time data augmentation (random horizontal flipping and cropping). The cross-entropy loss function was used, with an initial learning rate of 0.01, cosine annealing schedule, 50 training epochs, and a batch size of 128. After training, deep learning features were extracted from the penultimate global average pooling layer for each ROI configuration. These features were then combined with radiomics features to construct and compare models (Fig. 2 ). Model interpretability was evaluated using gradient-weighted class activation mapping (Grad-CAM)[ 31 ], which generates heatmaps highlighting regions contributing most to classification decisions. Radiomics-Deep Learning Model Construction Based on previous evidence that arterial-phase CT features are highly discriminative for ccRCC grading, radiomics features were extracted only from arterial-phase images. These included first-order statistics, shape features, and multiple texture descriptors. Deep learning features from the three ROI configurations were fused separately with arterial-phase radiomics features to form: comb-1: tumor-only ROI deep learning features + arterial-phase radiomics features; comb-2: conventional ROI deep learning features + arterial-phase radiomics features; comb-3: expanded ROI deep learning features + arterial-phase radiomics features. In the training set, feature selection was performed in two steps: Analysis of variance (ANOVA) to retain features with ICC > 0.9 and statistically significant differences; Least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-testing to identify features with non-zero coefficients. logistic regression (LR), support vector machine (SVM), and random forest (RF) were evaluated. Diagnostic performance was compared using the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), and the best-performing radiomics model was selected. Clinical-Radiomics-Deep Learning Model Integration To enhance clinical utility, the optimal radiomics–deep learning model was integrated with clinical variables to build a combined model. Model performance was assessed in both training and testing cohorts using AUC, accuracy, sensitivity, specificity, PPV, and NPV. Decision curve analysis (DCA) was conducted to evaluate net clinical benefit across a range of threshold probabilities, thereby assessing the potential clinical impact in ccRCC grading [ 32 ]. Statistical Analysis Categorical variables were compared using the chi-square test or Fisher’s exact test, as appropriate. Continuous variables were compared using the Mann-Whitney U test or the independent-samples t test. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. The 95% confidence interval (CI) for the AUC was calculated using the auc function in the pROC R package. Differences in AUCs were compared using DeLong’s test, with P < 0.05 considered statistically significant. All statistical analyses were performed using R software (version 4.3.3) and the scikit-learn library in Python (version 3.11.11). Results Patient Characteristics and Clinical Features A total of 496 patients were included in this study. The clinical information and imaging characteristics of the patients are summarized in Table 1. In the training cohort, there were 227 males and 120 females, while the testing cohort included 93 males and 56 females. The mean age was 60.89 ± 10.06 years in the training cohort and 59.28 ± 9.72 years in the testing cohort.Among the clinical variables, hematuria showed a statistically significant difference between the training and testing cohorts (P < 0.05). Among the imaging features, intratumoral necrosis and the presence of tumor interlobular arteries were significantly different between the two cohorts (P < 0.05). No significant differences were observed for age and calcification in the testing cohort, for sex in the training cohort, or for flank pain in either cohort.Univariate and multivariate analyses revealed that only age was significantly associated with pathological grade, while the remaining variables showed no significant associations (Table 2). A clinical model was subsequently developed based on these findings. Performance and Visualization of the Deep Learning Model ResNet50 demonstrates excellent stability and robustness in the development of medical imaging diagnostic models; therefore, this architecture was adopted for our deep learning analysis. As shown in Table 3 and Fig. 3 A-B, among the different ROI configurations, the 5mm enlarged ROI achieved the best predictive performance. In the training cohort, the AUC was 0.791 (95% CI: 0.740–0.842), with an accuracy of 0.775, sensitivity of 0.772, specificity of 0.833, PPV of 0.694, and NPV of 0.819. In the testing cohort, the AUC was 0.780 (95% CI: 0.700–0.859), with an accuracy of 0.718, sensitivity of 0.745, specificity of 0.702, PPV of 0.594, and NPV of 0.825. Model interpretability was assessed using gradient-weighted class activation mapping (Grad-CAM), which visualized the spatial distribution of pixel importance in different colors, highlighting differences between high-grade and low-grade ccRCC tumors. Representative cases from the training cohort were examined. High-grade tumors generally exhibited larger activation regions, while low-grade tumors showed more limited activation. In most cases, the network demonstrated sensitivity to intratumoral heterogeneity and tumor margins in high-grade lesions (Fig. 4 ). Comparison of Combined Model Performance Deep learning features were further extracted from the trained models and fused with arterial-phase radiomics features to construct combined models. After LASSO feature selection, the three combined models retained 24, 31, and 34 non-zero coefficient features, respectively (Supplement Fig. 1 A-I). As shown in Supplement Fig. 2 A-F and Supplement Table 1, although the conventional ROI slice based combined model achieved good performance in the training cohort, it showed lower AUC in the testing cohort, indicating a risk of overfitting. In contrast, the 5-mm expanded ROI slice–based combined model demonstrated better overall performance compared to the other two models. When comparing the three machine learning classifiers, logistic regression (LR), support vector machine (SVM), and random forest (RF). The RF classifier exhibited the most stable performance, with an AUC of 0.849 in the training cohort and 0.831 in the testing cohort. Although SVM achieved a higher AUC in the training cohort, its performance decreased in the testing cohort, suggesting potential overfitting. Therefore, the RF model was selected as the optimal classifier. Subsequently, imaging features from the optimal combined model were integrated with clinical variables to construct the clinical–imaging combined model (Fig. 5 A-B). In the training cohort, the combined model achieved an AUC of 0.855, and in the testing cohort, an AUC of 0.849. Its predictive ability was significantly higher than that of the clinical model alone and slightly better than the imaging combined model. Decision curve analysis (DCA) demonstrated (Fig. 6 A-B) that the clinical-imaging combined model provided a higher net benefit across a range of threshold probabilities, indicating substantial potential for clinical application in ccRCC pathological grading. Discussion Deep learning has demonstrated substantial advantages in multimodal fusion analysis within medical imaging[ 33 – 35 ]. By integrating complementary information from CT, MRI, PET, and other imaging modalities, deep learning frameworks can characterize lesions from multiple perspectives, enabling more precise diagnosis and supporting clinical decision-making in complex cases[ 36 – 38 ]. The ability to extract hierarchical, nonlinear features from heterogeneous data streams allows such models to capture subtle morphological and functional patterns that may be imperceptible to traditional analysis[ 39 ]. As algorithms mature, datasets diversify, and clinical testing deepens, these approaches hold strong potential for advancing precision medicine and personalized treatment strategies[ 40 – 42 ]. In the present study, we developed a clinical-imaging combined model for ccRCC pathological grading by integrating deep learning features, radiomics features, and clinical variables. Our results demonstrated that the multi-channel model based on a 5-mm peritumoral expansion, when fused with radiomics features, achieved superior performance compared with models using conventional slices or tumor-only slices. The expanded ROI setting likely captures peritumoral microenvironmental cues, including subtle invasion patterns, microvascular proliferation, and perilesional heterogeneity, all of which are known to influence tumor aggressiveness. Furthermore, the expanded ROI reduces the risk of omitting informative peritumoral signals while simultaneously suppressing irrelevant background structures, thereby enhancing discriminative power. These findings underscore the importance of spatial context in noninvasive tumor grading. Our work builds upon prior evidence supporting the utility of multi-dimensional imaging features for tumor classification. For example, Lin et al. [ 43 ] reported significant gains in renal tumor grading accuracy when combining radiomics with deep learning features. However, few previous studies have systematically compared multiple ROI ranges in conjunction with feature fusion, and even fewer have incorporated peritumoral expansion in a multi-phase CT setting. Yang et al. [ 44 ] and Xu et al. [ 45 ] demonstrated the predictive value of deep features extracted from contrast-enhanced imaging, particularly in fusion frameworks, but omitted the excretory CT phase. The excretory phase provides additional structural and functional information such as enhanced delineation of tumor parenchyma boundaries, assessment of delayed enhancement patterns, and indirect