Non-muscle invasive bladder cancer (NMIBC) is a highly recurrent malignancy with critical prognostic implications. This study aims to develop and validate a fusion radiomics-pathological omics predictive model for accurate prediction of postoperative recurrence and survival risk in NMIBC patients.

This study enrolled 193 patients with non-muscle-invasive bladder cancer (NMIBC) who received standard transurethral resection of bladder tumor (TURBT) prior to surgery, from 2017 to 2021 at Xuzhou Central Hospital and Nanjing Drum Tower Hospital Group Suqian Branch. Preoperative CT imaging data and postoperative histopathological sections were collected. CT images were annotated using ITC-SNAP software to extract radiomics features, with key characteristics identified through LASSO regression analysis. Pathomics features were generated via whole-slide imaging scanning, analyzed using multiple deep learning models combined with gradient-weighted category activation mapping (Grad-CAM). A predictive model was constructed by integrating radiomics, pathomics, and clinical features through various machine learning algorithms, validated using a randomly divided training set, internal validation set, and external test set. Model performance was evaluated using receiver-operating characteristic (ROC) curves and concordance indices. Additionally, Delong test, Hosmer-Lemeshow test, calibration curves, and clinical decision curves were employed to confirm the model's accuracy and clinical utility.

In radiomics feature selection, LASSO regression analysis identified 14 features significantly associated with recurrence risk, most of which were related to tumor-adjacent tissues, highlighting the tumor microenvironment's critical role in recurrence assessment. Pathological omics analysis demonstrated that the ResNet101 deep learning model achieved high accuracy in distinguishing tumor from non-tumor regions, with an area under the receiver operating characteristic (AUC) curve of 0.762 in the test cohort, indicating its effectiveness in pathological image analysis. The radiopathological omics model for predicting recurrence showed excellent performance across training and external test cohorts, achieving AUC values of 0.974 and 0.854 respectively. For survival prediction models, consistency indices reached 0.921,0.810, and 0.741 in training, validation, and test sets respectively, confirming their accuracy in predicting postoperative survival for patients with non-muscle-invasive bladder cancer.

This study has successfully developed two predictive models that integrate clinical, radiomics, and pathomics features to evaluate postoperative survival and recurrence risks in patients with non-muscle-invasive bladder cancer. The clinical application of these models is expected to provide physicians with more accurate risk assessment tools, guiding more rational patient management and personalized treatment strategies.