Journal: Scientific reports
This study presents a deep learning approach using convolutional neural networks and transformer-based models to classify normal, noninvasive, and invasive urothelial neoplasms from digitized histopathological images.
The models were trained on 12,500 whole-slide images from multiple institutions, incorporating stain normalization and patch extraction techniques.
Key results include:
- EfficientNet-B6 performed best, achieving:
- Accuracy: 91.3%
- Sensitivity: 90.9%
- Specificity: 95.6%
- F1-score: 90.6%
- AUC: 0.983
These findings highlight the potential of AI to effectively and reliably assist in bladder cancer classification.