Distinguishing normal and abnormal tissues in nonclinical toxicity studies using unsupervised representation...

Area of Science:

• Computational pathology
• Drug development
• Artificial intelligence in toxicology

Background:

• Nonclinical toxicity studies rely on manual pathology, which is time-consuming and labor-intensive.
• Automating the analysis of whole-slide images (WSIs) is crucial for efficient drug development.

Purpose of the Study:

• To develop and validate an AI-based workflow for automated abnormality detection and severity grading in rat liver WSIs.
• To reduce pathologist workload and improve the efficiency of nonclinical toxicity assessments.

Main Methods:

• Utilized an unsupervised Bidirectional Generative Adversarial Network (BiGAN) trained on normal histology to identify tissue abnormalities.
• Developed a machine learning-based severity grade classifier (SGC) using BiGAN's output for severity estimation.
• Employed heatmap visualizations to identify tile-level abnormalities.

Main Results:

• The BiGAN model achieved an area under the curve (AUC) of 0.77 for abnormality discrimination.
• The SGC model predicted slide-level severity grades with AUCs ranging from 0.68 to 0.96.
• Heatmap visualizations showed good to fair agreement with ground truth abnormalities in 70% of WSIs.

Conclusions:

• The AI-based workflow effectively identifies normal and abnormal tissues and estimates severity grades in nonclinical toxicity studies.
• This approach eliminates the need for manual annotation and can be rapidly adapted to new tissues and species.
• The developed AI models offer a scalable solution to streamline nonclinical pathology, reducing pathologist input and enhancing efficiency.