Aim: To develop a radiomics model to predict the muscle-invasive status of bladder cancer (BC) in contrast-enhanced computed tomography (CECT) images, compared with radiologists' interpretations.
Materials and methods: One hundred and eighty-eight CECT images with histopathologically confirmed BC were retrieved retrospectively from November 2018 to December 2019 and were divided randomly into the training (n=120) and test dataset (n=68). The BC were annotated manually and validated on the venous phase by a general radiologist and an experienced radiologist, respectively. The radiomics analysis included radiomics feature extraction and model development. The same images were also evaluated by two radiologists. The diagnostic performance of radiomics was evaluated using receiver operating characteristic (ROC) curve analysis and the area under the ROC curve (AUC), sensitivity, and specificity were calculated. The predictive performance of radiomics was then compared to visual assessments of the two radiologists.
Results: The radiomics model reached an AUC (95% confidence interval [CI]) of 0.979 (0.935-0.996) and 0.894 (0.796-0.956) in the training and test dataset, respectively. The radiomics model outperformed the visual assessment of radiologist A and B both in the training (0.865 [0.791-0.921], 0.894 [0.824-0.943]) and test dataset (0.766 [0.647-0.860], 0.826 [0.715-0.907]). Pairwise comparisons showed that the specificities of the radiomics model were higher than the radiologists (85.3-96.7% versus 47.1-58.3%, all p<0.05), but the sensitivities were comparable between the radiomics and the radiologists (79.4-90% versus 91.2-96.7%; all p>0.05).
Conclusions: A radiomics model was developed that outperformed the radiologists' visual assessment in predicting the muscle-invasive status of BC in the venous phase of CT images.
Copyright © 2022 The Royal College of Radiologists. Published by Elsevier Ltd. All rights reserved.