Comparative Evaluation of Transfer Learning Models for Detecting Malignant Cells in Urinary Cytology

Area of Science:

• Artificial Intelligence in Medicine
• Computational Pathology
• Uro-oncology Diagnostics

Background:

• Urine cytology is crucial for detecting urothelial cell carcinoma (UCC).
• Automated detection of malignant cells can enhance diagnostic efficiency.
• Transfer learning models offer potential for image-based medical diagnostics.

Purpose of the Study:

• To compare the diagnostic performance of six transfer learning models for detecting malignant cells in urine cytology.
• To evaluate the impact of ensemble learning with weighted soft voting on diagnostic accuracy.
• To assess the potential of these AI techniques for routine urine screening.

Main Methods:

• Utilized 104 histopathology-proven high-grade UCC cases and 86 benign cases.
• Applied six transfer learning models (DenseNet121, inception_v3, ResNet50, MobileNetV2, VGG16, Xception) to 1369 malignant and 446 benign urine images.
• Implemented dynamic training optimization and ensemble learning with soft voting using the top three models.

Main Results:

• The Xception model demonstrated the highest individual performance (sensitivity 88.57%, accuracy 86.55%).
• All individual transfer learning models achieved an Area Under the Curve (AUCROC) of ≥0.90.
• The ensemble model achieved superior performance: 92.10% accuracy, 95.41% sensitivity, 85.51% specificity, 91.23% precision, and 0.977 AUCROC.

Conclusions:

• Transfer learning models show high sensitivity, specificity, and accuracy in urine cytology for malignant cell detection.
• Ensemble learning with soft voting significantly enhances diagnostic accuracy compared to individual models.
• Transfer learning and ensemble methods hold substantial promise for routine urine cancer screening.