Bladder Urothelial Carcinoma: Machine Learning-based Computed Tomography Radiomics for Prediction of Histological Variant

Rationale and Objectives Histological variant (HV) of bladder urothelial carcinoma (UC) is a significant factor for therapy management. We aim to assess the predictive performance of machine learning (ML)-based Computed Tomography radiomics of UC for HV. Materials and Methods Volume of interest of 37 bladder UC tumors, of which 21 were pure and 16 were HV, were manually segmented. The extracted first- and second-order texture features (n = 117) using 3-D Slicer radiomics were compared to the radical cystectomy histopathological results. ML algorithms were performed to determine the significant models using Python 2.3, Pycaret library. The sample size was increased to 74 by synthetic data generation, and three outliers from the training set were removed (training dataset; n = 52, test dataset; n = 19). The predictive performances of 15 ML algorithms were compared. Then, the best two models were evaluated on the test set and ensembled by Voting Classifier. Results The ML algorithms demonstrated area under curve (AUC) and accuracy ranging 0.79-0.97 and 50%-90%, respectively on the train set. The best models were Gradient Boosting Classifier (AUC: 0.95, accuracy: 90%) and CatBoost Classifier (AUC: 0.97, accuracy: 85%). On the test set; the Voting Classifier of these two models demonstrated AUC, accuracy, recall, precision, and F1 scores as follows; 0.93, 79%, 86%, 67%, and 75%, respectively. Conclusion ML-based Computed Tomography radiomics of UC can predict HV, a prognostic factor that is indeterminable by qualitative radiological evaluation and can be missed in the preoperative histopathological specimens. Histological variant (HV) of bladder urothelial carcinoma (UC) is a significant factor for therapy management. We aim to assess the predictive performance of machine learning (ML)-based Computed Tomography radiomics of UC for HV. Volume of interest of 37 bladder UC tumors, of which 21 were pure and 16 were HV, were manually segmented. The extracted first- and second-order texture features (n = 117) using 3-D Slicer radiomics were compared to the radical cystectomy histopathological results. ML algorithms were performed to determine the significant models using Python 2.3, Pycaret library. The sample size was increased to 74 by synthetic data generation, and three outliers from the training set were removed (training dataset; n = 52, test dataset; n = 19). The predictive performances of 15 ML algorithms were compared. Then, the best two models were evaluated on the test set and ensembled by Voting Classifier. The ML algorithms demonstrated area under curve (AUC) and accuracy ranging 0.79-0.97 and 50%-90%, respectively on the train set. The best models were Gradient Boosting Classifier (AUC: 0.95, accuracy: 90%) and CatBoost Classifier (AUC: 0.97, accuracy: 85%). On the test set; the Voting Classifier of these two models demonstrated AUC, accuracy, recall, precision, and F1 scores as follows; 0.93, 79%, 86%, 67%, and 75%, respectively. ML-based Computed Tomography radiomics of UC can predict HV, a prognostic factor that is indeterminable by qualitative radiological evaluation and can be missed in the preoperative histopathological specimens.