OceanPath

OceanPath is a modernized MIL pathology pipeline built from OceanPath_v1 concepts.

What was migrated from OceanPath_v1

• WSI model definition (ABMIL, TransMIL, static pooling MIL)
• Lightning WSI training module (fine-tune + linear probe)
• k-fold cross-validation training/evaluation pipeline

New workflow features

• Visualization during evaluation (ROC/PR + probability histograms)
• Model export (TorchScript + ONNX)
• Statistical analysis (paired permutation test for AUC)

Expected training CSV format

Use a CSV containing at least:

• feature_path: path to .pt or .npy bag feature file
• label: integer class label
• k_fold: fold id (0..K-1 for train/val folds, -1 for held-out test)

Run k-fold training

python scripts/train.py \
	--csv /path/to/folds.csv \
	--output-dir outputs/cv_abmil \
	--mil ABMIL \
	--feature-dim 1024 \
	--n-classes 2

This writes:

• per-fold checkpoints in outputs/cv_abmil/fold_*/
• cv_results.json
• ensemble_test_predictions.csv (if test rows with k_fold=-1 exist)

Evaluate + visualize

python scripts/evaluate.py \
	--pred-csv outputs/cv_abmil/ensemble_test_predictions.csv \
	--output-dir outputs/cv_abmil/eval

Outputs:

• metrics.json
• roc_pr_curves.png
• probability_histogram.png

Export model

python scripts/export_model.py \
	--checkpoint outputs/cv_abmil/fold_0/best.ckpt \
	--output-dir outputs/cv_abmil/export \
	--feature-dim 1024

Outputs:

• wsi_model.ts
• wsi_model.onnx

Compare two models (analysis)

python scripts/analyze.py \
	--pred-a outputs/model_a_preds.csv \
	--pred-b outputs/model_b_preds.csv \
	--output outputs/significance.json