Radiomics-Based Prediction of Muscle-Invasive Bladder Cancer Using Multi-Parametric MRI

Accurate detection of muscle-invasive bladder cancer is critical for effective management and improved outcomes. This study aimed to characterize radiomic features from magnetic resonance images (MRI) to identify muscle invasive bladder cancer and to compare the predictive accuracy of various classification models trained on these features. We included a cohort of fifty-two bladder cancer subjects who underwent multiparametric magnetic resonance imaging. We manually segmented three-dimensional lesions on the T2-weighted images, Apparent Diffusion Coefficient images, and post-contrast T1-weighted images. From each segmented three-dimensional lesion for each image type, 105 radiomic features were extracted. Using selected radiomic features, we trained and tested five machine learning models: Decision Tree, Random Forest, Logistic Regression, Support Vector Machine, and Extreme Gradient Boosting. We selected the key radiomic features from each type of image, using Leave-One-Out cross-validation. Additionally, we trained the classifiers using the features selected from the combined images, including T2-weighted, Apparent Diffusion Coefficient, and postcontrast T1-weighted images. Pathology results served as the gold standard for the classification of muscle invasive bladder cancer. We measured the performance of the classifiers, calculating the accuracy, area under the ROC curve, sensitivity, and specificity. The radiomic features of the T2-weighted images yielded the highest average accuracy among individual image types. The features extracted from the post-contrast T1-weighted images alone yielded the lowest average accuracy. However, when combining all three image types, the combination produced the highest area under the curve, reaching 0.77, and an accuracy of 0.77. The Decision Tree model trained with the radiomic features from a combination of T2-weighted, Apparent Diffusion Coefficient, and post-contrast T1-weighted images provides improved accuracy in predicting muscle invasive bladder cancer.