Leveraging explainable AI and large-scale datasets for comprehensive classification of renal histologic types

Area of Science:

• Digital Pathology
• Artificial Intelligence in Oncology
• Genitourinary Pathology

Background:

• Increasing cancer incidence necessitates advanced diagnostic tools.
• Current AI research in renal pathology primarily focuses on cancer subtype classification.
• A gap exists in classifying broader renal tissue categories: normal, benign, and malignant, due to limited datasets.

Purpose of the Study:

• To introduce an artificial intelligence model for classifying renal tissue into three categories: normal (non-neoplastic), benign tumor, and malignant tumor.
• To address the understudied area of broad renal tissue categorization.
• To leverage digital pathology and machine learning for improved diagnostic capabilities.

Main Methods:

• Utilized a dataset of 12,223 whole slide images (WSIs) from 2,535 patients across multiple institutions.
• Employed the ResNet-18 architecture with a Multiple Instance Learning approach.
• Trained and validated the model on WSIs of normal, benign, and malignant renal tissues.

Main Results:

• The model achieved high F1-scores: 0.934 for normal tissue, 0.684 for benign tumors, and 0.878 for malignant tumors.
• Demonstrated a weighted average F1-score of 0.879 and a weighted average area under the receiver operating characteristic curve of 0.969.
• The model provides a foundational approach for distinguishing key renal tissue types.

Conclusions:

• The developed AI model significantly aids in the swift and accurate diagnosis of renal tissue.
• This approach enhances the ability to differentiate between non-neoplastic, benign, and malignant conditions.
• The study provides a robust tool for genitourinary pathology, addressing a critical need in cancer diagnostics.