Weakly Supervised Learning for Multi-class RCC Classification: Multicenter Validation with Biological Interpretability

Weakly Supervised Learning for Multi-class RCC Classification: Multicenter Validation with Biological Interpretability | 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 Article Weakly Supervised Learning for Multi-class RCC Classification: Multicenter Validation with Biological Interpretability Chengwei Chen, Bing Xia, Qinqin Kang, Na Li, Qianru Zhang, Yixuan Shen, and 15 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7282796/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 Accurate renal cell carcinoma (RCC) subtyping, especially challenging TFE3-rearranged RCC, is vital for treatment. We developed RCCNET (RCC Neural Enhancement Technology), a weakly supervised deep learning framework integrating a parallel cellular morphometric module for biological interpretability, for four-class classification (clear cell, papillary, chromophobe, TFE3-rearranged). Validated multicentrically on 340 patients (training n=233; external validation n=107), RCCNET achieved macro-average AUCs of 0.989 (training) and 0.966 (validation). For TFE3 RCC, AUC was 0.976 with 92.3% sensitivity, but a 66.7% positive predictive value necessitates molecular confirmation of all positive cases. Model predictions significantly correlated with quantitative morphological features, grounding decisions in histopathology. An economic analysis projected an RCCNET-assisted workflow could reduce costs by 83.2% and time by 45.2%. RCCNET provides an interpretable, cost-effective solution. We propose a confidence-based clinical integration framework, flagging uncertain TFE3 predictions for pathologist review to manage false positives and ensure safe deployment. Biological sciences/Cancer Biological sciences/Computational biology and bioinformatics Health sciences/Oncology Health sciences/Urology Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Renal cell carcinoma represents a heterogeneous group of malignancies with distinct morphological, molecular, and clinical characteristics that fundamentally shape therapeutic decisions and patient outcomes 1 . The complexity of accurate subtype classification has intensified with the 2022 World Health Organization classification, which recognizes over 20 distinct RCC entities, each requiring precise identification for optimal patient management 1 . Among these, clear cell RCC, papillary RCC, chromophobe RCC, and TFE3-rearranged RCC constitute the most clinically significant subtypes, yet their accurate differentiation remains one of the most challenging aspects of contemporary urological pathology 2 , 3 . The diagnostic complexity reaches its zenith with TFE3-rearranged RCC, an entity that exemplifies the intersection of morphological heterogeneity and molecular precision medicine 2 . Characterized by translocations involving the TFE3 gene, this subtype presents a morphological spectrum that can convincingly mimic clear cell, papillary, or even unclassified RCC patterns. This diagnostic ambiguity necessitates molecular confirmation through fluorescence in situ hybridization or next-generation sequencing, creating a cascade of increased costs, extended diagnostic timelines, and resource allocation challenges that particularly burden institutions with limited molecular pathology capabilities 2 . Traditional diagnostic paradigms, while foundational to pathological practice, face mounting pressures from several converging factors 4 , 5 . Inter-observer variability among pathologists, even those with subspecialty expertise, introduces diagnostic uncertainty that can directly impact patient care 6 . Resource constraints, particularly in molecular testing capabilities, create disparities in diagnostic access across different healthcare systems 7 . Perhaps most critically, the increasing complexity of RCC classification demands a level of subspecialty expertise that may not be readily available in all institutions, creating potential gaps in diagnostic accuracy and consistency 5 . The emergence of artificial intelligence in pathology represents a valuable technological advancement toward objective, reproducible, and scalable diagnostic solutions 3 , 4 . Recent developments in deep learning, particularly in computer vision and weakly supervised learning approaches, have demonstrated promising results in medical image analysis 5 , 8 . However, the translation of these technological capabilities to the specific challenges of RCC subtype classification has been limited by several factors: most existing approaches focus on binary classification tasks that oversimplify the clinical reality, lack comprehensive validation across diverse institutional settings, and fail to address the critical need for interpretability that is essential for clinical acceptance and regulatory approval 9 , 10 . To simultaneously address the dual challenges of diagnostic accuracy and clinical interpretability, we developed RCCNET (Renal Cell Carcinoma Neural Enhancement Technology), a comprehensive computational framework that represents a meaningful advance over conventional approaches 4 , 7 . Our innovation lies not merely in achieving high classification accuracy, but in creating a system that provides morphometric correlates for its decisions through quantitative cellular analysis. This dual-stream architecture addresses both the immediate clinical need for accurate subtype classification and the longer-term requirement for transparent, interpretable computational systems that can gain the trust and acceptance of the pathology community. Methods Study Design and Reporting Guidelines Adherence This investigation was conceived and executed as a rigorous multicenter diagnostic accuracy study, adhering to the Standards for Reporting of Diagnostic Accuracy Studies (STARD) 2015 guidelines 11 and the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis + Artificial Intelligence (TRIPOD + AI) statement 12 . The study protocol received approval from the Ethics Committees of both participating institutions and was conducted in strict accordance with the Declaration of Helsinki principles (Supplementary Method 1 and Figure S1 ). Study Setting and Participants This retrospective, multicenter study enrolled consecutive patients from two tertiary Grade A hospitals in China to ensure real-world generalizability. The training cohort was derived from XX Hospital (a major referral center with advanced digital pathology) for patients treated between June 2013 and December 2021. The external validation cohort was sourced from XXX Hospital (a center with standard pathology practices) for patients treated between January 2021 and June 2023. Inclusion criteria were: a history of radical or partial nephrectomy for a renal mass; a final histopathological diagnosis of clear cell, papillary, chromophobe, or TFE3-rearranged RCC according to the 2022 WHO classification; and available FFPE tissue with high-quality H&E slides 1 . Cases were excluded for mixed or uncertain histology, extensive necrosis (> 50%), significant artifacts, inadequate tumor tissue (< 75% tumor content or < 1 cm² area), or lack of molecular confirmation for TFE3-rearranged RCC 2 . Reference Standard and Molecular Confirmation The reference standard for RCC subtype classification was established through consensus review by two expert genitourinary pathologists, each with more than 10 years of subspecialty experience, using the 2022 WHO classification criteria 1 . This approach ensured diagnostic consistency while maintaining the highest standards of pathological expertise. All TFE3-rearranged RCC cases underwent molecular confirmation through fluorescence in situ hybridization using TFE3 break-apart probes or RNA sequencing when available, ensuring the diagnostic accuracy essential for training and validating our computational system 2 . RCCNET Framework Architecture Primary Classification Module (Stream 1) The architecture of the RCCNET framework is detailed in Supplementary Method 3 and Fig. 1 , with the full digital pathology workflow described in Supplementary Method 2 . The primary classification module (Stream 1) employs a sophisticated weakly supervised multiple instance learning approach, designed to process whole-slide images without requiring the pixel-level annotations that have historically limited computational development 13 , 14 . This choice circumvents the practical bottleneck of creating detailed annotations, especially for rare entities like TFE3 RCC 14 . The pipeline begins by dividing each WSI (approximately 100,000 × 50,000 pixels) into non-overlapping tiles measuring 256 µm × 256 µm, corresponding to 224 × 224-pixel patches at 0.5 µm/pixel. From these, a CTransPath encoder—chosen for its superior performance on histopathological images over general-purpose models—extracts 768-dimensional feature vectors for each tile. A Swin Transformer backbone then models the spatial dependencies and architectural patterns across the tissue 7 , capturing both local and global features to address the multi-scale nature of pathological diagnosis 15 . Finally, an attention-based aggregation mechanism weights each tile's contribution to the classification decision and generates interpretable attention heatmaps to highlight diagnostically relevant regions 5 , 7 . Figure 1. Overall Study Workflow Morphometric Analysis Module (Stream 2) To address interpretability challenges in computational pathology, the Morphometric Analysis Module (Stream 2) was designed to provide quantitative, biologically meaningful correlates for the model's classification decisions ( Supplementary Method 4 ). Nuclear segmentation was performed using the HoVer-Net model to classify individual nuclei into three categories (neoplastic, inflammatory, and dead cells) 16 , and its performance was rigorously validated on 200 annotated tissue regions to ensure the reliability of downstream analysis 8 . For each nucleus, we computed 11 morphometric descriptors across categories of size (e.g., area, volume), shape (e.g., roundness, eccentricity) 3 , and orientation (e.g., major/minor axes) 17 . Finally, five statistical descriptors (including mean, standard deviation, and entropy) were calculated for these features to transform qualitative pathological observations into quantitative, reproducible measurements for systematic comparison. Training and Validation Strategy Model training was conducted on the XX Hospital cohort (n = 233) using 5-fold cross-validation for hyperparameter optimization and to prevent overfitting. To address class imbalance, particularly for TFE3 RCC (13.3% of training data), we implemented a multi-faceted strategy including focal loss (γ = 2.0, α = 0.25), class-weighted sampling, and enhanced data augmentation. The model's generalizability was then assessed on the completely independent XXX Hospital cohort (n = 107) without any retraining or fine-tuning, providing an unbiased assessment. Performance Metrics and Clinical Integration Framework Primary diagnostic performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and predictive values for each subtype. Based on the model's confidence score distribution and error analysis, we developed a three-tier clinical integration framework to balance accuracy with practical implementation: high-confidence (> 0.8) cases for minimal review, moderate-confidence (0.4–0.8) cases for guided pathologist review, and low-confidence (< 0.4) cases directed to traditional workflows with molecular testing as indicated. Economic Analysis A comprehensive cost-effectiveness analysis, with detailed modeling assumptions provided in Supplementary Method 6 , was conducted to compare the RCCNET-assisted workflow with traditional diagnosis, incorporating both direct and indirect costs for a realistic assessment of economic impact 18 . Statistical Analysis Continuous and categorical variables were described and compared using appropriate statistical tests (e.g., Student's t-test, Mann-Whitney U test, chi-square test). Model performance was evaluated using ROC analysis, with DeLong's test for AUC comparisons. Morphometric correlations were assessed using Spearman correlation with Bonferroni correction (α = 0.0003), and misclassification analysis included effect size calculation (Cohen's d). Economic analysis employed decision tree modeling. A p-value < 0.05 was considered statistically significant. All analyses were performed using Python 3.11.0 and R 4.3.0, with technical details provided in Supplementary Method 5 . Results Cohort Characteristics and Baseline Demographics The study included 340 patients from two institutions, divided into a training cohort (n = 233) and a validation cohort (n = 107) ( Fig. 2 ) . Both cohorts showed a similar, representative distribution of RCC subtypes: the training set included 118 clear cell (ccRCC), 56 papillary (pRCC), 28 chromophobe (chRCC), and 31 TFE3-rearranged (TFE3 RCC) cases, while the validation set comprised 56, 26, 12, and 13 cases of each subtype, respectively. Baseline demographics and tumor characteristics were well-balanced between the cohorts, with no significant differences in age, sex, or tumor stage distribution. Tumor characteristics showed similar distributions, with the majority of cases presenting as Stage I disease (training: 68.2%, validation: 71.0%), providing a representative sample of contemporary RCC surgical practice ( Table 1 ) . Table 1 Baseline Demographic and Clinicopathological