OpenMIL

This repo provides:

• Open available slide-level models for whole-slide image (WSI) analysis, including classic MIL architectures and pre-trained slide encoders that operate on patch features.

• Open available dataset .json lists for WSI analysis, including classification and survival prediction datasets. You can find the lists in downstream_task_jsons/ and surv_downstream_task_jsons/.

• Easy-to-use benchmark scripts for downstream tasks (classification and survival prediction).

Models

MIL models

• abmil
• gated_abmil
• transmil
• amdmil
• clam_sb
• clam_mb
• dsmil
• 2dmamba
• retmil
• wikg
• ilramil

Pre-trained slide encoders

• titan
• prism
• gigapath
• madeleine
• chief
• care
• feather_uni_v1
• feather_uni_v2
• feather_conch_v1_5
• tangle_v2
• ticon

Usage

Run the demo to verify all models:

python demo_test_model.py

Create a model in Python:

from slide_encoder_models.model_registry import create_slide_encoder, list_models

print(list_models())
model = create_slide_encoder("abmil", dim_in=1024, dim_hidden=512, num_classes=2)

Benchmark (MIL downstream tasks)

Task definitions

• Classification tasks: downstream_task_jsons/
• Survival tasks: surv_downstream_task_jsons/

Run all benchmarks

Edit run_benchmark_all.sh first:

• DATA root: set DATA_ROOT="YOUR_DATA_ROOT" to your dataset root.
• Models: set PFM_NAMES=(...), SLIDE_NAME=..., DTYPE=...
• Paths (usually no change): JSON_DIR=./downstream_task_jsons, SURV_JSON_DIR=./surv_downstream_task_jsons, JOB_DIR=./results

Then run:

bash run_benchmark_all.sh

What gets executed

• Single-split classification: per-dataset, 5 seeds (run_five_seeds).
• 10-fold CV classification: cptac_* fold tasks, 10 folds × 5 seeds per fold.
• Survival prediction (OS): per-cohort folds × 5 seeds (run_five_seeds_surv).

Outputs

Metrics are written under results/ following the script arguments (dataset / pfm_name / slide_name / seed / benchmark), e.g.:

• results/<dataset>/<pfm_name>/<slide_name>/<seed>/benchmark/all_test_metrics.json
• For fold tasks: results/<dataset>/<task>/fold<k>/<pfm_name>/<slide_name>/<seed>/benchmark/all_test_metrics.json

---

Benchmark (pre-trained slide encoders)

This repo also provides lightweight benchmarks for pre-trained slide encoders (WSI encoders) that take patch features as input.

KNN benchmark

Script: knn_benchmark.py + run_knn_benchmark.sh

• Input: *.pth patch-feature files referenced by downstream_task_jsons/*.json (with <PFM_NAME> placeholder in the path).
• Model: --slide-encoder (pre-trained slide encoders only).
• Output: knn_results/<pfm_name>/<dataset>/<slide_encoder>/k{k}_{metric}/

Run:

# 1) Edit DATA_ROOT in the script
bash run_knn_benchmark.sh

Linear probe benchmark

Script: linear_probe_benchmark.py + run_linear_benchmark.sh

• Protocol: StratifiedKFold on train split; early-stop on val balanced-acc; evaluate on held-out test.
• Output: linear_probe_results/<pfm_name>/<dataset>/<slide_encoder>/lp_f{num_folds}_pat{patience}/

Run:

# 1) Edit DATA_ROOT in the script
bash run_linear_benchmark.sh

Patch encoder is fixed for each slide encoder

For these pre-trained slide encoders, the required patch encoder is fixed (do not mix), and the benchmark scripts enforce this mapping:

• titan: conch_v1_5
• care: conch_v1_5
• madeleine: conch_v1
• chief: ctranspath
• prism: virchow_1
• gigapath: prov_gigapath
• feather_uni_v1: uni_v1
• feather_uni_v2: uni_v2
• feather_conch_v1_5: conch_v1_5
• tangle_v2: uni_v1
• ticon: conchv15 by default (tile_encoder_key also supports hoptimus1, uni2h, gigapath, virchow2)

Jiawen Li, jw-li24@mails.tsinghua.edu.cn