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

EXAONE Path 2.0 rev EGFR

Introduction

EXAONE Path 2.0 rev EGFR is an EGFR predictor.
It takes a Whole Slide Image (WSI) as input and runs a three-stage pipeline to predict EGFR mutation status:

1. Patchify (§4.1-1): Segment tissue regions from the WSI and extract patch coordinates.
2. Patch Feature Extraction (§4.1-2): Encode each patch into a feature vector using the pretrained patch encoder, producing per-slide HDF5 feature files.
3. Slide Feature Extraction (§4.2): Aggregate patch-level features into a single slide-level representation using the slide encoder.
4. EGFR Classification (§4.3): 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 specifically uses CUDA-enabled PyTorch, making NVIDIA GPU mandatory for running the model.

2. Environment Setup

First, install Micromamba if you haven't already. You can find installation instructions here . Then create and activate the environment using the provided configuration:

micromamba create -n exaonepath python=3.12
micromamba activate exaonepath
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. Unified Hub Loading (Standard AutoModel Weights)

The Hub repo stores standard AutoModel weight files ( pytorch_model.bin or safetensors) with one AutoModel entrypoint.

from transformers import AutoModel

repo_id = "LGAI-EXAONE/EXAONE-Path-2.0-rev-EGFR"

bundle = AutoModel.from_pretrained(
    repo_id,
    trust_remote_code=True,
)
patch_encoder, slide_encoder, classifier = bundle
# Equivalent:
# patch_encoder = bundle.patch_encoder
# slide_encoder = bundle.slide_encoder
# classifier = bundle.classifier

4. Inference Workflow Overview

EXAONE Path 2.0 rev EGFR inference follows a three-stage pipeline. (1) Patch-level feature extraction : extract pretrained patch embeddings from full WSIs into HDF5 files. (2) Slide-level feature extraction : aggregate patch embeddings into slide representations. (3) EGFR classification : run the classifier head using slide/patch embeddings.

4.1. Patch Feature Extraction

This section focuses on full-WSI processing and HDF5 output for downstream slide encoding.

Full WSI patch-feature pipeline (HDF5 Output)

The step is further broken into smaller steps.

(1) Generate patch coordinates (and contour indices) with Python function API patchfy_batch to process every WSI in a folder:

• segment tissue regions
• extract patch coordinates
• (optionally) write a HDF5 file with coords + contour_index

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=out_dir,
    num_workers=8,
    csv_path=f"{out_dir}/patchfy_batch_status.csv",
    patch_size=256,
    step_size=256,
    patch_level=0,
    save_h5=True,
    save_mask=True,
    auto_skip=True,
)

Outputs

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

(2) Extract patch features for all successful WSIs from the Step (1) CSV:

python -m exaonepath.feature_extraction.extract_batch_patch_feature \
    --status_csv_path "<out_dir>/patchfy_batch_status.csv" \
    --coords_root_dir "<out_dir>/patches" \
    --feature_out_dir "<feature_out_dir>" \
    --batch_size_per_gpu 1024 \
    --gpu_ids "0,1" \
    --num_workers_per_gpu -1

Notes

• --status_csv_path : Step (1) patchfy_batch output status CSV.

• --coords_root_dir : directory containing coordinate HDF5 files ( <slide_id>.h5 ) from Step (1).

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

• --batch_size_per_gpu : batch size per GPU.

• --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 <feature_out_dir>/logs/ .

4.2. Slide Feature Extraction

Patch features, coordinates, (contour index) must be available.

Batch extraction (recommended)

python -m exaonepath.feature_extraction.extract_batch_slide_feature \
    --status_csv_path "<feature_out_dir>/logs/extract_patch_feature_batch_status.csv" \
    --patch_feature_root_dir "<feature_out_dir>" \
    --slide_feature_out_dir "<slide_feature_out_dir>" \
    --batch_size_per_gpu 32 \
    --gpu_ids "0,1" \
    --num_workers_per_gpu -1

Notes

• --status_csv_path : patch-feature extraction status CSV (Step 4.1-(2) output).

• --patch_feature_root_dir : directory containing patch feature files from Step 4.1-(2).

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

• --batch_size_per_gpu : batch size per GPU.

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

• --num_workers_per_gpu : DataLoader workers per GPU. Default: 12 (auto when <= 0 ).

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

4.3. EGFR Classification

Batch classification (recommended)

python -m exaonepath.feature_extraction.classify_batch_egfr \
    --status_csv_path "<slide_feature_out_dir>/logs/extract_slide_feature_batch_status.csv" \
    --prediction_csv_path "<slide_feature_out_dir>/logs/egfr_classification_results.csv" \
    --batch_size_per_gpu 8 \
    --gpu_ids "0,1" \
    --num_workers_per_gpu 8

Notes

• --status_csv_path : slide-feature extraction status CSV (Step 4.2 output).

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

• --batch_size_per_gpu : batch size per GPU.

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

• --num_workers_per_gpu : DataLoader workers per GPU. Default: 8 .

• Prediction CSV columns: wsi_name , score .

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

• Optional debug status CSV is written only if --status_log_csv_path is set.

License

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

Contact

LG AI Research Technical Support: [email protected]