ISUP (International Society of Urological Pathology) nuclear grade is one of the most important prognostic factors in clear cell renal cell carcinoma (ccRCC), yet grading requires invasive pathological examination and is subject to inter-observer variability. Non-invasive or AI-assisted grading could improve consistency and enable preoperative risk stratification.
CT radiomics captures tumor imaging phenotypes while pathomics from histological whole slide images (WSIs) quantifies tumor microarchitecture and nuclear morphology. Combining these two modalities in a multimodal predictive signature (MPS) has the potential to capture complementary aspects of tumor biology that neither modality captures alone.
This multicenter cohort study developed and validated a multimodal AI model combining CT radiomics (RBS) and pathomics (PBS) to predict ISUP nuclear grade and overall survival in ccRCC patients across three independent cohorts.
The study enrolled 729 patients across three cohorts: a General cohort of 512 patients used for primary training and validation, the publicly available TCGA cohort of 175 patients, and the CPTAC cohort of 42 patients. All patients had both CT imaging and H&E stained whole slide images available.
The Radiomic-Based Signature (RBS) model was built by extracting features from the CT corticomedullary phase and selecting 8 final features using LASSO regression. The Pathological-Based Signature (PBS) model used Hover-Net, a deep learning tool for nuclear instance segmentation and classification in H&E WSIs, extracting 16 morphological and spatial features from tumor nuclei.
The Multimodal Predictive Signature (MPS) combined RBS and PBS features as inputs to an XGBoost classifier. SHAP (SHapley Additive exPlanations) analysis was performed to identify and rank the most influential features in the final model, providing interpretability alongside predictive performance.
Model performance was assessed by AUC for ISUP grade classification and C-index for overall survival prediction, with both metrics computed separately for each of the three cohorts.
The RBS model achieved AUC of 0.93 in the General training cohort, with validation AUCs of 0.89, 0.83, and 0.88 in the General validation, TCGA, and CPTAC cohorts respectively. The PBS model performed comparably with AUC of 0.94 in training and 0.90, 0.88, and 0.89 in the three validation cohorts.
The combined MPS model substantially outperformed both individual modalities, achieving AUC of 0.97 in training and 0.95, 0.93, and 0.95 across the three external validation cohorts. This consistent performance gap confirms that CT and pathological features provide genuinely complementary predictive information.
For overall survival prediction, the MPS achieved C-index of 0.75 in TCGA and 0.71 in CPTAC. Univariate Cox regression showed hazard ratio of 2.542 (p below 0.0001), and multivariate analysis confirmed HR of 1.723 (p = 0.003) after adjusting for clinical covariates, establishing the MPS as an independent prognostic factor.
Hover-Net is a deep learning model designed for simultaneous nuclear instance segmentation and classification in H&E whole slide images. It produces accurate nucleus boundaries and cell type labels across large WSI tiles, enabling extraction of quantitative nuclear morphology features at scale without manual annotation.
From the 16 pathomics features extracted via Hover-Net, morphological features related to nuclear area, shape, and spatial distribution were captured. SHAP analysis identified Morph_Area (nuclear area) as the single most important feature in the MPS model, consistent with the known correlation between large nuclear size and high ISUP grade in ccRCC.
The 8 CT radiomic features selected by LASSO from the corticomedullary phase included texture and intensity-based descriptors capturing tumor heterogeneity. The complementarity between these macroscopic CT features and microscopic nuclear pathomics features explains the substantial performance gain achieved by the multimodal combination.
The PBS and MPS models require pathological WSI data that becomes available after biopsy or surgical resection, supporting postoperative grade verification and prognostic risk stratification. The automated Hover-Net pipeline could replace or supplement manual ISUP grading, reducing inter-pathologist variability.
The RBS model, using only preoperative CT, provides a purely non-invasive grading estimate that could inform surgical planning before biopsy. For patients under active surveillance for small renal masses, serial CT-based RBS scoring could track tumor grade progression without requiring repeated biopsy.
The MPS prognostic C-index of 0.75 in TCGA and 0.71 in CPTAC suggests meaningful ability to stratify patients by survival risk beyond standard clinicopathological staging. Patients identified as high MPS risk could be prioritized for adjuvant therapy enrollment or intensified surveillance protocols.
This study establishes that combining CT radiomics and WSI pathomics through an XGBoost multimodal framework provides substantially better ISUP grade prediction than either modality alone, with consistent performance across three independent cohorts including public TCGA and CPTAC datasets.
The SHAP-based feature attribution increases model transparency, an important factor for clinical acceptance. Understanding that nuclear area is the dominant driver aligns with established pathology principles and supports clinician trust in the model's predictions.
Future directions include expanding the model to distinguish all four ISUP grades rather than binary classification, incorporating genomic and clinical variables as additional modalities, and conducting prospective clinical trials to validate whether MPS-guided management decisions improve patient outcomes over current standard-of-care grading approaches.