CT-based multimodal fusion model for preoperative clear cell renal cell carcinoma grading: A multicenter study

To develop and evaluate a multimodal fusion framework for the preoperative World Health Organization/International Society of Urological Pathology (WHO/ISUP) low- and high-grade classification of clear cell renal cell carcinoma (ccRCC), integrating clinical characteristics, deep learning (DL), and radiomic signatures extracted from corticomedullary phase (CMP) contrast-enhanced computed tomography (CECT) images.

In this retrospective multicenter study, we evaluated a cohort of 1,052 patients with histologically verified ccRCC across four institutions. The data of the primary cohort (Center A, n = 665) were randomly divided into a training dataset (n = 465) and an internal validation dataset (n = 200). The data of 387 additional patients from three institutions (Centers B-D) constituted the independent external test set. Initial single-modality classifiers were developed via ensemble machine learning, utilizing clinical variables, multiscale radiomic data (capturing intratumoral, habitat, and peritumoral regions), and DL representations (2D, 2.5D, and 3D). A soft voting mechanism was subsequently applied to integrate the top-performing base models. Model interpretability was determined using SHapley Additive exPlanations (SHAP) values, and overall diagnostic efficacy was quantified through receiver operating characteristic (ROC) curve analysis.

The weighted soft voting model (SoftVote-Opt) showed the numerically highest AUC among the evaluated fusion strategies and single-modality models. This model yielded AUCs of 0.761 (95 % CI: 0.675-0.839) and 0.745 (95 % CI: 0.675-0.813), with corresponding MCCs of 0.347 (95 % CI: 0.217-0.475) and 0.302(95 % CI: 0.205 -0.387), within the internal validation and external test sets, respectively. SHAP analysis revealed peritumoral 5 mm radiomic features as the primary determinants of the model's output, followed by intratumoral radiomic features, whereas deep learning and clinical features offered lower relative contributions.

The multimodal fusion framework demonstrated moderate discriminatory performance across the internal validation and external test cohorts. These findings could be regarded as an exploratory study rather than evidence of established clinical superiority. The framework may provide complementary information for preoperative ccRCC grading.

This study evaluates an exploratory multimodal fusion framework for preoperative ccRCC grading by combining clinical data, deep learning, and multiscale radiomic features. The framework may provide complementary non-invasive information for risk stratification.