markers of renal function that can complement arterial and medullary phase data. By stacking arterial, medullary, and excretory phases into a three-channel input, our model captures a more comprehensive physiological and pathological profile, which proved beneficial for feature fusion and downstream classification. Our findings also align with a recent systematic review comparing radiomics-only, deep learning-only, and fusion models, which reported that fusion models achieved the highest performance in 63% of included studies [ 46 ]. Multi-channel imaging enables each channel to encode distinct aspects of tumor biology; when these channels are fused, the resulting feature space becomes richer and more robust to image noise, artifacts, or variability in a single phase. Moreover, this approach enhances the detection of subtle imaging patterns that may be overlooked by single-phase or handcrafted features alone.A key innovation of our approach lies in its dual-level integration: image-level fusion via multi-channel inputs and feature-level fusion with radiomics and clinical data. This hierarchical strategy enables the model to leverage low-level texture and intensity patterns while also incorporating handcrafted descriptors that may capture domain-specific cues. Compared with approaches relying solely on deep learning or radiomics, this hybrid method maximizes the complementary strengths of each while mitigating their individual limitations. The advantages of this design are further supported by our decision curve analysis, which demonstrated consistent net clinical benefit over a broad range of threshold probabilities, underscoring its potential for integration into routine diagnostic workflows. We observed a modest shortfall in sensitivity. This may be attributable to limited sample size and class imbalance specifically, the underrepresentation of high-grade tumors leading to a bias toward majority-class (low-grade) predictions and thus more false negatives for aggressive disease [ 47 ]. Previous literature has shown that such imbalance can substantially degrade machine learning performance, particularly in high-stakes clinical prediction tasks. Additional factors, including noise, artifacts, and inter-scanner resolution variability, can further challenge feature extraction and model generalization. To mitigate these issues, future work should consider advanced sampling strategies, synthetic minority oversampling, cost-sensitive or focal loss functions, and targeted augmentation to improve sensitivity for high-grade tumors. Moreover, incorporating harmonization algorithms to minimize scanner-related variability and integrating multi-center datasets could enhance the robustness of the model. From a clinical perspective, the integration of multiphasic CT-derived features with routine clinical data into a unified predictive framework is particularly appealing. Such models could serve as decision-support tools to identify high-risk patients, prioritize biopsy, or tailor surveillance intervals. Nevertheless, successful clinical adoption will require rigorous external testing across diverse populations and imaging systems, as well as the incorporation of interpretability tools capable of providing clinicians with both visual and quantitative explanations for model predictions. Future research could also investigate integration with complementary modalities such as MRI, histopathology, and genomics to further enrich the feature space and improve predictive accuracy. Several limitations warrant acknowledgment. First, this was a single-center retrospective study, which may limit the generalizability of the findings. Differences in scanner hardware, acquisition protocols, and patient demographics across institutions could influence model performance in external settings. Second, although our dataset was of reasonable size, expanding the cohort, particularly increasing the proportion of high-grade tumors would enable more balanced training and reduce variability in performance estimates. Third, while our fusion framework effectively integrated deep learning and radiomics features, the optimal strategies for feature selection, fusion, and hyperparameter tuning remain to be determined. Advanced interpretability methods beyond Grad-CAM, such as attention-based saliency mapping or SHAP value analysis, could yield deeper insights into model decision-making and facilitate its acceptance in clinical practice. In summary, we present a multi-channel, multiphase CT based deep learning and radiomics fusion model for noninvasive ccRCC grading. By incorporating peritumoral expansion and leveraging complementary phase information, our framework generates a more expressive and robust feature representation, achieving strong predictive performance while maintaining clinical interpretability. Despite certain limitations, the model holds considerable promise as a decision-support tool for risk stratification and treatment planning. Future work should focus on multi-institutional testing, integration with additional modalities, and exploration of real-time deployment within clinical workflows. Declarations Author Contribution Conceptualization: Chongyang sun;Qi chenMethodology: Chongyang sun;Qi chenSoftware: Ze zhangFormal Analysis: Chongyang sun;Qi chen; Meng gao; Shiqi HeInvestigation: Wei zhangData Curation: Chongyang sun;Wei zhangWriting – Original Draft Preparation: Chongyang sun;Qi chenWriting – Review & Editing: Chongyang sun;Qi chen; Meng gao; Shiqi HeVisualization: Xigang xiaoSupervision: Xigang xiaoFunding Acquisition: Qi chen References Liao C, Hu L, Zhang Q. Von Hippel-Lindau protein signalling in clear cell renal cell carcinoma. Nat Rev Urol. 2024;21(11):662–675. Sato Y, Yoshizato T, Shiraishi Y, et al. Integrated molecular analysis of clear-cell renal cell carcinoma. Nat Genet. 2013;45(8):860–7. Deng H, Gong X, Ji G, Li C, Cheng S. KIF2C promotes clear cell renal cell carcinoma progression via activating JAK2/STAT3 signaling pathway. Mol Cell Probes. 72:101938. Zhao Z, Lu J, Qu H, et al. Evaluation and prognostic significance of manganese superoxide dismutase in clear cell renal cell carcinoma. Hum Pathol. 80:87–93. Ishigami K, Leite LV, Pakalniskis MG, Lee DK, Holanda DG, Kuehn DM. Tumor grade of clear cell renal cell carcinoma assessed by contrast-enhanced computed tomography. Springerplus. 3:694. Published 2014 None. Soultati A, Stares M, Swanton C, Larkin J, Turajlic S. How should clinicians address intratumour heterogeneity in clear cell renal cell carcinoma? Curr Opin Urol. 2015;25(5):358–66. Lopez JI, Cortes JM. A divide-and-conquer strategy in tumor sampling enhances detection of intratumor heterogeneity in routine pathology: A modeling approach in clear cell renal cell carcinoma. F1000Res. 5:385. Published 2016 None. Pratt S, Bowen I, Hallowell G, Shipman E, Redpath A. Assessment of agreement using the equine glandular gastric disease grading system in 84 cases. Vet Med Sci. 2022;8(4):1472–1477. Pulvirenti A, Yamashita R, Chakraborty J, et al. Quantitative Computed Tomography Image Analysis to Predict Pancreatic Neuroendocrine Tumor Grade. JCO Clin Cancer Inform. 5:679–694. Park YW, Oh J, You SC, et al. Radiomics and machine learning may accurately predict the grade and histological subtype in meningiomas using conventional and diffusion tensor imaging. Eur Radiol. 2019;29(8):4068–4076. Huang WY, Wen LH, Wu G, et al. Comparison of Radiomics Analyses Based on Different Magnetic Resonance Imaging Sequences in Grading and Molecular Genomic Typing of Glioma. J Comput Assist Tomogr. 2021 Jan-Feb 01;45(1):110–120. Guo H, Tang HT, Hu WL, et al. The application of radiomics in esophageal cancer: Predicting the response after neoadjuvant therapy. Front Oncol. 13:1082960. Published 2023 None. Sumon RI, Mozumdar MAI, Akter S, et al. Comparative Study of Cell Nuclei Segmentation Based on Computational and Handcrafted Features Using Machine Learning Algorithms. Diagnostics (Basel). 2025;15(10). Published 2025 May 16. Jimenez-Del-Toro O, Aberle C, Bach M, et al. The Discriminative Power and Stability of Radiomics Features With Computed Tomography Variations: Task-Based Analysis in an Anthropomorphic 3D-Printed CT Phantom. Invest Radiol. 2021;56(12):820–825. Michallek F, Genske U, Niehues SM, Hamm B, Jahnke P. Deep learning reconstruction improves radiomics feature stability and discriminative power in abdominal CT imaging: a phantom study. Eur Radiol. 2022;32(7):4587–4595. Bai Y, Zhao H, Shi X, Chen L. Towards practical and privacy-preserving CNN inference service for cloud-based medical imaging analysis: A homomorphic encryption-based approach. Comput Methods Programs Biomed. 261:108599. Attique H, Shah S, Jabeen S, Khan FG, Khan A, ELAffendi M. Multiclass Cancer Prediction Based on Copy Number Variation Using Deep Learning. Comput Intell Neurosci. 2022:4742986. Published 2022 None. Clarke EL, Magee D, Newton-Bishop J, et al. The Development and Evaluation of a Convolutional Neural Network for Cutaneous Melanoma Detection in Whole Slide Images. Arch Pathol Lab Med. Published online Dec 12,2024. Zhao Q, Kong P, Min J, et al. [A review of deep learning methods for the detection and classification of pulmonary nodules]. Sheng Wu Yi Xue Gong Cheng Xue Za Zhi. 2019;36(6):1060–1068. Khan MSI, Rahman A, Debnath T, et al. Accurate brain tumor detection using deep convolutional neural network. Comput Struct Biotechnol J. 20:4733–4745. Li S, Liu J, Song Z. Brain tumor segmentation based on region of interest-aided localization and segmentation U-Net. Int J Mach Learn Cybern. 2022;13(9):2435–2445. Wei G, Fang G, Guo P, et al. Preoperative prediction of microvascular invasion risk in hepatocellular carcinoma with MRI: peritumoral versus tumor region. Insights Imaging. 2024;15(1):188. Mercan C, Balkenhol M, Salgado R, et al. Deep learning for fully-automated nuclear pleomorphism scoring in breast cancer. NPJ Breast Cancer. 