Related Experiment Video
Updated: Jul 20, 2025
13:01
Industrialized, Artificial Intelligence-guided Laser Microdissection for Microscaled Proteomic Analysis of the Tumor Microenvironment
Published on: June 3, 2022
3.8K
A Weakly Supervised Deep Learning Framework for Whole Slide Classification to Facilitate Digital Pathology in Animal
Nicole Bussola1, Joshua Xu2, Leihong Wu2
1Center for Integrative Biology, University of Trento, Trento 38123, Italy.
Chemical Research in Toxicology
|August 4, 2023
Summary
PathologAI uses artificial intelligence (AI) for automated animal pathology slide analysis, reducing time and cost. This weakly supervised AI system accurately predicts liver necrosis in rat studies without pixel-level annotations.
Area of Science:
- Digital pathology
- Toxicology
- Artificial intelligence in preclinical research
Background:
- Manual histopathological examination of animal study slides is time-consuming and prone to interobserver variability.
- Current AI model training requires costly pixel-level annotations, which are often unavailable for animal pathology.
- Automated analysis can enhance efficiency, consistency, and accuracy in toxicity evaluations.
Purpose of the Study:
- To develop and evaluate PathologAI, a weakly supervised artificial intelligence system for classifying whole slide images (WSIs) in rat liver pathology.
- To enable accurate prediction of liver necrosis without requiring pixel-level lesion annotations.
- To provide a rapid and cost-effective screening tool for preclinical toxicological studies.
Main Methods:
- Developed PathologAI, a weakly supervised approach for WSI classification using Generative Adversarial Network (GAN) for tile-level preprocessing.
- Utilized an ensemble model of 5 Convolutional Neural Network (CNN) classifiers trained on rat liver imaging data from the Open TG-GATEs system.
- Applied the system to predict necrosis in 816 WSIs, including 377 controls, from studies involving 170 compounds at various doses and time points.
Main Results:
- Achieved notable classification accuracy: 87% for controls without findings, 83% for controls with spontaneous necrosis.
- Demonstrated 67% accuracy for treated animals with spontaneous minimal/slight necrosis and 59% for treatment-induced minimal/slight necrosis.
- Successfully discriminated spontaneous from treatment-related necrosis, differentiated mild lesion levels, and identified treatment dose levels.
Conclusions:
- PathologAI offers an effective, weakly supervised method for automated WSI classification in animal pathology.
- The system provides an inexpensive and rapid screening tool for digital pathology analysis in preclinical and toxicological studies.
- AI-driven analysis can overcome limitations of manual slide examination, improving consistency and efficiency.
