Data Sheet 2_SpatialFinder: a human-in-the-loop vision-language framework for prioritizing

SpatialFinder is a human-in-the-loop vision-language framework for ranking high-value regions of interest (ROIs) on H&E tissue slides for spatial transcriptomics. The dataset, authored by Jonathan Xu and last updated in April 2026, contains evaluation results from four Visium HD tissue types. The framework aims to make spatial transcriptomics more cost-effective by predicting gene expression heterogeneity from standard histology images.

Use Cases

• Ranking high-value regions of interest (ROIs) for spatial transcriptomics based on H&E slide analysis
• Predicting gene expression heterogeneity from standard histology images using a vision-language model
• Optimizing human-AI collaboration for tumor-targeted or diversity-targeted ROI selection

Strengths

• Framework evaluated across four distinct Visium HD tissue types
• Achieved performance metrics up to Spearman's ρ of 0.89 and Overlap@10% of 78.8%
• Demonstrated an absolute 24.9 percentage-point gain over the strongest vision-language model baseline

Limitations

• Column-level documentation is absent; field semantics must be inferred after download
• Row count is unknown, which may limit suitability assessment
• The 2.8 MB size suggests the dataset is likely limited to framework outputs or metadata, not raw image data

Provenance

Source

Jonathan Xu via figshare

Collection Method

Results from a biomedical vision-language model (VLM) and human-in-the-loop optimization pipeline evaluated on Visium HD data.

Time Range

The dataset itself does not specify a temporal coverage for the underlying tissue samples.

Freshness

Last updated 2026-04-15 05:22:59

Geography

The dataset does not specify a geographic coverage.