Characteristics Across RCC Subtypes Characteristics Train Group Validation Group pRCC TFE3-RCC ccRCC chRCC P value pRCC TFE3-RCC ccRCC chRCC P value Number, n 56 31 118 28 26 13 56 12 Age, years (Mean ± SD) 53.8 ± 14.2 52.1 ± 13.5 54.9 ± 13.6 55.2 ± 14.1 0.342 54.2 ± 12.8 53.8 ± 11.9 56.1 ± 12.2 57.3 ± 13.1 0.445 Sex, n (%) 0.721 0.689 Male 35 (62.5) 19 (61.3) 78 (66.1) 19 (67.9) 16 (61.5) 8 (61.5) 36 (64.3) 7 (58.3) Female 21 (37.5) 12 (38.7) 40 (33.9) 9 (32.1) 10 (38.5) 5 (38.5) 20 (35.7) 5 (41.7) BMI (kg/m²± SD) 24.35 ± 3.19 24.60 ± 3.89 24.56 ± 3.44 24.49 ± 3.84 0.732 23.40 ± 3.20 23.64 ± 3.35 24.54 ± 3.44 24.01 ± 4.65 0.732 Tumor Stage, n (%) 0.654 0.598 Stage I 38 (67.9) 22 (71.0) 80 (67.8) 19 (67.9) 18 (69.2) 9 (69.2) 41 (73.2) 8 (66.7) Stage II 10 (17.9) 5 (16.1) 21 (17.8) 5 (17.9) 4 (15.4) 2 (15.4) 9 (16.1) 2 (16.7) Stage III 7 (12.5) 3 (9.7) 15 (12.7) 3 (10.7) 3 (11.5) 2 (15.4) 5 (8.9) 1 (8.3) Stage IV 1 (1.8) 1 (3.2) 2 (1.7) 1 (3.6) 1 (3.8) 0 (0.0) 1 (1.8) 1 (8.3) Tumor Grade, n (%) 0.445 0.512 Grade 1 11 (19.6) 6 (19.4) 23 (19.5) 5 (17.9) 4 (15.4) 2 (15.4) 10 (17.9) 2 (16.7) Grade 2 27 (48.2) 15 (48.4) 56 (47.5) 14 (50.0) 14 (53.8) 7 (53.8) 29 (51.8) 6 (50.0) Grade 3 14 (25.0) 8 (25.8) 30 (25.4) 6 (21.4) 6 (23.1) 3 (23.1) 13 (23.2) 3 (25.0) Grade 4 4 (7.1) 2 (6.5) 9 (7.6) 3 (10.7) 2 (7.7) 1 (7.7) 4 (7.1) 1 (8.3) Tumor Size, cm [Median (IQR)] 4.0 (2.8–5.8) 4.3 (3.2–6.2) 4.2 (3.1-6.0) 4.5 (3.3–6.5) 0.523 3.8 (2.6–5.6) 4.1 (2.9–5.9) 4.0 (2.9–5.8) 4.3 (3.1–6.1) 0.467 Primary Classification Performance RCCNET demonstrated strong diagnostic performance, achieving a macro-averaged AUC of 0.989 (95% CI: 0.985–0.993) in the training cohort and showing robust generalization with an AUC of 0.966 (95% CI: 0.951–0.981) in the external validation cohort. In validation, subtype-specific performance remained high ( Table 2 and Fig. 3 ) . The model effectively identified clear cell RCC (AUC 0.972, sensitivity 96.4%), papillary RCC (AUC 0.951, sensitivity 84.6%), and chromophobe RCC (AUC 0.979, sensitivity 83.3%). For the diagnostically challenging TFE3 RCC, the model achieved an AUC of 0.976 and 92.3% sensitivity. However, this was tempered by a critical limitation: the positive predictive value was only 66.7% (12 of 18 predictions), indicating a one-third false-positive rate that necessitates molecular confirmation due to significant clinical implications. Model visualization for diagnosis is shown in Figs. 4 and Supplementary Figure S2 . Table 2 RCCNET Performance Metrics Across Training and Validation Cohorts RCC Subtype n AUC (95% CI) Sensitivity (%) Specificity (%) PPV (%) NPV (%) Accuracy (%) Training Set Clear Cell RCC 118 0.991 (0.985–0.997) 96.6 (114/118) 95.7 (110/115) 95.0 (114/119) 97.0 (110/114) 96.1 (224/233) Papillary RCC 56 0.988 (0.979–0.997) 94.6 (53/56) 97.2 (172/177) 91.4 (53/58) 98.3 (172/175) 96.6 (225/233) Chromophobe RCC 28 0.995 (0.990-1.000) 92.9 (26/28) 99.0 (203/205) 92.9 (26/28) 99.0 (203/205) 98.3 (229/233) TFE3 RCC 31 0.989 (0.981–0.997) 90.3 (28/31) 98.0 (198/202) 87.5 (28/32) 98.5 (198/201) 97.0 (226/233) Macro-averaged 233 0.989 (0.985–0.993) 93.6 (218/233) 97.5 (681/699) 91.7 (218/238) 98.2 (681/694) 97.4 (227/233) Validation Set Clear Cell RCC 56 0.972 (0.951–0.993) 96.4 (54/56) 82.4 (42/51) 85.7 (54/63) 95.5 (42/44) 89.7 (96/107) Papillary RCC 26 0.951 (0.912–0.990) 84.6 (22/26) 95.1 (77/81) 84.6 (22/26) 95.1 (77/81) 92.5 (99/107) Chromophobe RCC 12 0.979 (0.958-1.000) 83.3 (10/12) 98.9 (94/95) 90.9 (10/11) 97.9 (94/96) 97.2 (104/107) TFE3 RCC 13 0.976 (0.943-1.000) 92.3 (12/13) 93.6 (88/94) 66.7 (12/18) 98.9 (88/89) 93.5 (100/107) Macro-averaged 107 0.966 (0.951–0.981) 89.2 (91/107) 92.5 (301/321) 81.9 (91/111) 96.9 (301/317) 85.0 (91/107) Note : Numbers in parentheses for validation cohort represent actual counts (true positives/total positives for sensitivity; true negatives/total negatives for specificity; true positives/predicted positives for PPV; true negatives/predicted negatives for NPV). AUC = area under the receiver operating characteristic curve; CI = confidence interval; PPV = positive predictive value; NPV = negative predictive value; RCC = renal cell carcinoma. Critical Clinical Note: The positive predictive value for TFE3 RCC in validation is 66.7%, indicating that one-third of cases predicted as TFE3 RCC are false positives. Molecular confirmation is essential for all positive TFE3 RCC predictions. A DeLong test confirmed the model's robustness, showing no significant differences in individual subtype AUCs between cohorts (p > 0.05 for all). Although the macro-average AUC showed a statistical difference (p = 0.045), the small effect size (95% CI: 0.001–0.045) reinforces the model's strong performance in external validation (Supplementary Table S1 ) . Nuclear Segmentation Performance and Morphometric Analysis Foundation The reliability of our morphometric analysis depended on accurate nuclear segmentation, which achieved high F1-scores for neoplastic (0.90 ± 0.04), inflammatory (0.83 ± 0.06), and dead cells (0.77 ± 0.09), establishing a robust foundation comparable to published benchmarks ( Supplementary Table S2 and Supplementary Figure S3 ). Segmentation performance varied across RCC subtypes, reflecting their distinct morphologies. Performance was highest in clear cell RCC due to its clear cytoplasm and contrast, and the model was particularly effective at identifying the elongated neoplastic cells of papillary RCC. Similarly, the uniform cell size and distinct borders of chromophobe RCC contributed to excellent detection performance. In contrast, the model’s moderate performance on TFE3 RCC mirrored this subtype's known morphological heterogeneity, which is also a diagnostic challenge for pathologists. Cellular Morphometric Analysis and Distribution Patterns Morphometric analysis revealed key quantitative differences among RCC subtypes, validating established pathological knowledge with new, measurable insights. For instance, nuclear area measurements followed subtype-specific log-normal distributions, and size heterogeneity was, consistent with pathological descriptions, highest in clear cell RCC and lowest in chromophobe RCC. Features such as cellular roundness clearly distinguished neoplastic from inflammatory populations, while eccentricity was highest in papillary RCC, reflecting its characteristic elongated morphology. Cellular compartment analysis provided further biological insights ( Supplementary Table S3 ). Neoplastic cells, comprising 45–75% of total nuclei, showed the greatest diversity in clear cell RCC and the most uniformity in chromophobe RCC. Notably, the proportion of inflammatory cells (15–35% of nuclei) was significantly reduced in TFE3-rearranged RCC, providing quantitative support for its described "immune-cold" phenotype, while dead cells accounted for a highly variable 5–15% of nuclei. Investigating Biological Correlates Through Computational-Morphometry Correlation Analysis To ensure RCCNET's accuracy was grounded in true histopathological features, we performed a post-hoc Spearman correlation analysis linking model predictions to quantitative cellular morphometrics. This analysis confirmed the model learned clinically relevant patterns Supplementary Figure S4 and Supplementary Table S4 ). For papillary RCC, predictions strongly correlated with features of its elongated cellular morphology (tumor cell roundness kurtosis: r = -0.509, p < 0.001) and complex architecture (tumor cell eccentricity entropy: r = 0.445, p < 0.001). The model identified the "immune-cold" phenotype of TFE3 RCC, showing negative correlations with inflammatory cell count (r = -0.270, p < 0.001) and volume variance (r = -0.245, p < 0.001). Furthermore, the model captured the known heterogeneity of clear cell RCC (positive correlation with eccentricity entropy: r = 0.515, p < 0.001) and the uniformity of chromophobe RCC (negative correlation with perimeter kurtosis: r = -0.472, p < 0.001). Collectively, these results provide compelling evidence that RCCNET's classifications are based on measurable, biologically meaningful features aligned with established pathology, thus enhancing its interpretability (Supplementary Figure S5 ). Misclassification Analysis and Clinical Integration Strategy Analysis of the 16 misclassified validation cases (15.0% error rate) revealed that errors typically occurred in morphologically ambiguous cases, with clear cell RCC being misclassified as TFE3 RCC most frequently (6 cases) ( Supplementary Figure S6 and Table S4 ). These false positive TFE3 predictions were predominantly known mimics, such as ccRCC with eosinophilic features, underscoring the need for molecular confirmation. Critically, RCCNET demonstrated a key safety feature, as confidence scores for misclassified cases were significantly lower than for correct predictions (mean 0.56 vs. 0.89, respectively; p 0.8 score; 75% of the cohort; 95.0% accuracy) are suitable for minimal review; 2) moderate-confidence cases (0.4–0.8 score; 21% of the cohort; 81.8% accuracy) are flagged for guided pathologist review ; and 3) low-confidence cases (< 0.4 score; 5% of the cohort; 60.0% accuracy) are directed to traditional workflows. Economic Impact Analysis Our comprehensive economic analysis revealed that the RCCNET-assisted workflow could yield substantial benefits by reducing per-case costs by 83.2% (from CNY 2,847 to CNY 479) and diagnostic time by 45.2% (from 58.5 to 32.0 minutes), significantly impacting laboratory throughput (Supplementary Table S7) . These efficiencies were primarily driven by projected decreases in the use of routine immunohistochemistry (70.6% reduction) and unnecessary molecular tests (44.4% reduction) for most RCC subtypes. However, these potential savings must be balanced against the clinical necessity of molecular confirmation for all cases predicted as TFE3 RCC, given the model's 66.7% positive predictive value for this specific subtype. Discussion This study presents RCCNET as a valuable advance toward a comprehensive, interpretable diagnostic system that addresses the complex realities of contemporary renal cell carcinoma (RCC) classification 5 . Our results demonstrate that accurate multi-class RCC subtype classification is achievable using a weakly supervised approach that circumvents the need for pixel-level annotations 19 . The framework shows particular strength for the diagnostically challenging TFE3-rearranged subtype, though we also highlight that careful attention to its false positive rate is essential for safe clinical implementation 2 . The strong performance of RCCNET, particularly for TFE3 RCC (AUC 0.976, sensitivity 92.3%), represents a meaningful advance addressing a critical diagnostic gap 2 . Unlike previous studies focusing on binary classification 20 , 21 , our multi-class framework directly addresses real-world diagnostic workflows where pathologists must simultaneously distinguish among multiple RCC entities 3 . While other groups have pioneered computational analysis for TFE3 RCC, our annotation-free method overcomes key scalability barriers, enabling broader clinical deployment 19 . However, this must be balanced against the 66.7% positive predictive value for TFE3 RCC 20 , which requires careful clinical interpretation. This finding indicates that one-third of positive TFE3 predictions are false positives—predominantly involving clear cell RCC with eosinophilic features or papillary RCC with solid patterns. This limitation necessitates molecular confirmation for positive TFE3 RCC predictions, particularly those with moderate confidence scores, and emphasizes the importance of integrating computational tools with traditional workflows rather than replacing them entirely 9 , 22 . Despite this, the model's robust generalization across institutions (macro-averaged AUC 0.966) validates its clinical applicability in diverse settings. A key innovation of this study is the integration of cellular morphometric analysis, which provides biological correlates for the model's predictions and addresses the fundamental "black box" challenge of interpretability in pathology 18 , 23 . The notable correlations between prediction probabilities and quantitative morphological features provide compelling evidence that our model learns clinically relevant patterns 18 . For example, the strong correlation between papillary RCC predictions and tumor cell eccentricity (r = 0.445) directly reflects the elongated cellular morphology that pathologists recognize as characteristic of this subtype 3 . Similarly, the reduced inflammatory infiltration found in TFE3 RCC cases (r = -0.270) provides quantitative support for its "immune-cold" phenotype, contributing new insights into RCC biology. While this post-hoc correlation does not establish causation, it provides strong support for the biological plausibility of the model's decisions. For clinical translation, we propose a confidence-based integration framework that provides a roadmap for safe, graduated implementation while maintaining appropriate human oversight 22 . This strategy is viable because of the model's intrinsic ability to identify its own limitations; misclassified cases had significantly lower confidence scores than correct ones (0.56 vs 0.89, p < 0.001), enabling workflows that automatically flag uncertain cases for expert review 9 . Our three-tier system stratifies cases for either minimal review (high confidence: 75% of cases, 95.0% accuracy), guided pathologist review (moderate confidence: 