2022;8(1):120. Published 2022 Nov 8. Lian J, Dong Z, Zhang H, Chen Y, Liu J. Multifocal region-assisted cross-modality learning for chest X-ray report generation. Comput Biol Med. 183:109187. Lu C, Zhang J, Zhang D, et al. Fine-Grained Hierarchical Progressive Modal-Aware Network for Brain Tumor Segmentation. IEEE J Biomed Health Inform. PP. Published online May 22,2025. Coppenrath EM, Mueller-Lisse UG. Multidetector CT of the kidney. Eur Radiol. 2006;16(11):2603–11. Li X, Guo Y, Huang S, et al. A CT-based intratumoral and peritumoral radiomics nomogram for postoperative recurrence risk stratification in localized clear cell renal cell carcinoma. BMC Med Imaging. 2025;25(1):167. Published 2025 May 16. Mao J, Li W, Sun X, et al. RPF-Net: A multimodal model for the postoperative UISS risk stratification of non-metastatic ccRCC based on CT and whole-slide images. Comput Methods Programs Biomed. 268:108836. Sim KC, Han NY, Cho Y, Sung DJ, Park BJ, Kim MJ, Han YE. Machine Learning-Based Magnetic Resonance Radiomics Analysis for Predicting Low- and High-Grade Clear Cell Renal Cell Carcinoma. J Comput Assist Tomogr. 2023 Nov-Dec 01;47(6):873–881. Cui E, Li Z, Ma C, Li Q, Lei Y, Lan Y, Yu J, Zhou Z, Li R, Long W, Lin F. Predicting the ISUP grade of clear cell renal cell carcinoma with multiparametric MR and multiphase CT radiomics. Eur Radiol. 2020;30(5):2912–2921. Selvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D (2020) Grad-CAM: visual explanations from deep networks via gradient-based localization. International Journal of Com-puter Vision 128:336–359 Parmar C, Grossmann P, Bussink J, Lambin P, Aerts HJWL (2015) Machine learning methods for quantitative radiomic biomarkers. Sci Rep 5:13087 Xu J, Zeng B, Egger J, et al. A review on AI-based medical image computing in head and neck surgery. Phys Med Biol. 2022;67(17). Dana J, Agnus V, Ouhmich F, Gallix B. Multimodality Imaging and Artificial Intelligence for Tumor Characterization: Current Status and Future Perspective. Semin Nucl Med. 2020;50(6):541–548. Vairetti C, Maldonado S, Cuitino L, Urzua CA. Interpretable multimodal classification for age-related macular degeneration diagnosis. PLoS One. 2024;19(11):e0311811. Qu J, Huang D, Shi Y, Liu J, Tang W. Diffusion-driven multi-modality medical image fusion. Med Biol Eng Comput. 2025;63(7):2105–2118. Liu R, Liu Y, Wang H, Li J, Hu K. SAFFusion: a saliency-aware frequency fusion network for multimodal medical image fusion. Biomed Opt Express. 2025;16(6):2459–2481. Wei L, Osman S, Hatt M, El Naqa I. Machine learning for radiomics-based multimodality and multiparametric modeling. Q J Nucl Med Mol Imaging. 2019;63(4):323–338. Smedley NF, Hsu W. USING DEEP NEURAL NETWORKS FOR RADIOGENOMIC ANALYSIS. Proc IEEE Int Symp Biomed Imaging. 2018:1529–1533. Shams A. Leveraging State-of-the-Art AI Algorithms in Personalized Oncology: From Transcriptomics to Treatment. Diagnostics (Basel). 2024;14(19). Stetson LC, Pearl T, Chen Y, Barnholtz-Sloan JS. Computational identification of multi-omic correlates of anticancer therapeutic response. BMC Genomics. 15 Suppl 7:S2. Kayikcioglu E, Onder AH, Bacak B, Serel TA. Machine learning for predicting colon cancer recurrence. Surg Oncol. 54:102079. Lin F, Ma C, Xu J, Lei Y, Li Q, Lan Y, Sun M, Long W, Cui E. A CT-based deep learning model for predicting the nuclear grade of clear cell renal cell carcinoma. Eur J Radiol. 2020;129:109079. Yang M, He X, Xu L, Liu M, Deng J, Cheng X, Wei Y, Li Q, Wan S, Zhang F, Wu L, Wang X, Song B, Liu M. CT-based transformer model for non-invasively predicting the Fuhrman nuclear grade of clear cell renal cell carcinoma. Front Oncol. 2022;12:961779. Xu L, Yang C, Zhang F, Cheng X, Wei Y, Fan S, Liu M, He X, Deng J, Xie T, Wang X, Liu M, Song B. Deep Learning Using CT Images to Grade Clear Cell Renal Cell Carcinoma: Development and Validation of a Prediction Model. Cancers (Basel). 2022;14(11):2574. Demircio˘glu A. Are deep models in radiomics performing better than generic models? A systematic review. Eur Radiol Exp.2023;7(1):11. Xu L, Yang C, Zhang F, Cheng X, Wei Y, Fan S, Liu M, He X, Deng J, Xie T, Wang X, Liu M, Song B. Deep Learning Using CT Images to Grade Clear Cell Renal Cell Carcinoma: Development and Validation of a Prediction Model. Cancers (Basel). 2022;14(11):2574. Tables Tables 1 to 3 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table2.docx Table3.docx Table1.docx SupplementFigure1C.jpg SupplementFigure1F.jpg SupplementFigure1E.jpg SupplementFigure1G.jpg SupplementFigure1A.jpg SupplementFigure1I.jpg SupplementFigure2A.jpg SupplementFigure1B.jpg SupplementFigure1D.jpg SupplementFigure2C.jpg SupplementTable1.docx SupplementFigure2D.jpg SupplementFigure2F.jpg SupplementFigure1H.jpg SupplementFigure2E.jpg SupplementFigure2B.jpg 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-7430046","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":509664457,"identity":"7cc593ab-0bfb-46fa-a246-6ff55f1119ab","order_by":0,"name":"chongyang sun","email":"","orcid":"","institution":"First Affiliated Hospital of Harbin Medical University","correspondingAuthor":false,"prefix":"","firstName":"chongyang","middleName":"","lastName":"sun","suffix":""},{"id":509664458,"identity":"946fb436-a462-49a9-8dca-3b93b7a32c97","order_by":1,"name":"qi chen","email":"","orcid":"","institution":"First Affiliated Hospital of Harbin Medical University","correspondingAuthor":false,"prefix":"","firstName":"qi","middleName":"","lastName":"chen","suffix":""},{"id":509664459,"identity":"bc169200-ebef-4039-bc75-fd0f7ed4e93d","order_by":2,"name":"ze zhang","email":"","orcid":"","institution":"First Affiliated Hospital of Harbin Medical University","correspondingAuthor":false,"prefix":"","firstName":"ze","middleName":"","lastName":"zhang","suffix":""},{"id":509664461,"identity":"c6021fe9-6ad9-4a32-a9ea-cd12ee772a7e","order_by":3,"name":"meng gao","email":"","orcid":"","institution":"First Affiliated Hospital of Harbin Medical University","correspondingAuthor":false,"prefix":"","firstName":"meng","middleName":"","lastName":"gao","suffix":""},{"id":509664463,"identity":"efb773ae-efb2-4eff-bf16-4230e7bfb7a6","order_by":4,"name":"shiqi he","email":"","orcid":"","institution":"First Affiliated Hospital of Harbin Medical University","correspondingAuthor":false,"prefix":"","firstName":"shiqi","middleName":"","lastName":"he","suffix":""},{"id":509664465,"identity":"83b2473f-eeac-470b-b031-f81cde712b05","order_by":5,"name":"wei zhang","email":"","orcid":"","institution":"First Affiliated Hospital of Harbin Medical University","correspondingAuthor":false,"prefix":"","firstName":"wei","middleName":"","lastName":"zhang","suffix":""},{"id":509664467,"identity":"feb6a66e-b72e-4b6f-9158-6db7db539ba3","order_by":6,"name":"xigang xiao","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1klEQVRIiWNgGAWjYBACfmbGhsN/KmwSGCTAfGbCWiTbmxsf8JxJI0GLQc/xZgPetsOkaJFIbJOQOHM+z3x2d5oEQ4V1YgP72QN4tZiDtBhU3C6WuXN2mwTDmfTEBp68BLxaLGcAtSScuZ04QyJ3mwRj2+HEBgkeA/wOuwHUcrDtHFTLP2K0nDnYbNjYdgCqpYEILZLtjY2PGc4kJ86QObvZIuFYunEbTw5+LfzM7A8OM1TYJc6Q7t1440ONtWw/+xn8WlBBAhCzkaB+FIyCUTAKRgEOAAD8aEq7z8uItwAAAABJRU5ErkJggg==","orcid":"","institution":"First Affiliated Hospital of Harbin Medical University","correspondingAuthor":true,"prefix":"","firstName":"xigang","middleName":"","lastName":"xiao","suffix":""}],"badges":[],"createdAt":"2025-08-22 02:38:18","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7430046/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7430046/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":90792506,"identity":"ff88ccff-ab0b-4213-b752-5fad45e3b059","added_by":"auto","created_at":"2025-09-08 08:29:30","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":107277,"visible":true,"origin":"","legend":"\u003cp\u003eStudy Population Enrollment Process.\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/c4fb7506634c5c2067750070.jpg"},{"id":90792510,"identity":"959d9589-e5df-41cb-95ba-2e54642dffbd","added_by":"auto","created_at":"2025-09-08 08:29:31","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":222118,"visible":true,"origin":"","legend":"\u003cp\u003eWorkflow of this study.\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/25037b218eb1badc7dc92a2f.jpg"},{"id":90792513,"identity":"ea244857-2dd3-4c21-b12a-0006533a37c9","added_by":"auto","created_at":"2025-09-08 08:29:31","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":769114,"visible":true,"origin":"","legend":"\u003cp\u003eA-B:The receiver operating characteristic (ROC) curves of three different roi deep learning model in the training cohort and testing cohort.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/9f0ca6cbe3ec5a0e938673aa.jpg"},{"id":90792512,"identity":"cbab0ead-a70f-43ae-8dd5-ecd84290fb21","added_by":"auto","created_at":"2025-09-08 08:29:31","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":39255,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative images and illustration of the ccRCC pathological grade prediction visualization. (a) Feature map extracted by high-grade on CT image. Grad-CAM plots highlighted large, multiple regions activated within the tumor of patients.(b)Feature map extracted by low-grade on CT image. Grad-CAM plots emphasized very few regions activated within the tumor of patients.\u003c/p\u003e","description":"","filename":"Figure4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/5dc5fddc793bd7fd1be8d302.jpg"},{"id":90792547,"identity":"be3936bb-84cf-4a64-aa7d-9cc99f1611b2","added_by":"auto","created_at":"2025-09-08 08:29:32","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":760394,"visible":true,"origin":"","legend":"\u003cp\u003e(E-F):The ROC curves of roi-enlarge Clinical-imaging integrated model in the training cohort and testing cohort.