Interactive phenotyping of large-scale histology imaging data with HistomicsML

Area of Science:

• Digital Pathology
• Computational Biology
• Machine Learning

Background:

• Whole-slide imaging generates high-resolution data of tissue microenvironments.
• Extracting quantitative features from millions of histologic objects presents a scalability challenge.
• Developing efficient methods for training classification rules is crucial for digital pathology research.

Purpose of the Study:

• To introduce HistomicsML, an interactive machine learning system for digital pathology imaging datasets.
• To enable efficient and scalable classifier training using active learning for large datasets.
• To facilitate the discovery of prognostic image biomarkers and genotype-phenotype associations.

Main Methods:

• Developed HistomicsML, an interactive machine learning framework.
• Utilized active learning to guide user feedback for classifier training.
• Applied the system to phenotype microvascular structures in gliomas and explore molecular pathways.

Main Results:

• Demonstrated efficient and scalable classifier training on datasets with over 10^8 histologic objects.
• Successfully phenotyped microvascular structures in gliomas to predict survival.
• Enabled exploration of molecular pathways associated with identified phenotypes.

Conclusions:

• HistomicsML enhances the utility of digital pathology data by enabling efficient feature extraction and analysis.
• The system unlocks phenotypic information for investigating prognostic image biomarkers.
• Facilitates genotype-phenotype association studies in cancer research.