21% of cases, 81.8% accuracy), or traditional workup (low confidence: 5% of cases, 60.0% accuracy) 22 . This approach, combined with substantial projected cost (83.2%) and time (45.2%) savings, presents a practical pathway for implementation, although the final economic advantages must be balanced against the necessary costs of TFE3 molecular confirmation 7 , 13 . This work has direct implications for making precision medicine more accessible, as RCCNET can help guide molecular testing strategies, particularly in resource-limited institutions. Future work should focus on exploring dynamic models that integrate molecular data and performing mechanistic studies to move beyond correlation toward causation. Several limitations also warrant acknowledgment. Our study was limited to four common RCC subtypes and the Chinese population, and its 2D-based morphometrics cannot fully capture three-dimensional nuclear architecture 17 . Most importantly, the 33.3% false positive rate and small validation sample size (n = 13) for TFE3 RCC are critical limitations that mandate caution. This reinforces that for rare and challenging subtypes, computational tools should serve to augment, not replace, expert pathological diagnosis and molecular confirmation 9 , 22 . Conclusion In conclusion, RCCNET represents a significant advance in computational RCC diagnosis by successfully combining technical innovation, morphometric correlation analysis, and a practical clinical integration strategy. The framework’s strong performance, robust cross-institutional generalization, and demonstrated economic benefits provide compelling evidence for its potential clinical impact, provided its limitations—particularly the false positive rate for TFE3 RCC—are carefully managed in a clinical setting. Our study demonstrates that comprehensive computational frameworks can address complex diagnostic challenges while maintaining the transparency and safety required for clinical practice. Ultimately, this work supports a vision of human-computational collaboration that enhances diagnostic capabilities and provides new insights into disease biology, while reinforcing the continued importance of expert pathological interpretation in ensuring optimal patient care. Declarations Disclosures: The authors declare no potential conflicts of interest. Ethics Approval and Consent to Participate This study was approved by the Institutional Review Boards of both participating institutions (Ethics Approval No.: CHEC-Y2024-077). Due to the retrospective nature of the study using de-identified data, the requirement for informed consent was waived by the ethics committees in accordance with local regulations and international guidelines for retrospective medical research. Consent for Publication Not applicable for this retrospective study using de-identified patient data. Competing Interests The authors declare that they have no competing interests, financial or otherwise, that could have influenced the design, conduct, or reporting of this study. Funding This work was supported in part by the National Science Foundation for Scientists of China (81871352, 82171915, 82171930, 82271972 and 82371955), The Natural Science Foundation of Shanghai Science and Technology Innovation Action Plan (21ZR1478500, 21Y11910300), and Clinical Research Plan of SHDC (SHDC2022CRD028), Shanghai Municipal Health Commission Seed Program for Research and Translation of Medical New Technologies Project (2024ZZ1015), and Plan for Promoting Scientific Research Paradigm Reform and Enhancing Disciplinary Advancement through Artificial Intelligence (2024RGZD001). Author Contribution C.C., B.X., Q.K., and N.L. designed the study and contributed equally to conceptual development.C.C., Q.Z., Y.S., D.Z., S.M., Y.Z., J.Y., and J.L. collected and curated the pathology and imaging data.X.F., F.L., and M.Y. developed the AI framework and conducted the computational analysis.T.W., L.W., and J.L. contributed to model interpretation and technical validation.S.Z. supervised pathology annotation and morphological interpretation.H.J., C.S., and Y.B. jointly supervised the project and provided critical revisions.C.C., B.X., and Y.B. wrote the initial draft.All authors reviewed and approved the final manuscript. Availability of Data and Materials The datasets used and analyzed during the current study are available from the corresponding author on reasonable request, subject to institutional data sharing policies, patient privacy regulations, and appropriate data use agreements. The AI models and analysis code are publicly available as described in the Data and Code Availability Statement. https://github.com/CHANGHAI-AILab/RCCNET References Goswami, P. R., Singh, G., Patel, T. & Dave, R. The WHO 2022 Classification of Renal Neoplasms (5th Edition): Salient Updates. Cureus 16 , e58470 (2024). https://doi.org:10.7759/cureus.58470 Pagarigan, A. K. L., Reyes-Murillo, P. D. & Carbonell, D. J. S. Approach to Diagnosis of TFE3-rearranged Renal Cell Carcinoma in a Limited Resource Setting: A Case Report. J Kidney Cancer VHL 11 , 40-44 (2024). https://doi.org:10.15586/jkcvhl.v11i3.338 Hsieh, J. J. et al. Renal cell carcinoma. Nat Rev Dis Primers 3 , 17009 (2017). https://doi.org:10.1038/nrdp.2017.9 Janowczyk, A. & Madabhushi, A. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. J Pathol Inform 7 , 29 (2016). https://doi.org:10.4103/2153-3539.186902 Verghese, G. et al. Computational pathology in cancer diagnosis, prognosis, and prediction - present day and prospects. J Pathol 260 , 551-563 (2023). https://doi.org:10.1002/path.6163 Abu Haeyeh, Y., Ghazal, M., El-Baz, A. & Talaat, I. M. Development and Evaluation of a Novel Deep-Learning-Based Framework for the Classification of Renal Histopathology Images. Bioengineering (Basel) 9 (2022). https://doi.org:10.3390/bioengineering9090423 Ardon, O. et al. Understanding the financial aspects of digital pathology: A dynamic customizable return on investment calculator for informed decision-making. J Pathol Inform 15 , 100376 (2024). https://doi.org:10.1016/j.jpi.2024.100376 Bankhead, P. et al. QuPath: Open source software for digital pathology image analysis. Sci Rep 7 , 16878 (2017). https://doi.org:10.1038/s41598-017-17204-5 Echle, A. et al. Deep learning in cancer pathology: a new generation of clinical biomarkers. Br J Cancer 124 , 686-696 (2021). https://doi.org:10.1038/s41416-020-01122-x Linehan, W. M. & Ricketts, C. J. The Cancer Genome Atlas of renal cell carcinoma: findings and clinical implications. Nat Rev Urol 16 , 539-552 (2019). https://doi.org:10.1038/s41585-019-0211-5 Bossuyt, P. M. et al. STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. BMJ 351 , h5527 (2015). https://doi.org:10.1136/bmj.h5527 Collins, G. S. et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 385 , e078378 (2024). https://doi.org:10.1136/bmj-2023-078378 Wang, Z. et al. Label Cleaning Multiple Instance Learning: Refining Coarse Annotations on Single Whole-Slide Images. IEEE Trans Med Imaging 41 , 3952-3968 (2022). https://doi.org:10.1109/TMI.2022.3202759 Lei, W. et al. One-Shot Weakly-Supervised Segmentation in 3D Medical Images. IEEE Trans Med Imaging 43 , 175-189 (2024). https://doi.org:10.1109/TMI.2023.3294975 Homeyer, A. et al. Artificial Intelligence in Pathology: From Prototype to Product. J Pathol Inform 12 , 13 (2021). https://doi.org:10.4103/jpi.jpi_84_20 Graham, S. et al. Hover-Net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Med Image Anal 58 , 101563 (2019). https://doi.org:10.1016/j.media.2019.101563 Veta, M., Pluim, J. P., van Diest, P. J. & Viergever, M. A. Breast cancer histopathology image analysis: a review. IEEE Trans Biomed Eng 61 , 1400-1411 (2014). https://doi.org:10.1109/TBME.2014.2303852 Srinidhi, C. L., Ciga, O. & Martel, A. L. Deep neural network models for computational histopathology: A survey. Med Image Anal 67 , 101813 (2021). https://doi.org:10.1016/j.media.2020.101813 Campanella, G. et al. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat Med 25 , 1301-1309 (2019). https://doi.org:10.1038/s41591-019-0508-1 Cheng, J. et al. Computational analysis of pathological images enables a better diagnosis of TFE3 Xp11.2 translocation renal cell carcinoma. Nat Commun 11 , 1778 (2020). https://doi.org:10.1038/s41467-020-15671-5 Tabibu, S., Vinod, P. K. & Jawahar, C. V. Pan-Renal Cell Carcinoma classification and survival prediction from histopathology images using deep learning. Sci Rep 9 , 10509 (2019). https://doi.org:10.1038/s41598-019-46718-3 Niazi, M. K. K., Parwani, A. V. & Gurcan, M. N. Digital pathology and artificial intelligence. Lancet Oncol 20 , e253-e261 (2019). https://doi.org:10.1016/S1470-2045(19)30154-8 Shmatko, A., Ghaffari Laleh, N., Gerstung, M. & Kather, J. N. Artificial intelligence in histopathology: enhancing cancer research and clinical oncology. Nat Cancer 3 , 1026-1038 (2022). https://doi.org:10.1038/s43018-022-00436-4 Additional Declarations No competing interests reported. Supplementary Files supplementary20250716.docx 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-7282796","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":514637215,"identity":"18205530-e9fa-4c9a-924b-765617d7c274","order_by":0,"name":"Chengwei Chen","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Chengwei","middleName":"","lastName":"Chen","suffix":""},{"id":514637216,"identity":"b9002426-ec2e-4f4c-83e9-81c4c148bb25","order_by":1,"name":"Bing Xia","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Bing","middleName":"","lastName":"Xia","suffix":""},{"id":514637217,"identity":"2a45c8c0-6d66-4b5c-891f-0842c8bf1456","order_by":2,"name":"Qinqin Kang","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Qinqin","middleName":"","lastName":"Kang","suffix":""},{"id":514637218,"identity":"9f76b2c1-1a18-422f-8e5c-ab05205c135f","order_by":3,"name":"Na Li","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Na","middleName":"","lastName":"Li","suffix":""},{"id":514637219,"identity":"b492a9ef-ff1a-4871-8b6b-d29d0878e67b","order_by":4,"name":"Qianru Zhang","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Qianru","middleName":"","lastName":"Zhang","suffix":""},{"id":514637220,"identity":"352bf723-3adc-454b-9fed-2e859d7b445a","order_by":5,"name":"Yixuan Shen","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yixuan","middleName":"","lastName":"Shen","suffix":""},{"id":514637221,"identity":"d758337f-2adb-4dfb-aa93-0d1a4a18f0a0","order_by":6,"name":"Danqun Zheng","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Danqun","middleName":"","lastName":"Zheng","suffix":""},{"id":514637222,"identity":"ab6fcbf3-314d-4938-925a-2f1f62898128","order_by":7,"name":"Shao jia Mo","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Shao","middleName":"jia","lastName":"Mo","suffix":""},{"id":514637223,"identity":"cf7cde00-275f-48a8-ab02-12dbcbb30648","order_by":8,"name":"Yilun Zheng","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yilun","middleName":"","lastName":"Zheng","suffix":""},{"id":514637224,"identity":"54fe4c09-4a0e-45d7-8a32-c875fdb4bd4b","order_by":9,"name":"Jieyu Yu","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jieyu","middleName":"","lastName":"Yu","suffix":""},{"id":514637225,"identity":"0456ca4e-d882-47a5-ab3a-56c3bfbe4f3a","order_by":10,"name":"Jing Li","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Li","suffix":""},{"id":514637226,"identity":"9a24e36e-5190-4945-a853-fbdd6500d736","order_by":11,"name":"Xu Fang","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xu","middleName":"","lastName":"Fang","suffix":""},{"id":514637227,"identity":"d02e1a47-c7e4-4666-b6d3-cd156f791593","order_by":12,"name":"Fang Liu","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Fang","middleName":"","lastName":"Liu","suffix":""},{"id":514637228,"identity":"fb6500d8-dad5-425f-aaed-a164aa7f84bd","order_by":13,"name":"Ming Yang","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ming","middleName":"","lastName":"Yang","suffix":""},{"id":514637229,"identity":"a82b9fdf-a003-430e-acaf-d1695beff91c","order_by":14,"name":"Tiegong Wang","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Tiegong","middleName":"","lastName":"Wang","suffix":""},{"id":514637230,"identity":"9b92687b-1837-4fa8-b996-06556ffe25f6","order_by":15,"name":"Li Wang","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Wang","suffix":""},{"id":514637231,"identity":"eadd4bd7-2218-4202-9399-e577696acdb9","order_by":16,"name":"Jianping Lu","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jianping","middleName":"","lastName":"Lu","suffix":""},{"id":514637232,"identity":"18566f36-7249-48d0-bc3e-744974c01f31","order_by":17,"name":"Shuqian Zhang","email":"","orcid":"","institution":"Hebei General Hospital","correspondingAuthor":false,"prefix":"","firstName":"Shuqian","middleName":"","lastName":"Zhang","suffix":""},{"id":514637233,"identity":"17f479ba-48d3-4f2a-b702-fbaf1ea6a7a6","order_by":18,"name":"Hui Jiang","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hui","middleName":"","lastName":"Jiang","suffix":""},{"id":514637234,"identity":"edf171c3-70aa-43fc-a3af-b48f3593c298","order_by":19,"name":"Chengwei Shao","email":"","orcid":"","institution":"Changhai Hospital","correspondingAuthor":false,"prefix":"","firstName":"Chengwei","middleName":"","lastName":"Shao","suffix":""},{"id":514637235,"identity":"68f62a06-b7a4-49ad-bfde-279c8c586659","order_by":20,"name":"Yun Bian","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+UlEQVRIiWNgGAWjYBACA2YILcfGzHzwQQWIyczcQJQWY352tmSDM2AtjAS0QOnEmf08ZhJgLQwEtJizMz97zNt2h3HDYQazigMVh6P524FaflRsw6nFspnN3Ji37RmzwWGGtBsHzqTlzjjM2MDYc+Y2bocBDZfmbTvMBmQcu/2xzSa3AaiFmbENnxb2byAtPAaHGdsKDrZJ5M4nrIUHbIuEZDMzG8NBoC0bCGmxbOYpk5xz7rABPzMbswTILxuBWg7i84s5//FtEm/KDte38Z//+AEYYrnzzh8++OBHBW4tIMDEgy5yAK96IGD8QUjFKBgFo2AUjGwAABsuWgqzJNLnAAAAAElFTkSuQmCC","orcid":"","institution":"Changhai Hospital","correspondingAuthor":true,"prefix":"","firstName":"Yun","middleName":"","lastName":"Bian","suffix":""}],"badges":[],"createdAt":"2025-08-03 10:38:22","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7282796/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7282796/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":91504896,"identity":"ec909a63-2047-4824-afac-52cd294eb68e","added_by":"auto","created_at":"2025-09-17 08:13:42","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":365489,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverall Study Workflow\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA flowchart illustrating the systematic methodology of the study. The workflow begins with the definition and acquisition of the training, validation, and test cohorts from multiple institutions. It then details the automated pipeline for whole-slide image (WSI) preprocessing, the extraction of multi-scale features, and the subsequent processes of feature selection and dimensionality reduction. Finally, it outlines the construction of the predictive model, its training, and its rigorous evaluation on the independent test cohort to assess diagnostic performance and generalizability.