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/4aab07ab9dc9abb85d7d54e7.jpg"},{"id":90792545,"identity":"ecde5ad1-cc6d-45cc-9336-dc340eb8db6d","added_by":"auto","created_at":"2025-09-08 08:29:32","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":760261,"visible":true,"origin":"","legend":"\u003cp\u003e(A-B):Decision curve analysis (DCA) for three models in classifying the diagnostic pathological grading in training and test cohorts, respectively.\u003c/p\u003e","description":"","filename":"6.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/95c6581105422c97c0196241.jpg"},{"id":90951717,"identity":"1024a6b9-76de-4591-8350-2feab1c8fb69","added_by":"auto","created_at":"2025-09-10 01:01:25","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3070139,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/70c86bf1-c7b9-4be8-8941-554107647f51.pdf"},{"id":90792540,"identity":"ea9c313b-e3c6-4c3a-a55f-bf663b18fb7c","added_by":"auto","created_at":"2025-09-08 08:29:32","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":13112,"visible":true,"origin":"","legend":"","description":"","filename":"Table2.docx","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/d11b7f16afdeea4b7f4a6215.docx"},{"id":90792508,"identity":"6b220f9d-9540-4f44-b4b0-0d438a4fbe2e","added_by":"auto","created_at":"2025-09-08 08:29:30","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":12872,"visible":true,"origin":"","legend":"","description":"","filename":"Table3.docx","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/c7dd39e10e0bd29c0cc6b81b.docx"},{"id":90792514,"identity":"55306f5c-bbab-4e74-a8aa-523d0c642d72","added_by":"auto","created_at":"2025-09-08 08:29:31","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":16337,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/1c1a5953cd48d264b2d184bd.docx"},{"id":90792563,"identity":"3a4684ed-5d99-4b05-a754-1ff7b07ee100","added_by":"auto","created_at":"2025-09-08 08:29:33","extension":"jpg","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":87987,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1C.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/cf5a863ba21a4a4b24b76e30.jpg"},{"id":90793074,"identity":"a3ffd347-e59c-47f3-840d-4ec79e20c04c","added_by":"auto","created_at":"2025-09-08 08:37:31","extension":"jpg","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":92760,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1F.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/f78eaa39bca729aa6cde1bf6.jpg"},{"id":90792527,"identity":"426f5739-e651-4668-a487-43fd99b0a2c4","added_by":"auto","created_at":"2025-09-08 08:29:31","extension":"jpg","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":108035,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1E.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/720bc23248179ec7dba9ef9c.jpg"},{"id":90792522,"identity":"f6e25769-2a24-47f6-b1e6-1dc885062236","added_by":"auto","created_at":"2025-09-08 08:29:31","extension":"jpg","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":120764,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1G.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/5af7744549704b7975616364.jpg"},{"id":90793083,"identity":"5a55a5d9-b5d7-457c-b04a-3a6625394550","added_by":"auto","created_at":"2025-09-08 08:37:33","extension":"jpg","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":102587,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1A.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/0ea8fd4388d50492a259bb58.jpg"},{"id":90793067,"identity":"d08c4f88-694b-4125-8f47-69eddbbf63bb","added_by":"auto","created_at":"2025-09-08 08:37:31","extension":"jpg","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":87550,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1I.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/6a93320c84c4f267f6b5f334.jpg"},{"id":90793072,"identity":"7c4994b7-ec94-4646-994b-98ccad7f4dd1","added_by":"auto","created_at":"2025-09-08 08:37:31","extension":"jpg","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":86799,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure2A.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/2ef816f64974d08b636d449f.jpg"},{"id":90792538,"identity":"a5dfe986-45f0-4133-974b-b21e4296d094","added_by":"auto","created_at":"2025-09-08 08:29:32","extension":"jpg","order_by":10,"title":"","display":"","copyAsset":false,"role":"supplement","size":98249,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1B.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/26921b22fd80c25a3e7595a0.jpg"},{"id":90792529,"identity":"48d4ef5e-8a12-446d-9667-781ca5aa8dba","added_by":"auto","created_at":"2025-09-08 08:29:32","extension":"jpg","order_by":11,"title":"","display":"","copyAsset":false,"role":"supplement","size":107366,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1D.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/bf3a181c57b4ed1ca8bac2e1.jpg"},{"id":90792560,"identity":"19673b50-83ad-484f-9f9d-f943fa493e14","added_by":"auto","created_at":"2025-09-08 08:29:33","extension":"jpg","order_by":12,"title":"","display":"","copyAsset":false,"role":"supplement","size":83778,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure2C.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/0580497ae569b12a413f7dec.jpg"},{"id":90792528,"identity":"d26a181e-bc8e-4218-a3b3-8ca01dc3ea26","added_by":"auto","created_at":"2025-09-08 08:29:32","extension":"docx","order_by":13,"title":"","display":"","copyAsset":false,"role":"supplement","size":17380,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementTable1.docx","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/4f3c996ccf5f2e8b488b5ad7.docx"},{"id":90794650,"identity":"e31da2ef-556e-4cee-814a-e08f111785a7","added_by":"auto","created_at":"2025-09-08 08:45:32","extension":"jpg","order_by":14,"title":"","display":"","copyAsset":false,"role":"supplement","size":86272,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure2D.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/367a21c797b8a6f309289269.jpg"},{"id":90792567,"identity":"4403cedb-72dc-4a79-b915-357705e15a2a","added_by":"auto","created_at":"2025-09-08 08:29:34","extension":"jpg","order_by":15,"title":"","display":"","copyAsset":false,"role":"supplement","size":84581,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure2F.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/2441a46f9d173c7d8c2b9218.jpg"},{"id":90792525,"identity":"efe570e0-1ac8-4404-af26-ea5c8e912cf6","added_by":"auto","created_at":"2025-09-08 08:29:31","extension":"jpg","order_by":16,"title":"","display":"","copyAsset":false,"role":"supplement","size":98115,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure1H.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/0722041a536200a8c8002c98.jpg"},{"id":90792532,"identity":"977cdc10-10ef-4644-b404-7af91d779500","added_by":"auto","created_at":"2025-09-08 08:29:32","extension":"jpg","order_by":17,"title":"","display":"","copyAsset":false,"role":"supplement","size":84690,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure2E.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/426f78e12ca23e050e149a25.jpg"},{"id":90792556,"identity":"1516ee7c-f6fc-4239-af0f-34a88d071f15","added_by":"auto","created_at":"2025-09-08 08:29:33","extension":"jpg","order_by":18,"title":"","display":"","copyAsset":false,"role":"supplement","size":83917,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementFigure2B.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7430046/v1/56b255f966e5a6d98c8a2170.jpg"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multi-Channel Multiphase CT-Based Deep Learning and Radiomics Fusion Model for Noninvasive Pathological Grading of Clear Cell Renal Cell Carcinoma","fulltext":[{"header":"Introduction","content":"\u003cp\u003eClear cell renal cell carcinoma (ccRCC) is the most common subtype of renal malignancy, accounting for approximately 70\u0026ndash;80% of all renal cell carcinoma cases[\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Pathological grading of ccRCC plays a pivotal role in assessing tumor aggressiveness, metastatic potential, and patient prognosis[\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Conventional grading methods rely on histopathological biopsy; however, due to the pronounced spatial and temporal heterogeneity of ccRCC, a single biopsy specimen often fails to fully capture the tumor\u0026rsquo;s biological characteristics, thereby limiting the accuracy of treatment planning[\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. This limitation underscores the urgent need for novel diagnostic strategies that can improve the accuracy and reliability of pathological grading[\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn recent years, imaging-based, noninvasive predictive models have emerged as promising tools for tumor grading, risk stratification, and treatment guidance[\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Among these, radiomics has attracted considerable attention by enabling the extraction of high-throughput quantitative features from standard-of-care medical images, revealing patterns beyond visual perception[\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. These handcrafted features\u0026mdash;covering intensity, texture, and shape\u0026mdash;can be used to train machine learning algorithms for disease classification and prognosis prediction[\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. However, the discriminative power of radiomics is often constrained by the predefinition of feature sets and susceptibility to variations in acquisition protocols[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eDeep learning, which simulates the cognitive processes of the human brain by constructing multi-layer neural networks, optimizes a vast number of parameters via the backpropagation algorithm and progressively extracts features through