Gradio

VOICE · H&E workbench

Single-cell expression and Stage 2 embeddings from an external H&E region.

H&E region · PNG/JPEG/TIFF · ≤1024 px per side

Actual resolution (µm/pixel)

Required. Supported range: 0.1–0.5. Images are never silently downsampled.

Gene symbols

Comma-separated symbols, or all for 18,580 genes. First gene colors the preview.

Tissue

Select the matching tissue. Choose Global if it is not listed. These choices come from the loaded fusion weights; reference retrieval still uses the full bank.

This image is de-identified and I approve processing on Hugging Face.

This image is external to the VOICE training/reference datasets.

Free shared GPU · up to 120 s GPU time · ≤512 nuclei. Queue and daily quota apply. Use the CLI for larger images or reference exclusions. Input copies are removed after processing; results are cleaned after approximately one hour while the app runs. Hugging Face/Gradio upload caches may remain until their cleanup cycle.

Predicted nuclear boundaries

Expression and Stage 2 embedding preview

Result summary

AnnData · expression + all embeddings

Embeddings · NPZ

All results · ZIP

Research use only. Expression values are predictions, not measured counts or a diagnosis. Nuclear boundaries are not complete cell boundaries. VOICE uses UNI2-h (Mahmood Lab), scGPT and Cellpose-SAM; see their original publications and terms.