GRASP: Graph-Structured Pyramidal Whole Slide Image Representation

📍 Early version Published at the LMRL Workshop @ ICLR 2025

GRASP is a lightweight, graph-based, multi-magnification framework for subtype prediction of whole slide images (WSIs) in computational pathology. It emulates how expert pathologists navigate through different magnification levels when diagnosing tissue samples.

⚡ GRASP sets a new SOTA for subtype classification across three cancer datasets, while being 5–15× smaller than other top models.

---

🔍 Key Features

• Multi-Magnification Pyramid: Jointly models 5×, 10×, and 20× patches in a graph structure.
• Lightweight & Interpretable: Just 0.378M parameters, with visualizable magnification-aware heatmap.
• Convergence Pooling: Replaces traditional pooling layers with a convergence mechanism—proven both empirically and theoretically.
• Clinically Validated: Behavior and interpretability confirmed by expert pathologists.

---

🖼️ Overview

Each WSI is encoded as a hierarchical graph:

• Nodes: Patch embeddings from different magnifications.
• Edges:
  • Intra-magnification (fully connected per mag).
  • Inter-magnification (links across magnification pyramid).

---

🧪 Setup

Directory layout

• assets/raw_features/ – Example feature files per slide (HDF5 format).
• assets/graphs/ – Generated DGL graphs will be placed here.
• assets/model_outputs/ – Model checkpoints and outputs are written here.
• assets/files/ – Supporting files including the manifest and data splits.
• bash_scripts/ – Bash scripts that illustrate the full pipeline.
• codes/ – Python sources for training and evaluation.

Requirements

• Python 3.9+
• PyTorch and DGL (GPU enabled is recommended)
• Use Conda to create the grasp environment defined in assets/grasp.yml

Run the following to create and activate the environment:

conda env create -f assets/grasp.yml
conda activate grasp

1. Graph construction

bash_scripts/run_1_graph_construction.sh builds graph files from the raw feature matrices. Features are expected under assets/raw_features/ and a manifest describing each slide is provided in assets/files/manifest.csv.

bash bash_scripts/run_1_graph_construction.sh

This expands to:

python app.py \
  --mags 5 10 20 \
  --feat_location ./assets/raw_features/ \
  --graph_location ./assets/graphs/ \
  --manifest_location ./assets/files/manifest.csv

Graphs are stored in assets/graphs/raw_features/*.bin.

2. Train GRASP

bash_scripts/run_2_submit_grasp.sh trains the GRASP model on the graphs. The script chooses an encoder (here KimiaNet) and sets the feature dimensionality accordingly. The command executed is similar to the following:

python3 ./codes/run_test_graph.py \
  --model_name GRASP \
  --split_name fold-1 \
  --batch_size 8 \
  --mags "5x 10x 20x" \
  --hidden_layers 256 128 \
  --lr 0.001 \
  --weight_decay 0.01 \
  --seed 256 \
  --epochs 5 \
  --feature_size 1024 \
  --classes CC:0 LGSC:1 \
  --spatial_gcn False \
  --conv_layer gcn \
  --path_to_folds ./assets/files/KimiaNet_data_folds_graph.json \
  --path_to_save ./assets/model_outputs/KimiaNet/gcn/

Modify the script to adjust magnifications, data splits, encoder name or other hyperparameters.

3. Report results

After training, run bash_scripts/run_3_result_reporter.sh to aggregate metrics across seeds and produce simple plots:

python3 ./codes/result_reporter.py \
  --batch_size 8 \
  --num_folds 1 \
  --lr 0.001 \
  --weight_decay 0.01 \
  --epochs 5 \
  --num_classes 2 \
  --models GRASP \
  --path_to_outputs ./assets/model_outputs/ \
  --encoder KimiaNet/gcn/ \
  --mags 5x 10x 20x

Metrics such as accuracy, balanced accuracy, F1, and AUC are printed to the console.

Notes

The provided data and scripts are for demonstration. Replace the sample feature files and manifest with your own dataset, following the same folder structure. Ensure the --classes argument in the training script matches your subtype labels.

📜 Citation

If you use GRASP in your work, please cite:

@inproceedings{
mirabadi2025grasp,
title={{GRASP}: {GRA}ph-Structured Pyramidal Whole Slide Image Representation},
author={Ali Khajegili Mirabadi and Graham AD Archibald and Amirali Darbandsari and Alberto Contreras-Sanz and Ramin Nakhli and Maryam Asadi and Allen W Zhang and Blake Gilks and Peter Colin Black and Gang Wang and Hossein Farahani and Ali Bashashati},
booktitle={Learning Meaningful Representations of Life (LMRL) Workshop at ICLR 2025},
year={2025},
url={https://openreview.net/forum?id=k6xlOfZnTC}
}