nonlinear activation functions, has achieved remarkable success in medical imaging[\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Its powerful feature-learning capability enables end-to-end modeling without the need for handcrafted feature engineering, allowing for the automatic discovery of complex hierarchical representations from raw data[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Convolutional neural networks (CNNs), in particular, have been extensively applied in tasks such as tumor detection, segmentation, and grading across various organ systems, demonstrating robustness and generalizability when adequately trained on diverse datasets[\u003cspan additionalcitationids=\"CR19\" citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn the development of deep learning models for tumor grading, different regions of interest (ROIs) allow networks to capture tumor biological characteristics at multiple spatial scales[\u003cspan additionalcitationids=\"CR22\" citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Narrow ROIs focus on the lesion core, potentially highlighting intra-tumoral heterogeneity, whereas expanded ROIs may include peritumoral tissue, which can harbor biologically relevant cues such as edema, angiogenesis, or subtle infiltration. By comparing model performance across different ROI configurations, one can evaluate the impact of spatial context on pathological grading accuracy and determine the optimal imaging analysis strategy.\u003c/p\u003e\u003cp\u003eMulti-channel deep learning represents a cutting-edge paradigm for cross-modal information integration, implemented through a parallel heterogeneous data-stream processing architecture[\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. Its core concept involves employing independent branch networks to extract modality-specific features, followed by a hierarchical fusion strategy to construct a complementary and enriched joint representation[\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. In the context of multiphasic contrast-enhanced CT, stacking arterial, medullary, and excretory phases as multi-channel inputs can enrich the feature space by simultaneously capturing tumor vascularity, parenchymal enhancement patterns, and delayed washout characteristics[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e]. Such fusion may enhance diagnostic accuracy, especially in conditions like ccRCC where vascular dynamics and tissue heterogeneity are key pathological hallmarks[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e].\u003c/p\u003e\u003cp\u003ePrevious studies, such as those by Ki Choon Sim et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] and Enming Cui et al. [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e], extracted radiomics features independently from MRI and CT sequences and performed feature-level fusion, without prior image-level fusion. While effective, these approaches may underutilize the synergistic potential of spatially aligned, multi-phase image data. In this study, we integrate multi-ROI images from three CT phases into a single three-channel representation and employ an end-to-end deep learning framework to directly learn high-level imaging features. These features are subsequently combined with arterial-phase radiomics features and clinical data to construct a more comprehensive and discriminative diagnostic model. We hypothesize that such integration will enable efficient, noninvasive pathological grading of ccRCC, ultimately facilitating more precise and individualized treatment strategies.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eStudy Population\u003c/p\u003e\u003cp\u003e This retrospective study was approved by our institutional review board, and the requirement for informed consent was waived. All procedures were conducted in accordance with the Declaration of Helsinki. From June 2021 to December 2024, consecutive patients with pathologically confirmed clear cell renal cell carcinoma (ccRCC) who underwent preoperative contrast-enhanced CT scans at our institution were enrolled.Inclusion criteria were as follows: (1) Availability of preoperative CT images in both Medullary phase and excretory phase; (2) Solitary tumor with no prior history of radiotherapy or chemotherapy; (3) Histologically confirmed diagnosis postoperatively with complete clinical documentation. Exclusion criteria were: (1)Images with severe motion artifacts or significant noise compromising diagnostic quality.(2)Incomplete radiological or pathological documentation.(3)Non-clear cell renal carcinomas. A total of 496 patients met the eligibility criteria and were randomly assigned to a training cohort (n\u0026thinsp;=\u0026thinsp;347) and a testing cohort (n\u0026thinsp;=\u0026thinsp;149) in a 7:3 ratio (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eClinical and Imaging Data\u003c/p\u003e\u003cp\u003eClinical variables including sex, age, hematuria, and flank pain were collected from the hospital information system (HIS). Pathological grading was recorded according to standard criteria, with grades 1\u0026ndash;2 defined as low grade and grades 3\u0026ndash;4 as high grade. Two radiologists (with 6 and 8 years of CT experience, respectively) independently reviewed preoperative triple-phase CT scans to assess imaging features, including presence of calcification, intratumoral necrosis, and tumor interlobular arteries during the arterial phase. Both readers were blinded to the pathological diagnosis. Each tumor was assessed once by one radiologist; discrepancies were resolved by consensus with a third senior radiologist.\u003c/p\u003e\u003cp\u003eCT Acquisition Protocol\u003c/p\u003e\u003cp\u003eAll CT examinations were performed using high-end multi-detector CT scanners (Philips Brilliance iCT 256, GE Revolution 128). Triple-phase contrast-enhanced scans included the arterial, medullary, and excretory phases. Scanning parameters were: tube voltage, 120 kV; tube current, 250 mA; slice thickness, 1 mm; and interslice gap, 1 mm. A nonionic iodinated contrast agent (70 mL) was injected at 4 mL/s, followed by a saline flush. The scan range covered the entire kidneys and tumors to ensure complete lesion capture.\u003c/p\u003e\u003cp\u003eImage Segmentation\u003c/p\u003e\u003cp\u003eAll CT data were resampled to an isotropic voxel size of 1 mm\u0026times;1 mm\u0026times;1 mm to reduce inter-scanner variability. Window level and width were standardized to 50 HU and 350 HU, respectively, and intensity normalization was performed using Z-score standardization. Two experienced radiologists manually delineated three-dimensional regions of interest (ROIs) for each tumor in the arterial, medullary, and excretory phases using ITK-SNAP v3.8.0 software. Discrepancies were resolved by a third senior radiologist. To assess interobserver reproducibility, 50 randomly selected cases were used to calculate the intraclass correlation coefficient (ICC), with ICC\u0026thinsp;\u0026gt;\u0026thinsp;0.75 indicating good agreement.\u003c/p\u003e\u003cp\u003e2D ROI Construction and Multi-channel Deep Learning Model Development\u003c/p\u003e\u003cp\u003eTumor ROIs were extracted from the original images using OneKey software. Three ROI configurations were generated: Originial ROI slice, encompassing the entire kidney region, providing a global context for tumor analysis; Tumor-only ROI slice, focusing solely on the tumor core to reduce background noise; Expanded ROI slice (Enlarge 5mm peritumoral region), incorporating peritumoral tissues to capture tumor microenvironmental features. All images were resized to 224\u0026times;224 pixels. For each case, ROIs from the three contrast phases were stacked along the channel dimension to form a 224\u0026times;224\u0026times;3 multi-channel image as input.\u003c/p\u003e\u003cp\u003eA ResNet50-based transfer learning framework was employed. Model training involved backpropagation for parameter optimization, with real-time data augmentation (random horizontal flipping and cropping). The cross-entropy loss function was used, with an initial learning rate of 0.01, cosine annealing schedule, 50 training epochs, and a batch size of 128. After training, deep learning features were extracted from the penultimate global average pooling layer for each ROI configuration. These features were then combined with radiomics features to construct and compare models (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eModel interpretability was evaluated using gradient-weighted class activation mapping (Grad-CAM)[\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e], which generates heatmaps highlighting regions contributing most to classification decisions.\u003c/p\u003e\u003cp\u003eRadiomics-Deep Learning Model Construction\u003c/p\u003e\u003cp\u003eBased on previous evidence that arterial-phase CT features are highly discriminative for ccRCC grading, radiomics features were extracted only from arterial-phase images. These included first-order statistics, shape features, and multiple texture descriptors. Deep learning features from the three ROI configurations were fused separately with arterial-phase radiomics features to form: comb-1: tumor-only ROI deep learning features\u0026thinsp;+\u0026thinsp;arterial-phase radiomics features; comb-2: conventional ROI deep learning features\u0026thinsp;+\u0026thinsp;arterial-phase radiomics features; comb-3: expanded ROI deep learning features\u0026thinsp;+\u0026thinsp;arterial-phase radiomics features. In the training set, feature selection was performed in two steps: Analysis of variance (ANOVA) to retain features with ICC\u0026thinsp;\u0026gt;\u0026thinsp;0.9 and statistically significant differences; Least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-testing to identify features with non-zero coefficients. logistic regression (LR), support vector machine (SVM), and random forest (RF) were evaluated. Diagnostic performance was compared using the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), and the best-performing radiomics model was selected.