Abstract
This study proposes a deep learning approach for classifying normal, noninvasive, and invasive urothelial neoplasms via digitized histopathologicalimages. Despite many artificial intelligence (AI) models for cancer diagnosis, few focus on bladder lesions or differentiate between these critical categories. We developed convolutional neural networks (CNNs) and transformer-based models, which were trained on 12,500 whole-slide images (WSIs) from five institutions, with preprocessing steps including stain normalization and patch extraction. Fivefold cross-validation was used for evaluation against expert-annotated labels. Among tested models, EfficientNet-B6 achieved the highest performance, with an accuracy of 0.913 (95% confidence interval (CI), 0.907-0.920), sensitivity of 0.909 (95% CI, 0.904-0.914), specificity of 0.956 (95% CI, 0.953-0.960), F1-score of 0.906 (95% CI, 0.901-0.911), and an area under the receiver operating characteristic curve (AUC) of 0.983 (95% CI, 0.982-0.984). These results demonstrate the effectiveness and generalizability of AI-based bladder cancer classification.