HistoNet: A Deep Learning-Based Model of Normal Histology.

Area of Science:

• Computational pathology
• Digital histology
• Machine learning in toxicology

Background:

• Histopathology is crucial for preclinical toxicology studies.
• Manual tissue analysis is time-consuming and subjective.
• Deep learning offers potential for automated image analysis.

Purpose of the Study:

• To develop and evaluate HistoNet, a deep neural network for automated tissue classification.
• To assess the performance of different network architectures (VGG-16, ResNet-50, Inception-v3) at various magnification levels.
• To explore the learned features for potential downstream applications.

Main Methods:

• Trained deep neural networks (VGG-16, ResNet-50, Inception-v3) on 1690 annotated rat tissue slides across 6 magnification levels.
• Utilized 4 studies for training and 2 for testing.
• Employed Uniform Manifold Approximation and Projection (UMAP) to visualize learned features.

Main Results:

• Inception-v3 and ResNet-50 outperformed VGG-16, with Inception-v3 achieving up to 83.4% accuracy in tissue identification.
• Misclassifications were primarily between histologically similar tissues.
• UMAP revealed meaningful subclusters within tissue embeddings, indicating deeper histological understanding.
• Models trained on rat tissues showed cross-species applicability to non-human primate and minipig tissues with minimal retraining.

Conclusions:

• HistoNet demonstrates high accuracy in classifying rat tissues, outperforming simpler models.
• The learned histological representations are robust and may serve as a foundation for other machine learning tasks.
• The model's ability to generalize across species highlights its potential utility in comparative toxicology.