\u003c/p\u003e\u003cp\u003eClinical-Radiomics-Deep Learning Model Integration\u003c/p\u003e\u003cp\u003eTo enhance clinical utility, the optimal radiomics\u0026ndash;deep learning model was integrated with clinical variables to build a combined model. Model performance was assessed in both training and testing cohorts using AUC, accuracy, sensitivity, specificity, PPV, and NPV. Decision curve analysis (DCA) was conducted to evaluate net clinical benefit across a range of threshold probabilities, thereby assessing the potential clinical impact in ccRCC grading [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e].\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eCategorical variables were compared using the chi-square test or Fisher\u0026rsquo;s exact test, as appropriate. Continuous variables were compared using the Mann-Whitney U test or the independent-samples t test. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. The 95% confidence interval (CI) for the AUC was calculated using the auc function in the pROC R package. Differences in AUCs were compared using DeLong\u0026rsquo;s test, with P\u0026thinsp;\u0026lt;\u0026thinsp;0.05 considered statistically significant. All statistical analyses were performed using R software (version 4.3.3) and the scikit-learn library in Python (version 3.11.11).\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003ePatient Characteristics and Clinical Features\u003c/p\u003e\u003cp\u003eA total of 496 patients were included in this study. The clinical information and imaging characteristics of the patients are summarized in Table\u0026nbsp;1. In the training cohort, there were 227 males and 120 females, while the testing cohort included 93 males and 56 females. The mean age was 60.89\u0026thinsp;\u0026plusmn;\u0026thinsp;10.06 years in the training cohort and 59.28\u0026thinsp;\u0026plusmn;\u0026thinsp;9.72 years in the testing cohort.Among the clinical variables, hematuria showed a statistically significant difference between the training and testing cohorts (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Among the imaging features, intratumoral necrosis and the presence of tumor interlobular arteries were significantly different between the two cohorts (P\u0026thinsp;\u0026lt;\u0026thinsp;0.05). No significant differences were observed for age and calcification in the testing cohort, for sex in the training cohort, or for flank pain in either cohort.Univariate and multivariate analyses revealed that only age was significantly associated with pathological grade, while the remaining variables showed no significant associations (Table\u0026nbsp;2). A clinical model was subsequently developed based on these findings.\u003c/p\u003e\u003cp\u003ePerformance and Visualization of the Deep Learning Model\u003c/p\u003e\u003cp\u003eResNet50 demonstrates excellent stability and robustness in the development of medical imaging diagnostic models; therefore, this architecture was adopted for our deep learning analysis. As shown in Table\u0026nbsp;3 and Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA-B, among the different ROI configurations, the 5mm enlarged ROI achieved the best predictive performance. In the training cohort, the AUC was 0.791 (95% CI: 0.740\u0026ndash;0.842), with an accuracy of 0.775, sensitivity of 0.772, specificity of 0.833, PPV of 0.694, and NPV of 0.819. In the testing cohort, the AUC was 0.780 (95% CI: 0.700\u0026ndash;0.859), with an accuracy of 0.718, sensitivity of 0.745, specificity of 0.702, PPV of 0.594, and NPV of 0.825.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eModel interpretability was assessed using gradient-weighted class activation mapping (Grad-CAM), which visualized the spatial distribution of pixel importance in different colors, highlighting differences between high-grade and low-grade ccRCC tumors. Representative cases from the training cohort were examined. High-grade tumors generally exhibited larger activation regions, while low-grade tumors showed more limited activation. In most cases, the network demonstrated sensitivity to intratumoral heterogeneity and tumor margins in high-grade lesions (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eComparison of Combined Model Performance\u003c/p\u003e\u003cp\u003eDeep learning features were further extracted from the trained models and fused with arterial-phase radiomics features to construct combined models. After LASSO feature selection, the three combined models retained 24, 31, and 34 non-zero coefficient features, respectively (Supplement Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA-I). As shown in Supplement Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA-F and Supplement Table\u0026nbsp;1, although the conventional ROI slice based combined model achieved good performance in the training cohort, it showed lower AUC in the testing cohort, indicating a risk of overfitting. In contrast, the 5-mm expanded ROI slice\u0026ndash;based combined model demonstrated better overall performance compared to the other two models.\u003c/p\u003e\u003cp\u003eWhen comparing the three machine learning classifiers, logistic regression (LR), support vector machine (SVM), and random forest (RF). The RF classifier exhibited the most stable performance, with an AUC of 0.849 in the training cohort and 0.831 in the testing cohort. Although SVM achieved a higher AUC in the training cohort, its performance decreased in the testing cohort, suggesting potential overfitting. Therefore, the RF model was selected as the optimal classifier. Subsequently, imaging features from the optimal combined model were integrated with clinical variables to construct the clinical\u0026ndash;imaging combined model (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eA-B). In the training cohort, the combined model achieved an AUC of 0.855, and in the testing cohort, an AUC of 0.849. Its predictive ability was significantly higher than that of the clinical model alone and slightly better than the imaging combined model.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eDecision curve analysis (DCA) demonstrated (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eA-B) that the clinical-imaging combined model provided a higher net benefit across a range of threshold probabilities, indicating substantial potential for clinical application in ccRCC pathological grading.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eDeep learning has demonstrated substantial advantages in multimodal fusion analysis within medical imaging[\u003cspan additionalcitationids=\"CR34\" citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e]. By integrating complementary information from CT, MRI, PET, and other imaging modalities, deep learning frameworks can characterize lesions from multiple perspectives, enabling more precise diagnosis and supporting clinical decision-making in complex cases[\u003cspan additionalcitationids=\"CR37\" citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e]. The ability to extract hierarchical, nonlinear features from heterogeneous data streams allows such models to capture subtle morphological and functional patterns that may be imperceptible to traditional analysis[\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e]. As algorithms mature, datasets diversify, and clinical testing deepens, these approaches hold strong potential for advancing precision medicine and personalized treatment strategies[\u003cspan additionalcitationids=\"CR41\" citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e].\u003c/p\u003e\u003cp\u003eIn the present study, we developed a clinical-imaging combined model for ccRCC pathological grading by integrating deep learning features, radiomics features, and clinical variables. Our results demonstrated that the multi-channel model based on a 5-mm peritumoral expansion, when fused with radiomics features, achieved superior performance compared with models using conventional slices or tumor-only slices. The expanded ROI setting likely captures peritumoral microenvironmental cues, including subtle invasion patterns, microvascular proliferation, and perilesional heterogeneity, all of which are known to influence tumor aggressiveness. Furthermore, the expanded ROI reduces the risk of omitting informative peritumoral signals while simultaneously suppressing irrelevant background structures, thereby enhancing discriminative power. These findings underscore the importance of spatial context in noninvasive tumor grading.\u003c/p\u003e\u003cp\u003eOur work builds upon prior evidence supporting the utility of multi-dimensional imaging features for tumor classification. For example, Lin et al. [\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e] reported significant gains in renal tumor grading accuracy when combining radiomics with deep learning features. However, few previous studies have systematically compared multiple ROI ranges in conjunction with feature fusion, and even fewer have incorporated peritumoral expansion in a multi-phase CT setting. Yang et al. [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e] and Xu et al. [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e] demonstrated the predictive value of deep features extracted from contrast-enhanced imaging, particularly in fusion frameworks, but omitted the excretory CT phase. The excretory phase provides additional structural and functional information such as enhanced delineation of tumor parenchyma boundaries, assessment of delayed enhancement patterns, and indirect markers of renal function that can complement arterial and medullary phase data. By stacking arterial, medullary, and excretory phases into a three-channel input, our model captures a more comprehensive physiological and pathological profile, which proved beneficial for feature fusion and downstream classification.