Histopathology OOD Classification

Setup (Kaggle)

!git clone https://github.com/TristanLecourtois/-histopathology_OOD_classification.git /kaggle/working/project
!pip install -r /kaggle/working/project/requirements.txt -q
!pip install git+https://github.com/sebastianffx/stainlib.git -q

from kaggle_secrets import UserSecretsClient
import os
os.environ['HF_TOKEN'] = UserSecretsClient().get_secret('HF_TOKEN')

---

Pipeline

Step 1 — Feature extraction Loads the foundation model and runs it on train, val and test images to extract embeddings. Applies optional Reinhard stain normalization and HED augmentation. Results are saved as .pt files in /kaggle/working/{model}/. If interrupted, relaunch the same command — already computed files are skipped.

!HF_TOKEN=$HF_TOKEN python /kaggle/working/project/scripts/extract_features.py --model MODEL [options]

Step 2 — Training Trains a linear classifier on top of the precomputed embeddings. Saves the best model weights based on validation accuracy.

!python /kaggle/working/project/scripts/train.py --model MODEL

Step 3 — Prediction Loads the trained classifier, runs it on all test embeddings (with TTA averaging), and saves the final predictions to /kaggle/working/submission_{MODEL}.csv.

!python /kaggle/working/project/scripts/predict.py --model MODEL

---

Models

For genbio, upload model.pth as a Kaggle dataset and set GENBIO_WEIGHTS_PATH in config.py.

---

Options — extract_features.py

Options — train.py / dann_train.py

---

DANN — Domain Adversarial Training

What it is: DANN forces the feature extractor to produce embeddings that are indistinguishable across hospitals (centers). It does this by adding a domain classifier that tries to predict which center an image comes from, and a gradient reversal layer (GRL) that flips the gradients from this classifier. The label classifier is thus trained on features that are both discriminative for tumor/non-tumor AND blind to the center.

How it works:

• The label classifier is trained normally (minimize label loss)
• The domain classifier tries to predict the center (0, 3 or 4 in train)
• The GRL reverses its gradients → the shared features are pushed to be domain-invariant
• Alpha (λ) starts at 0 and increases during training so domain adaptation kicks in gradually

Uses the same extracted features — no need to re-run extract_features.py.

Commands:

!python /kaggle/working/project/scripts/dann_train.py --model MODEL
!python /kaggle/working/project/scripts/dann_predict.py --model MODEL

Submission saved to /kaggle/working/submission_{MODEL}_dann.csv.

---

MixStyle

MixStyle mixes instance-level feature statistics (mean and std) between pairs of training samples to simulate cross-domain style variation. Enable it with --mixstyle on train.py or dann_train.py:

!python /kaggle/working/project/scripts/train.py --model MODEL --mixstyle
!python /kaggle/working/project/scripts/dann_train.py --model MODEL --mixstyle

---

Visualization

RGB distribution across centers:

!python /kaggle/working/project/plots/plot_rgb_distribution.py

Saves rgb_distribution.pdf to /kaggle/working/.

Stain normalization (before / after Reinhard):

!python /kaggle/working/project/plots/plot_stain_normalization.py

Saves stain_normalization.pdf to /kaggle/working/.