Data Sheet 3_SpatialFinder: a human-in-the-loop vision-language framework for prioritizing

SpatialFinder is a 3.5 MB dataset containing a framework that combines a biomedical vision-language model with a human-in-the-loop pipeline to predict gene expression heterogeneity from H&E tissue slides. It was authored by Jonathan Xu and last updated on April 15, 2026. The framework was evaluated across four Visium HD tissue types, achieving performance metrics like Spearman’s ρ up to 0.89.

Use Cases

• Rank high-value regions of interest (ROIs) for spatial transcriptomics based on H&E slide analysis.
• Predict gene expression heterogeneity from standard H&E tissue images using a vision-language model.
• Optimize a human-in-the-loop pipeline for targeting specific tissue microenvironments like tumor regions.
• Benchmark vision-language model performance for diversity- and tumor-targeted ROI ranking tasks.

Strengths

• Framework evaluated across four distinct Visium HD tissue types, providing multi-tissue validation.
• Achieved a Spearman’s ρ correlation of up to 0.89 and an Overlap@10% metric of 78.8% in benchmarks.
• Demonstrated an absolute 24.9 percentage-point gain over the strongest vision-language model baseline.
• Released under a permissive CC-BY-4.0 license, facilitating open reuse and modification.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• The 3.5 MB size suggests the dataset is likely a framework or code repository, not a large-scale primary data collection.

Provenance

Source

figshare

Collection Method

Likely contains the code, model, and evaluation data for the SpatialFinder framework as described in the associated research.

Time Range

null

Freshness

Last updated 2026-04-15 05:22:58; freshness should be verified.

Geography

null