Facetto: Combining Unsupervised and Supervised Learning for Hierarchical Phenotype Analysis in Multi-Channel Image...

Area of Science:

• Digital Pathology and Computational Biology
• High-dimensional data analysis in biomedical imaging

Background:

• Digital histology, using high-dimensional, multi-channel microscopy images, is revolutionizing cancer research and diagnosis.
• Analysis of complex tissue images (billions of pixels, 60+ channels, millions of cells) is challenging for manual and existing automated methods.
• Current automated approaches often fail to leverage expert knowledge from anatomic pathologists.

Purpose of the Study:

• To introduce Facetto, a scalable visual analytics application for semi-automated discovery of single-cell phenotypes in human tumor and tissue images.
• To enable effective integration of unsupervised and supervised learning for image and feature exploration in digital pathology.
• To provide tools for analytical provenance and hierarchical tracking of analysis steps for improved cell type identification.

Main Methods:

• Facetto integrates unsupervised and supervised machine learning techniques for cell type and state discovery.
• It allows users to cluster data for discovering new cell types and train convolutional neural networks for cell classification.
• A hierarchical approach is implemented for tracking analysis steps and data subsets, facilitating phenotype tree construction.

Main Results:

• Facetto enables semi-automated analysis, allowing experts to discover novel cancer and immune cell types through data clustering.
• The application facilitates the use of clustering results to train classifiers, enabling accurate cell classification.
• Users can cluster classifier outputs to identify aggregate patterns and phenotype subsets, gaining scientific insights into cancer biology.

Conclusions:

• Facetto provides a powerful platform for semi-automated analysis of complex, high-dimensional microscopy images in digital pathology.
• The application enhances the discovery of single-cell phenotypes and aids in understanding cancer biology by integrating expert knowledge with machine learning.
• Facetto assists domain scientists in steering analysis, inspecting results, and generating new scientific insights from large-scale imaging datasets.