Pathologic Image Classification of Flat Urothelial Lesions Using Pathologic Criteria-Based Deep Learning

Area of Science:

• Uropathology
• Artificial Intelligence in Medicine
• Computational Pathology

Background:

• Pathologic diagnosis of flat urothelial lesions exhibits considerable interobserver variability, impacting diagnostic consistency.
• Deep learning models offer potential for improving diagnostic accuracy but often function as "black boxes."

Purpose of the Study:

• To develop and validate a deep learning approach for classifying flat urothelial lesions that integrates the diagnostic reasoning of pathologists.
• To enhance the accuracy and consistency of pathologic diagnosis for urothelial lesions.

Main Methods:

• Trained six convolutional neural networks on 267 H&E-stained slides (127 cases) to classify images based on six criteria.
• Integrated these networks into a main training framework for final automated diagnosis, mimicking pathologist decision-making.

Main Results:

• The deep learning method significantly improved classification accuracy for flat urothelial lesions compared to manual analysis.
• Achieved near-perfect agreement (weighted κ = 0.98) with consensus diagnosis, demonstrating high reliability.
• The approach provided a reliable diagnosis that aligns with established histologic interpretation.

Conclusions:

• Developed an automated subtype classifier for flat urothelial lesions using deep learning.
• Successfully combined traditional morphologic analysis with advanced deep learning techniques.
• Created a learning mechanism that provides plausible and interpretable diagnostic insights for pathologists.