RaPaCL

RadiomicsFeature-Pathomics Contrastive Learning.

RaPaCL (RadiomicsFeature-Pathomics Contrastive Learning) is a multimodal representation learning framework designed to bridge handcrafted radiomics features and deep pathomics features derived from histopathology images. In this approach, radiomics features extracted from image patches serve as structured, interpretable signals, while deep learning based patch image encoders encode high-dimensional visual representations. RaPaCL leverages contrastive learning to align these two modalities in a shared latent space, encouraging consistency between radiomics-informed characteristics (e.g., texture, heterogeneity) and deep image embeddings. By doing so, the framework aims to enhance the biological relevance and interpretability of learned representations, ultimately improving downstream tasks such as spatial gene expression prediction and tumor characterization in whole-slide images.

---

Prepare data

(i) Hugging Face 토큰 환경변수로 설정

touch .env
HF_TOKEN=hf_xxxxxxxxxxxxxxxxxxxxx
echo ".env" >> .gitignore

(ii) HEST / HEST-Bench 다운로드

python -m src.data.download_hest --config configs/data/download_hest.yaml

(iii) gene list 추출

python -m src.data.extract_genes --config configs/data/extract_genes.yaml

---

RaPaCL

Run RaPaCL

...

Pretrain Radiomics TransTab

cd RaPaCL/

(i) prepare tabular custom data

• from H5 (HEST-style)

  python -m src.pretrain_transtab.prepare_tabular \
    --config configs/data/prepare_tabular.yaml 

(ii) pretrain

• mode explanation:

  • train: contrastive pretraining
  • eval: classifier finetuning + test classification metric (accuracy / macro F1 / AUROC)
  • eval_detailed: representation quality 평가 전용 (표현 공간 자체 평가)
    • embedding 추출
    • clustering metrics: Silhouette / NMI / ARI
    • UMAP / t-SNE visualization
    • label cluster visualization
    • radiomics feature representative points
    • feature quantile section visualization

• pretrain with single GPU

  python -m src.pretrain_transtab.pretrain_transtab \
    --config configs/pretrain_transtab/idc_allxenium.yaml \
    --distributed false \
    --mode train

• pretrain with multi GPU

  CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node=2 \
    -m src.pretrain_transtab.pretrain_transtab \
    --config configs/pretrain_transtab/idc_allxenium.yaml \
    --mode train

• simple evaluation of pretraining

  python -m src.pretrain_transtab.pretrain_transtab \
    --config configs/pretrain_transtab/idc_allxenium.yaml \
    --distributed false \
    --mode eval

• detailed evaluation of pretraining

  python -m src.pretrain_transtab.pretrain_transtab \
    --config configs/pretrain_transtab/idc_allxenium.yaml \
    --distributed false \
    --mode eval_detailed

---

Baselines

Run stnet

cd RaPaCL/

(i) train

python -m src.baselines.stnet.run \
  --config configs/stnet.yaml \
  --mode train

(ii) eval

python -m src.baselines.stnet.run \
  --config configs/stnet.yaml \
  --mode eval

(iii) tuning

python -m src.baselines.stnet.run \
  --config configs/stnet.yaml \
  --mode tuning

Run Img2Rad

cd RaPaCL/

(i) run both train & eval

python -m src.baselines.img2rad.main \
  --config configs/img2rad.yaml \
  --mode all

(ii) run train or eval individually

python -m src.baselines.img2rad.main \
  --config configs/img2rad.yaml \
  --mode train

python -m src.baselines.img2rad.main \
  --config configs/img2rad.yaml \
  --mode eval

(iii) inspect parquet

python -m src.baselines.img2rad.inspect \
  --config configs/img2rad.yaml \
  --mode parquet \
  --show_columns

(iv) run ablation studies

• see scripts/run_img2rad_*.sh.

---