\u003c/p\u003e\u003cp\u003eOur findings also align with a recent systematic review comparing radiomics-only, deep learning-only, and fusion models, which reported that fusion models achieved the highest performance in 63% of included studies [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. Multi-channel imaging enables each channel to encode distinct aspects of tumor biology; when these channels are fused, the resulting feature space becomes richer and more robust to image noise, artifacts, or variability in a single phase. Moreover, this approach enhances the detection of subtle imaging patterns that may be overlooked by single-phase or handcrafted features alone.A key innovation of our approach lies in its dual-level integration: image-level fusion via multi-channel inputs and feature-level fusion with radiomics and clinical data. This hierarchical strategy enables the model to leverage low-level texture and intensity patterns while also incorporating handcrafted descriptors that may capture domain-specific cues. Compared with approaches relying solely on deep learning or radiomics, this hybrid method maximizes the complementary strengths of each while mitigating their individual limitations. The advantages of this design are further supported by our decision curve analysis, which demonstrated consistent net clinical benefit over a broad range of threshold probabilities, underscoring its potential for integration into routine diagnostic workflows.\u003c/p\u003e\u003cp\u003eWe observed a modest shortfall in sensitivity. This may be attributable to limited sample size and class imbalance specifically, the underrepresentation of high-grade tumors leading to a bias toward majority-class (low-grade) predictions and thus more false negatives for aggressive disease [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e]. Previous literature has shown that such imbalance can substantially degrade machine learning performance, particularly in high-stakes clinical prediction tasks. Additional factors, including noise, artifacts, and inter-scanner resolution variability, can further challenge feature extraction and model generalization. To mitigate these issues, future work should consider advanced sampling strategies, synthetic minority oversampling, cost-sensitive or focal loss functions, and targeted augmentation to improve sensitivity for high-grade tumors. Moreover, incorporating harmonization algorithms to minimize scanner-related variability and integrating multi-center datasets could enhance the robustness of the model.\u003c/p\u003e\u003cp\u003eFrom a clinical perspective, the integration of multiphasic CT-derived features with routine clinical data into a unified predictive framework is particularly appealing. Such models could serve as decision-support tools to identify high-risk patients, prioritize biopsy, or tailor surveillance intervals. Nevertheless, successful clinical adoption will require rigorous external testing across diverse populations and imaging systems, as well as the incorporation of interpretability tools capable of providing clinicians with both visual and quantitative explanations for model predictions. Future research could also investigate integration with complementary modalities such as MRI, histopathology, and genomics to further enrich the feature space and improve predictive accuracy.\u003c/p\u003e\u003cp\u003eSeveral limitations warrant acknowledgment. First, this was a single-center retrospective study, which may limit the generalizability of the findings. Differences in scanner hardware, acquisition protocols, and patient demographics across institutions could influence model performance in external settings. Second, although our dataset was of reasonable size, expanding the cohort, particularly increasing the proportion of high-grade tumors would enable more balanced training and reduce variability in performance estimates. Third, while our fusion framework effectively integrated deep learning and radiomics features, the optimal strategies for feature selection, fusion, and hyperparameter tuning remain to be determined. Advanced interpretability methods beyond Grad-CAM, such as attention-based saliency mapping or SHAP value analysis, could yield deeper insights into model decision-making and facilitate its acceptance in clinical practice.\u003c/p\u003e\u003cp\u003eIn summary, we present a multi-channel, multiphase CT based deep learning and radiomics fusion model for noninvasive ccRCC grading. By incorporating peritumoral expansion and leveraging complementary phase information, our framework generates a more expressive and robust feature representation, achieving strong predictive performance while maintaining clinical interpretability. Despite certain limitations, the model holds considerable promise as a decision-support tool for risk stratification and treatment planning. Future work should focus on multi-institutional testing, integration with additional modalities, and exploration of real-time deployment within clinical workflows.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eConceptualization: Chongyang sun;Qi chenMethodology: Chongyang sun;Qi chenSoftware: Ze zhangFormal Analysis: Chongyang sun;Qi chen; Meng gao; Shiqi HeInvestigation: Wei zhangData Curation: Chongyang sun;Wei zhangWriting \u0026ndash; Original Draft Preparation: Chongyang sun;Qi chenWriting \u0026ndash; Review \u0026amp; Editing: Chongyang sun;Qi chen; Meng gao; Shiqi HeVisualization: Xigang xiaoSupervision: Xigang xiaoFunding Acquisition: Qi chen\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLiao C, Hu L, Zhang Q. Von Hippel-Lindau protein signalling in clear cell renal cell carcinoma. Nat Rev Urol. 2024;21(11):662\u0026ndash;675.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSato Y, Yoshizato T, Shiraishi Y, et al. Integrated molecular analysis of clear-cell renal cell carcinoma. Nat Genet. 2013;45(8):860\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDeng H, Gong X, Ji G, Li C, Cheng S. KIF2C promotes clear cell renal cell carcinoma progression via activating JAK2/STAT3 signaling pathway. Mol Cell Probes. 72:101938.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao Z, Lu J, Qu H, et al. Evaluation and prognostic significance of manganese superoxide dismutase in clear cell renal cell carcinoma. Hum Pathol. 80:87\u0026ndash;93.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eIshigami K, Leite LV, Pakalniskis MG, Lee DK, Holanda DG, Kuehn DM. Tumor grade of clear cell renal cell carcinoma assessed by contrast-enhanced computed tomography. Springerplus. 3:694. Published 2014 None.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSoultati A, Stares M, Swanton C, Larkin J, Turajlic S. How should clinicians address intratumour heterogeneity in clear cell renal cell carcinoma? Curr Opin Urol. 2015;25(5):358\u0026ndash;66.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLopez JI, Cortes JM. A divide-and-conquer strategy in tumor sampling enhances detection of intratumor heterogeneity in routine pathology: A modeling approach in clear cell renal cell carcinoma. F1000Res. 5:385. Published 2016 None.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePratt S, Bowen I, Hallowell G, Shipman E, Redpath A. Assessment of agreement using the equine glandular gastric disease grading system in 84 cases. Vet Med Sci. 2022;8(4):1472\u0026ndash;1477.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePulvirenti A, Yamashita R, Chakraborty J, et al. Quantitative Computed Tomography Image Analysis to Predict Pancreatic Neuroendocrine Tumor Grade. JCO Clin Cancer Inform. 5:679\u0026ndash;694.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003ePark YW, Oh J, You SC, et al. Radiomics and machine learning may accurately predict the grade and histological subtype in meningiomas using conventional and diffusion tensor imaging. Eur Radiol. 2019;29(8):4068\u0026ndash;4076.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang WY, Wen LH, Wu G, et al. Comparison of Radiomics Analyses Based on Different Magnetic Resonance Imaging Sequences in Grading and Molecular Genomic Typing of Glioma. J Comput Assist Tomogr. 2021 Jan-Feb 01;45(1):110\u0026ndash;120.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGuo H, Tang HT, Hu WL, et al. The application of radiomics in esophageal cancer: Predicting the response after neoadjuvant therapy. Front Oncol. 13:1082960. Published 2023 None.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSumon RI, Mozumdar MAI, Akter S, et al. Comparative Study of Cell Nuclei Segmentation Based on Computational and Handcrafted Features Using Machine Learning Algorithms. Diagnostics (Basel). 2025;15(10). Published 2025 May 16.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJimenez-Del-Toro O, Aberle C, Bach M, et al. The Discriminative Power and Stability of Radiomics Features With Computed Tomography Variations: Task-Based Analysis in an Anthropomorphic 3D-Printed CT Phantom. Invest Radiol. 2021;56(12):820\u0026ndash;825.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMichallek F, Genske U, Niehues SM, Hamm B, Jahnke P. Deep learning reconstruction improves radiomics feature stability and discriminative power in abdominal CT imaging: a phantom study. Eur Radiol. 2022;32(7):4587\u0026ndash;4595.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBai Y, Zhao H, Shi X, Chen L. Towards practical and privacy-preserving CNN inference service for cloud-based medical imaging analysis: A homomorphic encryption-based approach. Comput Methods Programs Biomed. 261:108599.