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7282796/v1/7b09f82bf3a9c993a50f26b3.png"},{"id":91504878,"identity":"786d8427-27c8-4882-bb73-cac23e7303e2","added_by":"auto","created_at":"2025-09-17 08:13:41","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":150287,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSchematic of the RCCNET (Renal Cell Carcinoma Neural Enhancement Technology) AI Framework\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA schematic illustration of the dual-resolution RCCNET framework architecture. The framework takes a whole-slide image (WSI) as input, which undergoes tessellation into numerous tiles. Stream 1 (Classification) processes these tiles through a CTransPath feature extractor and a Swin Transformer backbone to model spatial dependencies, followed by an attention-based pooling mechanism to generate a final subtype classification. Stream 2 (Interpretability) runs in parallel, using a HoVer-Net based model for instance-level nuclear segmentation and classification (tumor, inflammatory, dead cells), from which quantitative morphometric features are extracted. The attention mechanism from Stream 1 also generates attention heatmaps, providing a direct method for visualizing the model's decision-making process.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7282796/v1/a1df84b7cff0c4c591cd66ee.png"},{"id":91504894,"identity":"c120fb9f-0574-4317-a69a-4a0ab52a4f01","added_by":"auto","created_at":"2025-09-17 08:13:42","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":316083,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDiagnostic Performance of the RCCNET Framework\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A, B)\u003c/strong\u003e Receiver operating characteristic (ROC) curves for the four-class classification task in the training (A, n=233) and validation (B, n=107) cohorts. Each curve represents the performance of RCCNET for one RCC subtype, with area under the curve (AUC) values indicating discriminative ability. Training cohort AUCs: Clear Cell RCC 0.991, Papillary RCC 0.988, Chromophobe RCC 0.995, TFE3 RCC 0.989. Validation cohort AUCs: Clear Cell RCC 0.972, Papillary RCC 0.951, Chromophobe RCC 0.979, TFE3 RCC 0.976.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(C, D)\u003c/strong\u003e Confusion matrices detailing the classification results for the training (C, n=233) and validation (D, n=107) sets, with diagonal elements representing correct predictions. Training cohort sample distribution: 118 Clear Cell RCC, 56 Papillary RCC, 28 Chromophobe RCC, 31 TFE3 RCC. Validation cohort sample distribution: 56 Clear Cell RCC, 26 Papillary RCC, 12 Chromophobe RCC, 13 TFE3 RCC. The validation cohort demonstrates 16 total misclassifications (15.0% error rate) with 91 correct predictions (85.0% overall accuracy).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(E-H)\u003c/strong\u003e Bar charts quantifying key performance metrics—sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1-score—for each RCC subtype. Darker bars represent the training set performance, and lighter bars represent the validation set performance. Error bars indicate 95% confidence intervals. Critical limitation: TFE3 RCC demonstrates a PPV of only 66.7% in the validation cohort, indicating that one-third of positive predictions are false positives, necessitating molecular confirmation for clinical implementation. All other subtypes maintain PPV \u0026gt;80% in validation.\u003c/p\u003e\n\u003cp\u003eAbbreviations: AUC, area under the curve; ccRCC, clear cell renal cell carcinoma; chRCC, chromophobe renal cell carcinoma; NPV, negative predictive value; PPV, positive predictive value; pRCC, papillary renal cell carcinoma; TFE3 RCC, TFE3-rearranged renal cell carcinoma.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-7282796/v1/b72261ba74a6d825affd38f4.png"},{"id":91504895,"identity":"312e9636-8ce3-42fc-8581-83710d1da3f2","added_by":"auto","created_at":"2025-09-17 08:13:42","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1157210,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAttention-based interpretability of RCCNET for multi-class RCC subtype classification\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAttention heatmaps demonstrate the model's focus on diagnostically relevant histopathological features for accurate subtype classification. Heat intensity (color scale: blue = low attention, red = high attention) indicates the relative importance of tissue regions for classification decisions. Numbered boxes on heatmaps correspond to regions of interest (ROIs) shown in adjacent H\u0026amp;E images at higher magnification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(A) \u003c/strong\u003eTFE3-rearranged RCC attention heatmap with corresponding H\u0026amp;E sections (A1-A3) from high-attention regions. The model correctly focuses on areas containing characteristic morphological features including voluminous eosinophilic cytoplasm, prominent nucleoli, and nested growth patterns typical of TFE3 RCC.\u003cstrong\u003e(B) \u003c/strong\u003eClear cell RCC attention heatmap with corresponding H\u0026amp;E sections (B1-B3) from high-attention regions. Attention is concentrated on regions displaying pathognomonic clear cell features including abundant clear cytoplasm, distinct cell borders, and delicate capillary networks characteristic of ccRCC. \u003cstrong\u003e(C) \u003c/strong\u003ePapillary RCC attention heatmap with corresponding H\u0026amp;E sections (C1-C3) from high-attention regions. The model appropriately highlights areas with complex papillary architecture, fibrovascular cores, and the characteristic cellular arrangement of papillary RCC. \u003cstrong\u003e(D)\u003c/strong\u003e Chromophobe RCC attention heatmap with corresponding H\u0026amp;E sections (D1-D3) from high-attention regions. Attention accurately identifies regions containing cells with distinctive plant cell-like borders, perinuclear halos, and the uniform cellular morphology characteristic of chromophobe RCC.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-7282796/v1/e8a1fd41e66a2d6f9d3ba536.png"},{"id":92668688,"identity":"c9e4b012-743d-413d-97c2-bdc94f12dd71","added_by":"auto","created_at":"2025-10-02 17:31:36","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3466021,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7282796/v1/4a26cc51-6c3b-4a5b-833e-8ed8bff019bc.pdf"},{"id":91504888,"identity":"7b0ece30-bafd-4930-8cbe-13227208bcdb","added_by":"auto","created_at":"2025-09-17 08:13:41","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":2765720,"visible":true,"origin":"","legend":"","description":"","filename":"supplementary20250716.docx","url":"https://assets-eu.researchsquare.com/files/rs-7282796/v1/0c6056b1a19256b89c1cba02.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Weakly Supervised Learning for Multi-class RCC Classification: Multicenter Validation with Biological Interpretability","fulltext":[{"header":"Introduction","content":"\u003cp\u003eRenal cell carcinoma represents a heterogeneous group of malignancies with distinct morphological, molecular, and clinical characteristics that fundamentally shape therapeutic decisions and patient outcomes \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. The complexity of accurate subtype classification has intensified with the 2022 World Health Organization classification, which recognizes over 20 distinct RCC entities, each requiring precise identification for optimal patient management \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Among these, clear cell RCC, papillary RCC, chromophobe RCC, and TFE3-rearranged RCC constitute the most clinically significant subtypes, yet their accurate differentiation remains one of the most challenging aspects of contemporary urological pathology \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e,\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eThe diagnostic complexity reaches its zenith with TFE3-rearranged RCC, an entity that exemplifies the intersection of morphological heterogeneity and molecular precision medicine\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Characterized by translocations involving the TFE3 gene, this subtype presents a morphological spectrum that can convincingly mimic clear cell, papillary, or even unclassified RCC patterns. This diagnostic ambiguity necessitates molecular confirmation through fluorescence in situ hybridization or next-generation sequencing, creating a cascade of increased costs, extended diagnostic timelines, and resource allocation challenges that particularly burden institutions with limited molecular pathology capabilities \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eTraditional diagnostic paradigms, while foundational to pathological practice, face mounting pressures from several converging factors \u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Inter-observer variability among pathologists, even those with subspecialty expertise, introduces diagnostic uncertainty that can directly impact patient care\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Resource constraints, particularly in molecular testing capabilities, create disparities in diagnostic access across different healthcare systems \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Perhaps most critically, the increasing complexity of RCC classification demands a level of subspecialty expertise that may not be readily available in all institutions, creating potential gaps in diagnostic accuracy and consistency\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eThe emergence of artificial intelligence in pathology represents a valuable technological advancement toward objective, reproducible, and scalable diagnostic solutions\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Recent developments in deep learning, particularly in computer vision and weakly supervised learning approaches, have demonstrated promising results in medical image analysis\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. However, the translation of these technological capabilities to the specific challenges of RCC subtype classification has been limited by several factors: most existing approaches focus on binary classification tasks that oversimplify the clinical reality, lack comprehensive validation across diverse institutional settings, and fail to address the critical need for interpretability that is essential for clinical acceptance and regulatory approval \u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eTo simultaneously address the dual challenges of diagnostic accuracy and clinical interpretability, we developed RCCNET (Renal Cell Carcinoma Neural Enhancement Technology), a comprehensive computational framework that represents a meaningful advance over conventional approaches \u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Our innovation lies not merely in achieving high classification accuracy, but in creating a system that provides morphometric correlates for its decisions through quantitative cellular analysis. This dual-stream architecture addresses both the immediate clinical need for accurate subtype classification and the longer-term requirement for transparent, interpretable computational systems that can gain the trust and acceptance of the pathology community.