Modelpublic

idso-fa1-pathology/ROI-ST

sourceHugging Facemitupdated 6mo agoView on Hugging Face

3likes

Model Card

ROI-ST: Automated ROI Selection for Spatial Transcriptomics Whole-Slide Images

Model Summary

ROI-ST is a multi-scenario framework for automated Region of Interest (ROI) selection in whole-slide images (WSIs) intended for downstream spatial transcriptomics (ST) analysis. It combines a trained tile-level classifier, TME segmentation masks, and generalist/specialist foundation model embeddings to identify spatially informative ROIs and benchmark them against manual pathologist annotations.

The framework comprises three independent scenarios of increasing complexity, each targeting a different combination of inputs and use cases. It is released to support reproducibility of the results reported in [paper title].

Scenarios Overview

Scenario	Approach	Key Input	Key Output
Scenario 1 — GFM	Tile-level classifier on foundation model embeddings	HDF5 embeddings + WSI	Probability heatmap, predicted ROI window

Intended Use

Primary Use

This framework is intended for research use in computational pathology and spatial transcriptomics, specifically:

—Automated ROI candidate identification from WSIs prior to ST placement
—Benchmarking AI-driven ROI selection against manual pathologist annotations
—Reproducibility of results reported in [paper title]

Out-of-Scope Use

—Clinical diagnostic decision-making
—Use without appropriate pre-computed embeddings or segmentation masks
—Deployment outside the H&E / SVS / OME-TIFF formats described below

System Requirements

Component	Version
Python	3.11.0
PyTorch	2.1.2 + CUDA 12.1
TorchVision	0.16.2
pathologyfoundation (PLIP)	0.1.14
trident (UNI-V2)	0.2.0
tmesegformer	0.1.0
OpenSlide-Python	1.3.1
NumPy / Pandas / Matplotlib / SciPy / scikit-learn	see `requirements.txt`

GPU requirements: ≥ 24 GB VRAM recommended for WSI-scale inference.

Hardware	CUDA	Status
NVIDIA H100	12.x	✅ Validated
NVIDIA A100	12.x	✅ Validated
NVIDIA RTX 4090	12.x	✅ Validated

Installation

bash

python3.11 -m venv roi_env
source roi_env/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
python -m pip check

Scenario 1 — GFM: Classifier on Foundation Model Embeddings

Uses a trained classifier (tile_classifier.pkl) applied to tile-level embeddings (HDF5) to compute per-tile probabilities and identify the highest-probability ROI window.

Inputs

Argument	Description
`--model_dir`	Directory containing `classifier.pkl`
`--test_h5`	HDF5 file with `coords`, `features`, `patch_size_level0`
`--test_tif`	Whole-slide `.tif` image
`--roi_json`	JSON file with manual ROI coordinates `{x0, y0, x1, y1}`
`--out_dir`	Output directory
`--thumb_width`	Thumbnail width in pixels (default: 1000)

Outputs

File	Description
`probability_heatmap.png`	Tile probability heatmap with predicted ROI overlay
`probability_histogram.png`	Distribution of tile-level probabilities
`thumbnail_overlay.png`	Thumbnail with manual vs predicted ROI rectangles

Usage

bash

python3 scenario_1.py \
  --model_dir "./model" \
  --test_h5 "./uni_embeddings/gbm_xxx.h5" \
  --test_tif "./raw_tif/gbm_xxx.tif" \
  --roi_json "./roi_json/gbm_xxx_roi.json" \
  --out_dir "./output"

Citation

If you use this framework, please cite:

bibtex

@article{[citation_key],
  title   = {[Paper title]},
  author  = {[Authors]},
  journal = {[Journal]},
  year    = {[Year]},
  doi     = {[DOI]}
}

License

This model is released under the MIT License. See LICENSE for details.