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAttique H, Shah S, Jabeen S, Khan FG, Khan A, ELAffendi M. Multiclass Cancer Prediction Based on Copy Number Variation Using Deep Learning. Comput Intell Neurosci. 2022:4742986. Published 2022 None.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eClarke EL, Magee D, Newton-Bishop J, et al. The Development and Evaluation of a Convolutional Neural Network for Cutaneous Melanoma Detection in Whole Slide Images. Arch Pathol Lab Med. Published online Dec 12,2024.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhao Q, Kong P, Min J, et al. [A review of deep learning methods for the detection and classification of pulmonary nodules]. Sheng Wu Yi Xue Gong Cheng Xue Za Zhi. 2019;36(6):1060\u0026ndash;1068.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKhan MSI, Rahman A, Debnath T, et al. Accurate brain tumor detection using deep convolutional neural network. Comput Struct Biotechnol J. 20:4733\u0026ndash;4745.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi S, Liu J, Song Z. Brain tumor segmentation based on region of interest-aided localization and segmentation U-Net. Int J Mach Learn Cybern. 2022;13(9):2435\u0026ndash;2445.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWei G, Fang G, Guo P, et al. Preoperative prediction of microvascular invasion risk in hepatocellular carcinoma with MRI: peritumoral versus tumor region. Insights Imaging. 2024;15(1):188.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMercan C, Balkenhol M, Salgado R, et al. Deep learning for fully-automated nuclear pleomorphism scoring in breast cancer. NPJ Breast Cancer. 2022;8(1):120. Published 2022 Nov 8.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLian J, Dong Z, Zhang H, Chen Y, Liu J. Multifocal region-assisted cross-modality learning for chest X-ray report generation. Comput Biol Med. 183:109187.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLu C, Zhang J, Zhang D, et al. Fine-Grained Hierarchical Progressive Modal-Aware Network for Brain Tumor Segmentation. IEEE J Biomed Health Inform. PP. Published online May 22,2025.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCoppenrath EM, Mueller-Lisse UG. Multidetector CT of the kidney. Eur Radiol. 2006;16(11):2603\u0026ndash;11.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi X, Guo Y, Huang S, et al. A CT-based intratumoral and peritumoral radiomics nomogram for postoperative recurrence risk stratification in localized clear cell renal cell carcinoma. BMC Med Imaging. 2025;25(1):167. Published 2025 May 16.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMao J, Li W, Sun X, et al. RPF-Net: A multimodal model for the postoperative UISS risk stratification of non-metastatic ccRCC based on CT and whole-slide images. Comput Methods Programs Biomed. 268:108836.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSim KC, Han NY, Cho Y, Sung DJ, Park BJ, Kim MJ, Han YE. Machine Learning-Based Magnetic Resonance Radiomics Analysis for Predicting Low- and High-Grade Clear Cell Renal Cell Carcinoma. J Comput Assist Tomogr. 2023 Nov-Dec 01;47(6):873\u0026ndash;881.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCui E, Li Z, Ma C, Li Q, Lei Y, Lan Y, Yu J, Zhou Z, Li R, Long W, Lin F. Predicting the ISUP grade of clear cell renal cell carcinoma with multiparametric MR and multiphase CT radiomics. Eur Radiol. 2020;30(5):2912\u0026ndash;2921.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSelvaraju RR, Cogswell M, Das A, Vedantam R, Parikh D, Batra D (2020) Grad-CAM: visual explanations from deep networks via gradient-based localization. International Journal of Com-puter Vision 128:336\u0026ndash;359\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eParmar C, Grossmann P, Bussink J, Lambin P, Aerts HJWL (2015) Machine learning methods for quantitative radiomic biomarkers. Sci Rep 5:13087\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu J, Zeng B, Egger J, et al. A review on AI-based medical image computing in head and neck surgery. Phys Med Biol. 2022;67(17).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDana J, Agnus V, Ouhmich F, Gallix B. Multimodality Imaging and Artificial Intelligence for Tumor Characterization: Current Status and Future Perspective. Semin Nucl Med. 2020;50(6):541\u0026ndash;548.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eVairetti C, Maldonado S, Cuitino L, Urzua CA. Interpretable multimodal classification for age-related macular degeneration diagnosis. PLoS One. 2024;19(11):e0311811.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eQu J, Huang D, Shi Y, Liu J, Tang W. Diffusion-driven multi-modality medical image fusion. Med Biol Eng Comput. 2025;63(7):2105\u0026ndash;2118.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu R, Liu Y, Wang H, Li J, Hu K. SAFFusion: a saliency-aware frequency fusion network for multimodal medical image fusion. Biomed Opt Express. 2025;16(6):2459\u0026ndash;2481.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWei L, Osman S, Hatt M, El Naqa I. Machine learning for radiomics-based multimodality and multiparametric modeling. Q J Nucl Med Mol Imaging. 2019;63(4):323\u0026ndash;338.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSmedley NF, Hsu W. USING DEEP NEURAL NETWORKS FOR RADIOGENOMIC ANALYSIS. Proc IEEE Int Symp Biomed Imaging. 2018:1529\u0026ndash;1533.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShams A. Leveraging State-of-the-Art AI Algorithms in Personalized Oncology: From Transcriptomics to Treatment. Diagnostics (Basel). 2024;14(19).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eStetson LC, Pearl T, Chen Y, Barnholtz-Sloan JS. Computational identification of multi-omic correlates of anticancer therapeutic response. BMC Genomics. 15 Suppl 7:S2.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKayikcioglu E, Onder AH, Bacak B, Serel TA. Machine learning for predicting colon cancer recurrence. Surg Oncol. 54:102079.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLin F, Ma C, Xu J, Lei Y, Li Q, Lan Y, Sun M, Long W, Cui E. A CT-based deep learning model for predicting the nuclear grade of clear cell renal cell carcinoma. Eur J Radiol. 2020;129:109079.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eYang M, He X, Xu L, Liu M, Deng J, Cheng X, Wei Y, Li Q, Wan S, Zhang F, Wu L, Wang X, Song B, Liu M. CT-based transformer model for non-invasively predicting the Fuhrman nuclear grade of clear cell renal cell carcinoma. Front Oncol. 2022;12:961779.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu L, Yang C, Zhang F, Cheng X, Wei Y, Fan S, Liu M, He X, Deng J, Xie T, Wang X, Liu M, Song B. Deep Learning Using CT Images to Grade Clear Cell Renal Cell Carcinoma: Development and Validation of a Prediction Model. Cancers (Basel). 2022;14(11):2574.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDemircio˘glu A. Are deep models in radiomics performing better than generic models? A systematic review. Eur Radiol Exp.2023;7(1):11.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eXu L, Yang C, Zhang F, Cheng X, Wei Y, Fan S, Liu M, He X, Deng J, Xie T, Wang X, Liu M, Song B. Deep Learning Using CT Images to Grade Clear Cell Renal Cell Carcinoma: Development and Validation of a Prediction Model. Cancers (Basel). 2022;14(11):2574.\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":"","lastPublishedDoi":"10.21203/rs.3.rs-7430046/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7430046/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePurposes: To develop a combined model integrating multi-channel deep learning features, radiomics features, and clinical variables for noninvasive pathological grading of clear cell renal cell carcinoma (ccRCC).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMaterial and methods: A retrospective study was conducted on 496 patients with pathologically confirmed ccRCC who underwent preoperative triple-phase contrast-enhanced CT. Multi-channel deep learning features were extracted from three ROI settings (conventional, tumor-only, and 5-mm expansion) by stacking arterial, medullary, and excretory phases. These were fused with arterial-phase radiomics features and clinical data to construct and compare predictive models.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eResults: In the ResNet50 model, the expanded 5mm ROI slice model had an AUC of 0.791 in the training group and 0.780 in the test group, indicating that the model could effectively predict the pathological grading of ccRCC. By combining deep learning features with radiomics and clinical features, the integrated model achieved AUCs of 0.855 in the training group and 0.849 in the test group, significantly outperforming the individual radiomics and clinical models. Decision curve analysis (DCA) further showed that the clinical-imaging combined model provided a higher net benefit.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eConclusion: Multi-channel, multiphase CT fusion, when integrated with radiomics and clinical features, can significantly enhance predictive accuracy for ccRCC grading, providing a promising and interpretable noninvasive tool to support individualized treatment planning.\u003c/p\u003e","manuscriptTitle":"Multi-Channel Multiphase CT-Based Deep Learning and Radiomics Fusion Model for Noninvasive Pathological Grading of Clear Cell Renal Cell Carcinoma","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-08 08:29:25","doi":"10.21203/rs.3.rs-7430046/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":"fc877a3c-a32d-4f29-8a7b-8d475df8bca7","owner":[],"postedDate":"September 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-09-10T00:53:17+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-08 08:29:25","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7430046","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7430046","identity":"rs-7430046","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}