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStudy Design and Reporting Guidelines Adherence\u003c/h2\u003e\u003cp\u003eThis investigation was conceived and executed as a rigorous multicenter diagnostic accuracy study, adhering to the Standards for Reporting of Diagnostic Accuracy Studies (STARD) 2015 guidelines\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e and the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis\u0026thinsp;+\u0026thinsp;Artificial Intelligence (TRIPOD\u0026thinsp;+\u0026thinsp;AI) statement \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. The study protocol received approval from the Ethics Committees of both participating institutions and was conducted in strict accordance with the Declaration of Helsinki principles \u003cb\u003e(Supplementary Method 1 and Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/b\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eStudy Setting and Participants\u003c/h3\u003e\n\u003cp\u003eThis retrospective, multicenter study enrolled consecutive patients from two tertiary Grade A hospitals in China to ensure real-world generalizability. The training cohort was derived from XX Hospital (a major referral center with advanced digital pathology) for patients treated between June 2013 and December 2021. The external validation cohort was sourced from XXX Hospital (a center with standard pathology practices) for patients treated between January 2021 and June 2023.\u003c/p\u003e\u003cp\u003eInclusion criteria were: a history of radical or partial nephrectomy for a renal mass; a final histopathological diagnosis of clear cell, papillary, chromophobe, or TFE3-rearranged RCC according to the 2022 WHO classification; and available FFPE tissue with high-quality H\u0026amp;E slides\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Cases were excluded for mixed or uncertain histology, extensive necrosis (\u0026gt;\u0026thinsp;50%), significant artifacts, inadequate tumor tissue (\u0026lt;\u0026thinsp;75% tumor content or \u0026lt;\u0026thinsp;1 cm\u0026sup2; area), or lack of molecular confirmation for TFE3-rearranged RCC\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003e\u003cb\u003eReference Standard and Molecular Confirmation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe reference standard for RCC subtype classification was established through consensus review by two expert genitourinary pathologists, each with more than 10 years of subspecialty experience, using the 2022 WHO classification criteria \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. This approach ensured diagnostic consistency while maintaining the highest standards of pathological expertise. All TFE3-rearranged RCC cases underwent molecular confirmation through fluorescence in situ hybridization using TFE3 break-apart probes or RNA sequencing when available, ensuring the diagnostic accuracy essential for training and validating our computational system \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\n\u003ch3\u003eRCCNET Framework Architecture\u003c/h3\u003e\n\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003ePrimary Classification Module (Stream 1)\u003c/h2\u003e\u003cp\u003eThe architecture of the RCCNET framework is detailed in \u003cb\u003eSupplementary Method 3\u003c/b\u003e and \u003cb\u003eFig.\u0026nbsp;1\u003c/b\u003e, with the full digital pathology workflow described in \u003cb\u003eSupplementary Method 2\u003c/b\u003e. The primary classification module (Stream 1) employs a sophisticated weakly supervised multiple instance learning approach, designed to process whole-slide images without requiring the pixel-level annotations that have historically limited computational development \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. This choice circumvents the practical bottleneck of creating detailed annotations, especially for rare entities like TFE3 RCC \u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. The pipeline begins by dividing each WSI (approximately 100,000 \u0026times; 50,000 pixels) into non-overlapping tiles measuring 256 \u0026micro;m \u0026times; 256 \u0026micro;m, corresponding to 224 \u0026times; 224-pixel patches at 0.5 \u0026micro;m/pixel. From these, a CTransPath encoder\u0026mdash;chosen for its superior performance on histopathological images over general-purpose models\u0026mdash;extracts 768-dimensional feature vectors for each tile. A Swin Transformer backbone then models the spatial dependencies and architectural patterns across the tissue \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e, capturing both local and global features to address the multi-scale nature of pathological diagnosis\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. Finally, an attention-based aggregation mechanism weights each tile's contribution to the classification decision and generates interpretable attention heatmaps to highlight diagnostically relevant regions \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003e\u003cb\u003eFigure 1. Overall Study Workflow\u003c/b\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eMorphometric Analysis Module (Stream 2)\u003c/h3\u003e\n\u003cp\u003eTo address interpretability challenges in computational pathology, the Morphometric Analysis Module (Stream 2) was designed to provide quantitative, biologically meaningful correlates for the model's classification decisions (\u003cb\u003eSupplementary Method 4\u003c/b\u003e). Nuclear segmentation was performed using the HoVer-Net model to classify individual nuclei into three categories (neoplastic, inflammatory, and dead cells) \u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, and its performance was rigorously validated on 200 annotated tissue regions to ensure the reliability of downstream analysis \u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. For each nucleus, we computed 11 morphometric descriptors across categories of size (e.g., area, volume), shape (e.g., roundness, eccentricity) \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, and orientation (e.g., major/minor axes) \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Finally, five statistical descriptors (including mean, standard deviation, and entropy) were calculated for these features to transform qualitative pathological observations into quantitative, reproducible measurements for systematic comparison.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eTraining and Validation Strategy\u003c/h2\u003e\u003cp\u003eModel training was conducted on the XX Hospital cohort (n\u0026thinsp;=\u0026thinsp;233) using 5-fold cross-validation for hyperparameter optimization and to prevent overfitting. To address class imbalance, particularly for TFE3 RCC (13.3% of training data), we implemented a multi-faceted strategy including focal loss (γ\u0026thinsp;=\u0026thinsp;2.0, α\u0026thinsp;=\u0026thinsp;0.25), class-weighted sampling, and enhanced data augmentation. The model's generalizability was then assessed on the completely independent XXX Hospital cohort (n\u0026thinsp;=\u0026thinsp;107) without any retraining or fine-tuning, providing an unbiased assessment.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003ePerformance Metrics and Clinical Integration Framework\u003c/h3\u003e\n\u003cp\u003ePrimary diagnostic performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and predictive values for each subtype. Based on the model's confidence score distribution and error analysis, we developed a three-tier clinical integration framework to balance accuracy with practical implementation: high-confidence (\u0026gt;\u0026thinsp;0.8) cases for minimal review, moderate-confidence (0.4\u0026ndash;0.8) cases for guided pathologist review, and low-confidence (\u0026lt;\u0026thinsp;0.4) cases directed to traditional workflows with molecular testing as indicated.\u003c/p\u003e\n\u003ch3\u003eEconomic Analysis\u003c/h3\u003e\n\u003cp\u003eA comprehensive cost-effectiveness analysis, with detailed modeling assumptions provided in \u003cb\u003eSupplementary Method 6\u003c/b\u003e, was conducted to compare the RCCNET-assisted workflow with traditional diagnosis, incorporating both direct and indirect costs for a realistic assessment of economic impact \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eContinuous and categorical variables were described and compared using appropriate statistical tests (e.g., Student's t-test, Mann-Whitney U test, chi-square test). Model performance was evaluated using ROC analysis, with DeLong's test for AUC comparisons. Morphometric correlations were assessed using Spearman correlation with Bonferroni correction (α\u0026thinsp;=\u0026thinsp;0.0003), and misclassification analysis included effect size calculation (Cohen's d). Economic analysis employed decision tree modeling. A p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant. All analyses were performed using Python 3.11.0 and R 4.3.0, with technical details provided in \u003cb\u003eSupplementary Method 5\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eCohort Characteristics and Baseline Demographics\u003c/h2\u003e\u003cp\u003eThe study included 340 patients from two institutions, divided into a training cohort (n\u0026thinsp;=\u0026thinsp;233) and a validation cohort (n\u0026thinsp;=\u0026thinsp;107) \u003cb\u003e(\u003c/b\u003eFig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e2\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e. Both cohorts showed a similar, representative distribution of RCC subtypes: the training set included 118 clear cell (ccRCC), 56 papillary (pRCC), 28 chromophobe (chRCC), and 31 TFE3-rearranged (TFE3 RCC) cases, while the validation set comprised 56, 26, 12, and 13 cases of each subtype, respectively. Baseline demographics and tumor characteristics were well-balanced between the cohorts, with no significant differences in age, sex, or tumor stage distribution. Tumor characteristics showed similar distributions, with the majority of cases presenting as Stage I disease (training: 68.2%, validation: 71.0%), providing a representative sample of contemporary RCC surgical practice \u003cb\u003e(\u003c/b\u003eTable\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eBaseline Demographic and Clinicopathological Characteristics Across RCC Subtypes\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"12\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c11\" colnum=\"11\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c12\" colnum=\"12\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e\u003cp\u003eCharacteristics\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"5\" nameend=\"c6\" namest=\"c2\"\u003e\u003cp\u003eTrain Group\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"5\" nameend=\"c12\" namest=\"c8\"\u003e\u003cp\u003eValidation Group\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003epRCC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTFE3-RCC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eccRCC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003echRCC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eP value\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003epRCC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eTFE3-RCC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c10\"\u003e\u003cp\u003eccRCC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c11\"\u003e\u003cp\u003echRCC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c12\"\u003e\u003cp\u003eP value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNumber, n\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e118\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e28\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e26\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAge, years (Mean\u0026thinsp;\u0026plusmn;\u0026thinsp;SD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e53.8\u0026thinsp;\u0026plusmn;\u0026thinsp;14.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e52.1\u0026thinsp;\u0026plusmn;\u0026thinsp;13.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e54.9\u0026thinsp;\u0026plusmn;\u0026thinsp;13.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e55.2\u0026thinsp;\u0026plusmn;\u0026thinsp;14.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.342\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e54.2\u0026thinsp;\u0026plusmn;\u0026thinsp;12.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e53.8\u0026thinsp;\u0026plusmn;\u0026thinsp;11.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e56.1\u0026thinsp;\u0026plusmn;\u0026thinsp;12.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e57.3\u0026thinsp;\u0026plusmn;\u0026thinsp;13.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.445\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSex, n (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.721\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.689\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e35 (62.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e19 (61.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e78 (66.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e19 (67.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e16 (61.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e8 (61.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e36 (64.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e7 (58.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFemale\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e21 (37.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e12 (38.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e40 (33.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e9 (32.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e10 (38.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e5 (38.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e20 (35.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e5 (41.