LGAI-EXAONE/EXAONE-Path-2.0-rev-EGFR

EXAONE Path 2.0 rev EGFR

EGFR Viewer Demo

Introduction

EXAONE Path 2.0 rev EGFR predicts EGFR mutation status from Whole Slide Images (WSIs) using a four-stage pipeline:

1. Patchify (§3.1): Segment tissue regions from the WSI and extract patch coordinates.
2. Patch Feature Extraction (§3.2): Encode each patch into a feature vector using the pretrained patch encoder, producing per-slide HDF5 feature files.
3. Slide Feature Extraction (§3.3): Aggregate patch-level features into a single slide-level representation using the slide encoder.
4. EGFR Classification (§3.4): Feed the slide representation into the classifier head to output an EGFR mutation probability score.

Quickstart

Load EXAONE Path 2.0 rev EGFR and extract features.

1. Hardware Requirements

• NVIDIA GPU with 12GB+ VRAM
• NVIDIA driver version >= 525.60.13 required

Note: This implementation requires NVIDIA GPU and drivers. The provided environment setup uses CUDA-enabled PyTorch, so NVIDIA GPU is mandatory for running the model.

2. Environment Setup

Install Micromamba first if needed (installation guide), then create and activate the environment:

micromamba create -n exaone-path-egfr python=3.12
micromamba activate exaone-path-egfr
pip install -U "huggingface_hub[cli]"
hf download LGAI-EXAONE/EXAONE-Path-2.0-rev-EGFR --repo-type model --local-dir exaone-path-2.0-rev-egfr
cd exaone-path-2.0-rev-egfr
pip install -r requirements.txt

3. Inference Workflow Overview

Run the following stages in order.

3.1. Patchify

Generate patch coordinates with the Python API patchfy_batch:

• segment tissue regions
• extract patch coordinates
• save an HDF5 file with coords and contour_index (save_h5=True; required for Steps 3.2-3.4)

from exaonepath.patches import patchfy_batch

data_folder = "path/to/your/wsi_folder"
out_dir = "path/to/output_dir"

results = patchfy_batch(
    data_folder=data_folder,  # scans common WSI formats
    out=f"{out_dir}/patches",
    num_workers=8,  # number of CPU cores to use
    csv_path=f"{out_dir}/patches/logs/patchfy_batch_status.csv",
    patch_size=256,
    step_size=256,
    patch_level=0,
    save_h5=True,
    save_mask=True,
)

Outputs

• Patch coordinate HDF5: <out_dir>/patches/patches/<slide_id>.h5
• Status CSV: <out_dir>/patches/logs/patchfy_batch_status.csv
• (Optional) masks: <out_dir>/patches/masks/<slide_id>.jpg

3.2. Patch Feature Extraction

Extract patch features for all successful WSIs from the patchify status CSV:

python -m exaonepath.feature_extraction.extract_batch_patch_feature \
    --status_csv_path "<out_dir>/patches/logs/patchfy_batch_status.csv" \
    --coords_root_dir "<out_dir>/patches/patches" \
    --feature_out_dir "<out_dir>/patch_features" \
    --batch_size_per_gpu 1024 \
    --gpu_ids "" \
    --num_workers_per_gpu -1

Notes

• --status_csv_path: patchfy_batch output status CSV.

• --coords_root_dir: directory containing coordinate HDF5 files (<slide_id>.h5) from Step 3.1.

• --feature_out_dir: output directory root for patch feature HDF5 files and logs.

• --batch_size_per_gpu: batch size per GPU. Adjust based on available GPU memory.

• --gpu_ids: comma-separated GPU indices to use. Default: "" (all visible GPUs).

• --num_workers_per_gpu: DataLoader workers per GPU. Default: -1 (auto).

• Output status CSV is written under <out_dir>/patch_features/logs/.

3.3. Slide Feature Extraction

Patch features, coordinates, and contour index must be available.

