SEAL : Spatially-resolved Embedding Analysis with Linked Imaging Data.

Area of Science:

• Data Visualization
• Computational Biology
• Image Analysis

Background:

• Dimensionality reduction techniques project high-dimensional spatial data into 2D, but often lose crucial spatial and morphological context.
• Interpreting complex imaging datasets like tissue, satellite, or astronomical data is challenging due to abstracted spatial information.

Purpose of the Study:

• To introduce SEAL, an interactive visual analytics system that bridges the gap between 2D embeddings and spatial imaging context.
• To enable analysts to preserve and utilize image and morphological information within dimensionality reduction visualizations.
• To facilitate the identification, visualization, and comparison of data subsets within both embedding and spatial views.

Main Methods:

• SEAL employs a novel hybrid-embedding visualization combining 2D projections with original image data.
• It adapts set visualization methods for interactive selection and comparison of data subsets.
• A scalable surrogate model calculates feature importance scores to explain embedding positions and differences between selections.

Main Results:

• SEAL preserves critical spatial, positional, and morphological information lost in traditional dimensionality reduction.
• The system allows for interactive exploration and comparative analysis of selected data subsets using set operations.
• Feature importance scores help identify key attributes driving data organization within embeddings.

Conclusions:

• Integrating image context into embedding spaces is crucial for interpreting high-dimensional imaging datasets.
• SEAL enhances interpretability and insight generation for complex spatial data analysis.
• The system's versatility is demonstrated across biological and astronomical case studies.