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eBMI (kg/m\u0026sup2;\u0026plusmn; SD)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e24.35\u0026thinsp;\u0026plusmn;\u0026thinsp;3.19\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e24.60\u0026thinsp;\u0026plusmn;\u0026thinsp;3.89\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e24.56\u0026thinsp;\u0026plusmn;\u0026thinsp;3.44\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e24.49\u0026thinsp;\u0026plusmn;\u0026thinsp;3.84\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.732\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e23.40\u0026thinsp;\u0026plusmn;\u0026thinsp;3.20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e23.64\u0026thinsp;\u0026plusmn;\u0026thinsp;3.35\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e24.54\u0026thinsp;\u0026plusmn;\u0026thinsp;3.44\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e24.01\u0026thinsp;\u0026plusmn;\u0026thinsp;4.65\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.732\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTumor Stage, n (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.654\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.598\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eStage I\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e38 (67.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e22 (71.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e80 (67.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e19 (67.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e18 (69.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e9 (69.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e41 (73.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e8 (66.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eStage II\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e10 (17.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e5 (16.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e21 (17.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5 (17.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e4 (15.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e2 (15.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e9 (16.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e2 (16.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eStage III\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7 (12.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e3 (9.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e15 (12.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3 (10.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e3 (11.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e2 (15.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e5 (8.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e1 (8.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eStage IV\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1 (1.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1 (3.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2 (1.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1 (3.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e1 (3.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0 (0.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e1 (1.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e1 (8.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTumor Grade, n (%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.445\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.512\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrade 1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e11 (19.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e6 (19.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e23 (19.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5 (17.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e4 (15.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e2 (15.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e10 (17.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e2 (16.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrade 2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e27 (48.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e15 (48.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e56 (47.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e14 (50.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e14 (53.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e7 (53.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e29 (51.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e6 (50.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrade 3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e14 (25.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e8 (25.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e30 (25.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e6 (21.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e6 (23.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e3 (23.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e13 (23.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e3 (25.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eGrade 4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4 (7.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2 (6.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e9 (7.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e3 (10.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e2 (7.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e1 (7.7)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e4 (7.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e1 (8.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c12\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTumor Size, cm [Median (IQR)]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4.0 (2.8\u0026ndash;5.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e4.3 (3.2\u0026ndash;6.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e4.2 (3.1-6.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e4.5 (3.3\u0026ndash;6.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.523\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e3.8 (2.6\u0026ndash;5.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e4.1 (2.9\u0026ndash;5.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c10\"\u003e\u003cp\u003e4.0 (2.9\u0026ndash;5.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c11\"\u003e\u003cp\u003e4.3 (3.1\u0026ndash;6.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c12\"\u003e\u003cp\u003e0.467\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003ePrimary Classification Performance\u003c/h2\u003e\u003cp\u003eRCCNET demonstrated strong diagnostic performance, achieving a macro-averaged AUC of 0.989 (95% CI: 0.985\u0026ndash;0.993) in the training cohort and showing robust generalization with an AUC of 0.966 (95% CI: 0.951\u0026ndash;0.981) in the external validation cohort.\u003c/p\u003e\u003cp\u003eIn validation, subtype-specific performance remained high \u003cb\u003e(\u003c/b\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e \u003cb\u003eand\u003c/b\u003e Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e3\u003c/span\u003e\u003cb\u003e)\u003c/b\u003e. The model effectively identified clear cell RCC (AUC 0.972, sensitivity 96.4%), papillary RCC (AUC 0.951, sensitivity 84.6%), and chromophobe RCC (AUC 0.979, sensitivity 83.3%). For the diagnostically challenging TFE3 RCC, the model achieved an AUC of 0.976 and 92.3% sensitivity. However, this was tempered by a critical limitation: the positive predictive value was only 66.7% (12 of 18 predictions), indicating a one-third false-positive rate that necessitates molecular confirmation due to significant clinical implications. Model visualization for diagnosis is shown in Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e4\u003c/span\u003e \u003cb\u003eand Supplementary Figure S2\u003c/b\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eRCCNET Performance Metrics Across Training and Validation Cohorts\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRCC Subtype\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003en\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAUC (95% CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSensitivity (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSpecificity (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003ePPV (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eNPV (%)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eAccuracy (%)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTraining Set\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClear Cell RCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e118\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.991 (0.985\u0026ndash;0.997)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e96.6 (114/118)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e95.7 (110/115)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e95.0 (114/119)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e97.0 (110/114)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e96.1 (224/233)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePapillary RCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.988 (0.979\u0026ndash;0.997)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e94.6 (53/56)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e97.2 (172/177)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e91.4 (53/58)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e98.3 (172/175)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e96.6 (225/233)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eChromophobe RCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e28\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.995 (0.990-1.000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e92.9 (26/28)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e99.0 (203/205)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e92.9 (26/28)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e99.0 (203/205)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e98.3 (229/233)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTFE3 RCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.989 (0.981\u0026ndash;0.997)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e90.3 (28/31)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e98.0 (198/202)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e87.5 (28/32)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e98.5 (198/201)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e97.0 (226/233)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMacro-averaged\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e233\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.989 (0.985\u0026ndash;0.993)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e93.6 (218/233)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e97.5 (681/699)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e91.7 (218/238)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e98.2 (681/694)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e97.4 (227/233)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eValidation Set\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClear Cell RCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.972 (0.951\u0026ndash;0.993)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e96.4 (54/56)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e82.4 (42/51)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e85.7 (54/63)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e95.5 (42/44)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e89.7 (96/107)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePapillary RCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e26\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.951 (0.912\u0026ndash;0.990)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e84.6 (22/26)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e95.1 (77/81)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e84.6 (22/26)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e95.1 (77/81)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e92.5 (99/107)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eChromophobe RCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.979 (0.958-1.000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e83.3 (10/12)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e98.9 (94/95)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e90.9 (10/11)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e97.9 (94/96)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e97.2 (104/107)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTFE3 RCC\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.976 (0.943-1.000)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e92.3 (12/13)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e93.6 (88/94)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e66.7 (12/18)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e98.9 (88/89)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e93.5 (100/107)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eMacro-averaged\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e107\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.966 (0.951\u0026ndash;0.981)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e89.2 (91/107)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e92.5 (301/321)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e81.9 (91/111)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e96.9 (301/317)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e85.0 (91/107)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"8\" nameend=\"c8\" namest=\"c1\"\u003e\u003cp\u003e\u003cb\u003eNote\u003c/b\u003e: Numbers in parentheses for validation cohort represent actual counts (true positives/total positives for sensitivity; true negatives/total negatives for specificity; true positives/predicted positives for PPV; true negatives/predicted negatives for NPV). AUC\u0026thinsp;=\u0026thinsp;area under the receiver operating characteristic curve; CI\u0026thinsp;=\u0026thinsp;confidence interval; PPV\u0026thinsp;=\u0026thinsp;positive predictive value; NPV\u0026thinsp;=\u0026thinsp;negative predictive value; RCC\u0026thinsp;=\u0026thinsp;renal cell carcinoma.