Radiomics has been widely studied and used to predict muscle invasive bladder cancer (MIBC). However, it is not known whether adding manually quantified 3D morphological features to radiomics could improve the diagnostic performance of MIBC. This study aims to examine the diagnostic performance of integrating quantified 3D morphological features with radiomics for MIBC diagnosis.

We conducted a study with 106 bladder cancer patients diagnosed pathologically, who underwent preoperative MRI from August 2021 to July 2023. We manually extracted 3D radiomics features from T2WI using 3D Slicer (4.8.1) and PyRadiomics (3.1.0). We also manually quantified the 3D morphological features and added them to the software-extracted radiomics features. The MIBC diagnostic performance of several machine learning models was validated in JMP Pro (18.0.2). The dataset was randomly divided into a training set and a validation set at a ratio of 4:1. Multiple models were applied, and each was trained and evaluated using 10-fold stratified cross-validation with 5 repetitions. For each fold, training and validation AUCs were calculated, and an average ROC curve was generated by interpolating across all folds.

This study included 106 patients (88 men, 18 women; median age: 73.0 years), of whom 32 had MIBC and 74 had NMIBC. Compared to the results using radiomics alone, the integration of manually quantified 3D morphological features improved the diagnostic performance of each model. AUC values increased in Generalized Regression Lasso (0.900 to 0.945), Support Vector Machine (0.865 to 0.904), and Boosted Tree (0.903 to 0.919), indicating enhanced discriminative ability. Neural Boosting (0.939 to 0.951) and Bootstrap Forest (0.951 to 0.963) also showed improvement while maintaining high accuracy.

The integration of software-derived and manually quantified radiomics features from T2WI-MRI demonstrated high diagnostic performance for predicting MIBC using multiple machine learning models.