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

Area of Science:

• * Computational Biology
• * Data Visualization
• * Bioinformatics

Background:

• * High-dimensional spatial datasets (e.g., multiplexed imaging, satellite data) are challenging to interpret.
• * Dimensionality reduction techniques often lose crucial spatial and morphological context.
• * Existing methods limit the understanding of complex imaging data.

Purpose of the Study:

• * To present SEAL, an interactive visual analytics system.
• * To bridge the gap between abstract 2D embeddings and spatial imaging context.
• * To enhance the interpretability of high-dimensional spatial datasets.

Main Methods:

• * Developed a novel hybrid-embedding visualization preserving image and morphological information.
• * Adapted set visualization for interactive selection and comparison in embedding and spatial views.
• * Employed a scalable surrogate model to calculate feature importance scores for selected sets.

Main Results:

• * SEAL enables identification and comparison of data subsets in both embedding and spatial contexts.
• * Feature importance scores highlight key attributes driving data distribution in embeddings.
• * Case studies in cancer research and astronomy demonstrate SEAL's effectiveness.

Conclusions:

• * Integrating image context with embedding spaces is crucial for interpreting complex imaging datasets.
• * SEAL significantly enhances interpretability and insight generation for high-dimensional spatial data.
• * The system offers a versatile platform for spatially informed data exploration.