\u003c/p\u003e\u003cp\u003eCritical Clinical Note: The positive predictive value for TFE3 RCC in validation is 66.7%, indicating that one-third of cases predicted as TFE3 RCC are false positives. Molecular confirmation is essential for all positive TFE3 RCC predictions.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eA DeLong test confirmed the model's robustness, showing no significant differences in individual subtype AUCs between cohorts (p\u0026thinsp;\u0026gt;\u0026thinsp;0.05 for all). Although the macro-average AUC showed a statistical difference (p\u0026thinsp;=\u0026thinsp;0.045), the small effect size (95% CI: 0.001\u0026ndash;0.045) reinforces the model's strong performance in external validation \u003cb\u003e(Supplementary Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e)\u003c/b\u003e.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eNuclear Segmentation Performance and Morphometric Analysis Foundation\u003c/h2\u003e\u003cp\u003eThe reliability of our morphometric analysis depended on accurate nuclear segmentation, which achieved high F1-scores for neoplastic (0.90\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04), inflammatory (0.83\u0026thinsp;\u0026plusmn;\u0026thinsp;0.06), and dead cells (0.77\u0026thinsp;\u0026plusmn;\u0026thinsp;0.09), establishing a robust foundation comparable to published benchmarks (\u003cb\u003eSupplementary Table S2 and Supplementary Figure S3\u003c/b\u003e). Segmentation performance varied across RCC subtypes, reflecting their distinct morphologies. Performance was highest in clear cell RCC due to its clear cytoplasm and contrast, and the model was particularly effective at identifying the elongated neoplastic cells of papillary RCC. Similarly, the uniform cell size and distinct borders of chromophobe RCC contributed to excellent detection performance. In contrast, the model\u0026rsquo;s moderate performance on TFE3 RCC mirrored this subtype's known morphological heterogeneity, which is also a diagnostic challenge for pathologists.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003eCellular Morphometric Analysis and Distribution Patterns\u003c/h2\u003e\u003cp\u003eMorphometric analysis revealed key quantitative differences among RCC subtypes, validating established pathological knowledge with new, measurable insights. For instance, nuclear area measurements followed subtype-specific log-normal distributions, and size heterogeneity was, consistent with pathological descriptions, highest in clear cell RCC and lowest in chromophobe RCC. Features such as cellular roundness clearly distinguished neoplastic from inflammatory populations, while eccentricity was highest in papillary RCC, reflecting its characteristic elongated morphology. Cellular compartment analysis provided further biological insights (\u003cb\u003eSupplementary Table S3\u003c/b\u003e). Neoplastic cells, comprising 45\u0026ndash;75% of total nuclei, showed the greatest diversity in clear cell RCC and the most uniformity in chromophobe RCC. Notably, the proportion of inflammatory cells (15\u0026ndash;35% of nuclei) was significantly reduced in TFE3-rearranged RCC, providing quantitative support for its described \"immune-cold\" phenotype, while dead cells accounted for a highly variable 5\u0026ndash;15% of nuclei.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003eInvestigating Biological Correlates Through Computational-Morphometry Correlation Analysis\u003c/h2\u003e\u003cp\u003eTo ensure RCCNET's accuracy was grounded in true histopathological features, we performed a post-hoc Spearman correlation analysis linking model predictions to quantitative cellular morphometrics. This analysis confirmed the model learned clinically relevant patterns \u003cb\u003eSupplementary Figure S4\u003c/b\u003e and \u003cb\u003eSupplementary Table S4\u003c/b\u003e). For papillary RCC, predictions strongly correlated with features of its elongated cellular morphology (tumor cell roundness kurtosis: r = -0.509, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and complex architecture (tumor cell eccentricity entropy: r\u0026thinsp;=\u0026thinsp;0.445, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). The model identified the \"immune-cold\" phenotype of TFE3 RCC, showing negative correlations with inflammatory cell count (r = -0.270, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and volume variance (r = -0.245, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Furthermore, the model captured the known heterogeneity of clear cell RCC (positive correlation with eccentricity entropy: r\u0026thinsp;=\u0026thinsp;0.515, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and the uniformity of chromophobe RCC (negative correlation with perimeter kurtosis: r = -0.472, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Collectively, these results provide compelling evidence that RCCNET's classifications are based on measurable, biologically meaningful features aligned with established pathology, thus enhancing its interpretability \u003cb\u003e(Supplementary Figure S5\u003c/b\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003eMisclassification Analysis and Clinical Integration Strategy\u003c/h2\u003e\u003cp\u003eAnalysis of the 16 misclassified validation cases (15.0% error rate) revealed that errors typically occurred in morphologically ambiguous cases, with clear cell RCC being misclassified as TFE3 RCC most frequently (6 cases) (\u003cb\u003eSupplementary Figure S6 and Table S4\u003c/b\u003e). These false positive TFE3 predictions were predominantly known mimics, such as ccRCC with eosinophilic features, underscoring the need for molecular confirmation. Critically, RCCNET demonstrated a key safety feature, as confidence scores for misclassified cases were significantly lower than for correct predictions (mean 0.56 vs. 0.89, respectively; p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). This self-awareness underpins our proposed three-tier clinical integration framework (\u003cb\u003eSupplementary Table S6\u003c/b\u003e): 1) high-confidence cases (\u0026gt;\u0026thinsp;0.8 score; 75% of the cohort; 95.0% accuracy) are suitable for minimal review; 2) moderate-confidence cases (0.4\u0026ndash;0.8 score; 21% of the cohort; 81.8% accuracy) are flagged for guided pathologist review ; and 3) low-confidence cases (\u0026lt;\u0026thinsp;0.4 score; 5% of the cohort; 60.0% accuracy) are directed to traditional workflows.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec19\" class=\"Section2\"\u003e\u003ch2\u003eEconomic Impact Analysis\u003c/h2\u003e\u003cp\u003eOur comprehensive economic analysis revealed that the RCCNET-assisted workflow could yield substantial benefits by reducing per-case costs by 83.2% (from CNY 2,847 to CNY 479) and diagnostic time by 45.2% (from 58.5 to 32.0 minutes), significantly impacting laboratory throughput \u003cb\u003e(Supplementary Table S7)\u003c/b\u003e. These efficiencies were primarily driven by projected decreases in the use of routine immunohistochemistry (70.6% reduction) and unnecessary molecular tests (44.4% reduction) for most RCC subtypes. However, these potential savings must be balanced against the clinical necessity of molecular confirmation for all cases predicted as TFE3 RCC, given the model's 66.7% positive predictive value for this specific subtype.\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis study presents RCCNET as a valuable advance toward a comprehensive, interpretable diagnostic system that addresses the complex realities of contemporary renal cell carcinoma (RCC) classification \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Our results demonstrate that accurate multi-class RCC subtype classification is achievable using a weakly supervised approach that circumvents the need for pixel-level annotations \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. The framework shows particular strength for the diagnostically challenging TFE3-rearranged subtype, though we also highlight that careful attention to its false positive rate is essential for safe clinical implementation \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eThe strong performance of RCCNET, particularly for TFE3 RCC (AUC 0.976, sensitivity 92.3%), represents a meaningful advance addressing a critical diagnostic gap \u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Unlike previous studies focusing on binary classification \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, our multi-class framework directly addresses real-world diagnostic workflows where pathologists must simultaneously distinguish among multiple RCC entities \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. While other groups have pioneered computational analysis for TFE3 RCC, our annotation-free method overcomes key scalability barriers, enabling broader clinical deployment \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. However, this must be balanced against the 66.7% positive predictive value for TFE3 RCC \u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e, which requires careful clinical interpretation. This finding indicates that one-third of positive TFE3 predictions are false positives\u0026mdash;predominantly involving clear cell RCC with eosinophilic features or papillary RCC with solid patterns. This limitation necessitates molecular confirmation for positive TFE3 RCC predictions, particularly those with moderate confidence scores, and emphasizes the importance of integrating computational tools with traditional workflows rather than replacing them entirely \u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. Despite this, the model's robust generalization across institutions (macro-averaged AUC 0.966) validates its clinical applicability in diverse settings.\u003c/p\u003e\u003cp\u003eA key innovation of this study is the integration of cellular morphometric analysis, which provides biological correlates for the model's predictions and addresses the fundamental \"black box\" challenge of interpretability in pathology \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. The notable correlations between prediction probabilities and quantitative morphological features provide compelling evidence that our model learns clinically relevant patterns \u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. For example, the strong correlation between papillary RCC predictions and tumor cell eccentricity (r\u0026thinsp;=\u0026thinsp;0.445) directly reflects the elongated cellular morphology that pathologists recognize as characteristic of this subtype \u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Similarly, the reduced inflammatory infiltration found in TFE3 RCC cases (r = -0.270) provides quantitative support for its \"immune-cold\" phenotype, contributing new insights into RCC biology. While this post-hoc correlation does not establish causation, it provides strong support for the biological plausibility of the model's decisions.\u003c/p\u003e\u003cp\u003eFor clinical translation, we propose a confidence-based integration framework that provides a roadmap for safe, graduated implementation while maintaining appropriate human oversight \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. This strategy is viable because of the model's intrinsic ability to identify its own limitations; misclassified cases had significantly lower confidence scores than correct ones (0.56 vs 0.89, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), enabling workflows that automatically flag uncertain cases for expert review \u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. Our three-tier system stratifies cases for either minimal review (high confidence: 75% of cases, 95.0% accuracy), guided pathologist review (moderate confidence: 21% of cases, 81.8% accuracy), or traditional workup (low confidence: 5% of cases, 60.0% accuracy) \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. This approach, combined with substantial projected cost (83.2%) and time (45.2%) savings, presents a practical pathway for implementation, although the final economic advantages must be balanced against the necessary costs of TFE3 molecular confirmation \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eThis work has direct implications for making precision medicine more accessible, as RCCNET can help guide molecular testing strategies, particularly in resource-limited institutions. Future work should focus on exploring dynamic models that integrate molecular data and performing mechanistic studies to move beyond correlation toward causation. Several limitations also warrant acknowledgment. Our study was limited to four common RCC subtypes and the Chinese population, and its 2D-based morphometrics cannot fully capture three-dimensional nuclear architecture \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Most importantly, the 33.3% false positive rate and small validation sample size (n\u0026thinsp;=\u0026thinsp;13) for TFE3 RCC are critical limitations that mandate caution. This reinforces that for rare and challenging subtypes, computational tools should serve to augment, not replace, expert pathological diagnosis and molecular confirmation \u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, RCCNET represents a significant advance in computational RCC diagnosis by successfully combining technical innovation, morphometric correlation analysis, and a practical clinical integration strategy. The framework\u0026rsquo;s strong performance, robust cross-institutional generalization, and demonstrated economic benefits provide compelling evidence for its potential clinical impact, provided its limitations\u0026mdash;particularly the false positive rate for TFE3 RCC\u0026mdash;are carefully managed in a clinical setting. Our study demonstrates that comprehensive computational frameworks can address complex diagnostic challenges while maintaining the transparency and safety required for clinical practice. Ultimately, this work supports a vision of human-computational collaboration that enhances diagnostic capabilities and provides new insights into disease biology, while reinforcing the continued importance of expert pathological interpretation in ensuring optimal patient care.