Related Experiment Video
Updated: May 5, 2026
Magnetic Resonance Imaging Assessment of Carcinogen-induced Murine Bladder Tumors
Published on: March 29, 2019
An Artificial Intelligence-Based Method for Risk Stratification of Urothelial Carcinoma from Liquid-Based Urine
Lei Xiong1, Jia Li1, Xinyi Jin2
1Department of Oncology, The First Affiliated Hospital of Dalian Medical University, Dalian, China.
Introduction:
Urine cytology is a noninvasive and widely used approach for the early detection of urothelial carcinoma (UC), but its diagnostic accuracy is limited, particularly for low-grade lesions. This study aimed to develop a novel artificial intelligence (AI)-based framework for risk stratification of UC from whole-slide images (WSIs), offering a promising solution to enhance the diagnostic accuracy of urine cytology.
Methods:
A total of 385 urine cytology slides were included and stratified into three diagnostic groups based on cytological evaluation: negative for high-grade urothelial carcinoma (NHGUC), low risk (including atypical urothelial cells and low-grade urothelial carcinoma [LGUC]), and high risk (including suspicious for high-grade urothelial carcinoma and high-grade urothelial carcinoma). Following digitization into WSIs, expert pathologists conducted detailed cell-level annotation. Cell detection and segmentation were performed using RTMDet and DuckNet, and the extracted features were aggregated into slide-level representations for training and evaluation of classification models.
Results:
Support vector machine demonstrated the highest overall performance among the classifiers, with an accuracy of 79%, recall of 79%, and a specificity of 90%. The model demonstrated strong classification performance across three risk stratifications. The high-risk group achieved a sensitivity of 73.1% and specificity of 90.2%, while the low-risk group showed a sensitivity of 81.8% and specificity of 89.1%. Precision-recall curves indicated that the NHGUC group achieved the highest average precision, reaching 0.93, followed by the high-risk group at 0.85 and the low-risk group at 0.82. ROC analysis further demonstrated strong discriminative capability for three risk groups, with the area under the curve measured at 0.95 for NHGUC and 0.91 for both the low-risk and High-risk groups.
Conclusion:
The proposed AI-assisted framework shows robust and interpretable performance in stratifying UC cytological categories from WSIs. It holds strong potential as a supportive tool in urine cytology, especially in assisting with the diagnosis of high-risk UC cases.