Batch extraction (recommended)

python -m exaonepath.feature_extraction.extract_batch_slide_feature \
    --status_csv_path "<out_dir>/patch_features/logs/extract_patch_feature_batch_status.csv" \
    --patch_feature_root_dir "<out_dir>/patch_features" \
    --slide_feature_out_dir "<out_dir>/slide_features" \
    --batch_size_per_gpu 32 \
    --gpu_ids "" \
    --num_workers_per_gpu -1

Notes

• --status_csv_path: patch-feature extraction status CSV (Step 3.2 output).

• --patch_feature_root_dir: directory containing patch feature files from Step 3.2.

• --slide_feature_out_dir: output directory root for slide feature files.

• --batch_size_per_gpu: batch size per GPU. Adjust based on available GPU memory.

• --gpu_ids: comma-separated GPU indices to use. Default: "" (all visible GPUs).

• --num_workers_per_gpu: DataLoader workers per GPU. Default: -1 (auto).

• Output status CSV is always written to <out_dir>/slide_features/logs/extract_slide_feature_batch_status.csv.

3.4. EGFR Classification

Batch classification (recommended)

python -m exaonepath.feature_extraction.classify_batch_egfr \
    --slide_feature_status_csv_path "<out_dir>/slide_features/logs/extract_slide_feature_batch_status.csv" \
    --prediction_csv_path "<out_dir>/slide_features/logs/egfr_classification_results.csv" \
    --heatmap_out_dir "<out_dir>/heatmaps" \
    --gpu_ids ""

Batch classification with optional viewer artifacts

python -m exaonepath.feature_extraction.classify_batch_egfr \
    --slide_feature_status_csv_path "<out_dir>/slide_features/logs/extract_slide_feature_batch_status.csv" \
    --prediction_csv_path "<out_dir>/slide_features/logs/egfr_classification_results.csv" \
    --heatmap_out_dir "<out_dir>/heatmaps" \
    --viewer_out_dir "<out_dir>/viewer" \
    --gpu_ids ""

Notes

• --slide_feature_status_csv_path: slide-feature extraction status CSV (Step 3.3 output), expected at <out_dir>/slide_features/logs/extract_slide_feature_batch_status.csv.

• --prediction_csv_path: output CSV path for EGFR prediction results.

• --gpu_ids: comma-separated GPU indices to use. Default: "" (all visible GPUs).

• Prediction CSV columns: wsi_name, score.

• score is class-1 probability (probs[1]), i.e., EGFR mutation score.

Heatmap options

• --heatmap_out_dir: output directory for slide-level attention heatmaps. If omitted, defaults to a sibling heatmaps/ directory next to the slide-feature root.
• --heatmap_max_side: maximum width/height in pixels for the saved heatmap image (default: 2048).
• --heatmap_alpha: blend weight of the heatmap overlay over the slide thumbnail, 0=transparent, 1=fully opaque (default: 0.4).
• --heatmap_blur: apply Gaussian blur to smooth patch boundaries; applied twice, once on the score map and once on the final RGB image (default: True).
• --heatmap_blur_strength: blur strength multiplier relative to the auto-derived sigma. 1.0=default strength, <1.0=weaker, >1.0=stronger (default: 1.0).
• --heatmap_cmap: matplotlib colormap name for the attention heatmap, e.g. coolwarm, jet, viridis (default: coolwarm).
• --top_patches_n: also save a grid of the highest-attention patches alongside the heatmap (0 to disable).

Viewer artifact options

• --viewer_out_dir: required to generate the viewer artifacts used by the zoomable EGFR viewer.

3.5. EGFR Viewer

Before running the viewer, generate viewer artifacts in Step 3.4 by setting --viewer_out_dir "<out_dir>/viewer".

Run the local viewer

python -m exaonepath.viewer.serve \
    --viewer_manifest_csv_path "<out_dir>/viewer/logs/egfr_viewer_manifest.csv" \
    --host "127.0.0.1" \
    --port 8000

Then open http://127.0.0.1:8000.

License

The model is licensed under EXAONEPath AI Model License Agreement 1.0 - NC

Contact

LG AI Research Technical Support: contact_us1@lgresearch.ai