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eDisclosures:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no potential conflicts of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics Approval and Consent to Participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Institutional Review Boards of both participating institutions (Ethics Approval No.: CHEC-Y2024-077). Due to the retrospective nature of the study using de-identified data, the requirement for informed consent was waived by the ethics committees in accordance with local regulations and international guidelines for retrospective medical research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for Publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable for this retrospective study using de-identified patient data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests, financial or otherwise, that could have influenced the design, conduct, or reporting of this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported in part by the National Science Foundation for Scientists of China (81871352, 82171915, 82171930, 82271972 and 82371955), The Natural Science Foundation of Shanghai Science and Technology Innovation Action Plan (21ZR1478500, 21Y11910300), and Clinical Research Plan of SHDC (SHDC2022CRD028), Shanghai Municipal Health Commission Seed Program for Research and Translation of Medical New Technologies Project (2024ZZ1015), and Plan for Promoting Scientific Research Paradigm Reform and Enhancing Disciplinary Advancement through Artificial Intelligence (2024RGZD001).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contribution\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eC.C., B.X., Q.K., and N.L. designed the study and contributed equally to conceptual development.C.C., Q.Z., Y.S., D.Z., S.M., Y.Z., J.Y., and J.L. collected and curated the pathology and imaging data.X.F., F.L., and M.Y. developed the AI framework and conducted the computational analysis.T.W., L.W., and J.L. contributed to model interpretation and technical validation.S.Z. supervised pathology annotation and morphological interpretation.H.J., C.S., and Y.B. jointly supervised the project and provided critical revisions.C.C., B.X., and Y.B. wrote the initial draft.All authors reviewed and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of Data and Materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and analyzed during the current study are available from the corresponding author on reasonable request, subject to institutional data sharing policies, patient privacy regulations, and appropriate data use agreements. The AI models and analysis code are publicly available as described in the Data and Code Availability Statement. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/CHANGHAI-AILab/RCCNET\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eGoswami, P. R., Singh, G., Patel, T. \u0026amp; Dave, R. The WHO 2022 Classification of Renal Neoplasms (5th Edition): Salient Updates. \u003cem\u003eCureus\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, e58470 (2024). https://doi.org:10.7759/cureus.58470\u003c/li\u003e\n\u003cli\u003ePagarigan, A. K. L., Reyes-Murillo, P. D. \u0026amp; Carbonell, D. J. S. Approach to Diagnosis of TFE3-rearranged Renal Cell Carcinoma in a Limited Resource Setting: A Case Report. \u003cem\u003eJ Kidney Cancer VHL\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 40-44 (2024). https://doi.org:10.15586/jkcvhl.v11i3.338\u003c/li\u003e\n\u003cli\u003eHsieh, J. J.\u003cem\u003e et al.\u003c/em\u003e Renal cell carcinoma. \u003cem\u003eNat Rev Dis Primers\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 17009 (2017). https://doi.org:10.1038/nrdp.2017.9\u003c/li\u003e\n\u003cli\u003eJanowczyk, A. \u0026amp; Madabhushi, A. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. \u003cem\u003eJ Pathol Inform\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 29 (2016). https://doi.org:10.4103/2153-3539.186902\u003c/li\u003e\n\u003cli\u003eVerghese, G.\u003cem\u003e et al.\u003c/em\u003e Computational pathology in cancer diagnosis, prognosis, and prediction - present day and prospects. \u003cem\u003eJ Pathol\u003c/em\u003e \u003cstrong\u003e260\u003c/strong\u003e, 551-563 (2023). https://doi.org:10.1002/path.6163\u003c/li\u003e\n\u003cli\u003eAbu Haeyeh, Y., Ghazal, M., El-Baz, A. \u0026amp; Talaat, I. M. Development and Evaluation of a Novel Deep-Learning-Based Framework for the Classification of Renal Histopathology Images. \u003cem\u003eBioengineering (Basel)\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e (2022). https://doi.org:10.3390/bioengineering9090423\u003c/li\u003e\n\u003cli\u003eArdon, O.\u003cem\u003e et al.\u003c/em\u003e Understanding the financial aspects of digital pathology: A dynamic customizable return on investment calculator for informed decision-making. \u003cem\u003eJ Pathol Inform\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 100376 (2024). https://doi.org:10.1016/j.jpi.2024.100376\u003c/li\u003e\n\u003cli\u003eBankhead, P.\u003cem\u003e et al.\u003c/em\u003e QuPath: Open source software for digital pathology image analysis. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 16878 (2017). https://doi.org:10.1038/s41598-017-17204-5\u003c/li\u003e\n\u003cli\u003eEchle, A.\u003cem\u003e et al.\u003c/em\u003e Deep learning in cancer pathology: a new generation of clinical biomarkers. \u003cem\u003eBr J Cancer\u003c/em\u003e \u003cstrong\u003e124\u003c/strong\u003e, 686-696 (2021). https://doi.org:10.1038/s41416-020-01122-x\u003c/li\u003e\n\u003cli\u003eLinehan, W. M. \u0026amp; Ricketts, C. J. The Cancer Genome Atlas of renal cell carcinoma: findings and clinical implications. \u003cem\u003eNat Rev Urol\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 539-552 (2019). https://doi.org:10.1038/s41585-019-0211-5\u003c/li\u003e\n\u003cli\u003eBossuyt, P. M.\u003cem\u003e et al.\u003c/em\u003e STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies. \u003cem\u003eBMJ\u003c/em\u003e \u003cstrong\u003e351\u003c/strong\u003e, h5527 (2015). https://doi.org:10.1136/bmj.h5527\u003c/li\u003e\n\u003cli\u003eCollins, G. S.\u003cem\u003e et al.\u003c/em\u003e TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. \u003cem\u003eBMJ\u003c/em\u003e \u003cstrong\u003e385\u003c/strong\u003e, e078378 (2024). https://doi.org:10.1136/bmj-2023-078378\u003c/li\u003e\n\u003cli\u003eWang, Z.\u003cem\u003e et al.\u003c/em\u003e Label Cleaning Multiple Instance Learning: Refining Coarse Annotations on Single Whole-Slide Images. \u003cem\u003eIEEE Trans Med Imaging\u003c/em\u003e \u003cstrong\u003e41\u003c/strong\u003e, 3952-3968 (2022). https://doi.org:10.1109/TMI.2022.3202759\u003c/li\u003e\n\u003cli\u003eLei, W.\u003cem\u003e et al.\u003c/em\u003e One-Shot Weakly-Supervised Segmentation in 3D Medical Images. \u003cem\u003eIEEE Trans Med Imaging\u003c/em\u003e \u003cstrong\u003e43\u003c/strong\u003e, 175-189 (2024). https://doi.org:10.1109/TMI.2023.3294975\u003c/li\u003e\n\u003cli\u003eHomeyer, A.\u003cem\u003e et al.\u003c/em\u003e Artificial Intelligence in Pathology: From Prototype to Product. \u003cem\u003eJ Pathol Inform\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 13 (2021). https://doi.org:10.4103/jpi.jpi_84_20\u003c/li\u003e\n\u003cli\u003eGraham, S.\u003cem\u003e et al.\u003c/em\u003e Hover-Net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. \u003cem\u003eMed Image Anal\u003c/em\u003e \u003cstrong\u003e58\u003c/strong\u003e, 101563 (2019). https://doi.org:10.1016/j.media.2019.101563\u003c/li\u003e\n\u003cli\u003eVeta, M., Pluim, J. P., van Diest, P. J. \u0026amp; Viergever, M. A. Breast cancer histopathology image analysis: a review. \u003cem\u003eIEEE Trans Biomed Eng\u003c/em\u003e \u003cstrong\u003e61\u003c/strong\u003e, 1400-1411 (2014). https://doi.org:10.1109/TBME.2014.2303852\u003c/li\u003e\n\u003cli\u003eSrinidhi, C. L., Ciga, O. \u0026amp; Martel, A. L. Deep neural network models for computational histopathology: A survey. \u003cem\u003eMed Image Anal\u003c/em\u003e \u003cstrong\u003e67\u003c/strong\u003e, 101813 (2021). https://doi.org:10.1016/j.media.2020.101813\u003c/li\u003e\n\u003cli\u003eCampanella, G.\u003cem\u003e et al.\u003c/em\u003e Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 1301-1309 (2019). https://doi.org:10.1038/s41591-019-0508-1\u003c/li\u003e\n\u003cli\u003eCheng, J.\u003cem\u003e et al.\u003c/em\u003e Computational analysis of pathological images enables a better diagnosis of TFE3 Xp11.2 translocation renal cell carcinoma. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 1778 (2020). https://doi.org:10.1038/s41467-020-15671-5\u003c/li\u003e\n\u003cli\u003eTabibu, S., Vinod, P. K. \u0026amp; Jawahar, C. V. Pan-Renal Cell Carcinoma classification and survival prediction from histopathology images using deep learning. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 10509 (2019). https://doi.org:10.1038/s41598-019-46718-3\u003c/li\u003e\n\u003cli\u003eNiazi, M. K. K., Parwani, A. V. \u0026amp; Gurcan, M. N. Digital pathology and artificial intelligence. \u003cem\u003eLancet Oncol\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, e253-e261 (2019). https://doi.org:10.1016/S1470-2045(19)30154-8\u003c/li\u003e\n\u003cli\u003eShmatko, A., Ghaffari Laleh, N., Gerstung, M. \u0026amp; Kather, J. N. Artificial intelligence in histopathology: enhancing cancer research and clinical oncology. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 1026-1038 (2022). https://doi.org:10.1038/s43018-022-00436-4\u003c/li\u003e\n\u003c/ol\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-7282796/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7282796/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAccurate renal cell carcinoma (RCC) subtyping, especially challenging TFE3-rearranged RCC, is vital for treatment. We developed RCCNET (RCC Neural Enhancement Technology), a weakly supervised deep learning framework integrating a parallel cellular morphometric module for biological interpretability, for four-class classification (clear cell, papillary, chromophobe, TFE3-rearranged). Validated multicentrically on 340 patients (training n=233; external validation n=107), RCCNET achieved macro-average AUCs of 0.989 (training) and 0.966 (validation). For TFE3 RCC, AUC was 0.976 with 92.3% sensitivity, but a 66.7% positive predictive value necessitates molecular confirmation of all positive cases. Model predictions significantly correlated with quantitative morphological features, grounding decisions in histopathology. An economic analysis projected an RCCNET-assisted workflow could reduce costs by 83.2% and time by 45.2%. RCCNET provides an interpretable, cost-effective solution. We propose a confidence-based clinical integration framework, flagging uncertain TFE3 predictions for pathologist review to manage false positives and ensure safe deployment.\u003c/p\u003e","manuscriptTitle":"Weakly Supervised Learning for Multi-class RCC Classification: Multicenter Validation with Biological Interpretability","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-09-17 08:13:35","doi":"10.21203/rs.3.rs-7282796/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":"3e6565d5-937f-4ac3-9480-28e36769cb19","owner":[],"postedDate":"September 17th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":54672124,"name":"Biological sciences/Cancer"},{"id":54672125,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":54672126,"name":"Health sciences/Oncology"},{"id":54672127,"name":"Health sciences/Urology"}],"tags":[],"updatedAt":"2025-10-02T17:23:28+00:00","versionOfRecord":[],"versionCreatedAt":"2025-09-17 08:13:35","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7282796","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7282796","identity":"rs-7282796","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}