[ISBI 2025] Rethinking Pre-trained Feature Extractor Selection in Multiple Instance Learning for Whole Slide Image Classification
IEEE International Symposium on Biomedical Imaging (ISBI) 2025
Multiple Instance Learning (MIL) is widely used for gigapixel Whole Slide Image (WSI) classification without requiring patch-level annotations. Despite its popularity, there’s little consensus on what makes a good pre-trained feature extractor for MIL—most pipelines rely on default configurations without deeper evaluation.
To address this gap, our study systematically investigates what actually matters in choosing a feature extractor. We analyze three critical dimensions—pre-training dataset, backbone architecture, and self-supervised learning method—and uncover several insights that challenge common assumptions.
- The choice of self-supervised learning (SSL) method has a greater impact than in-domain dataset selection
- Transformer-based backbones with deeper architectures outperform CNNs in generalization
- Larger and more diverse pre-training datasets significantly enhance feature quality and downstream performance
- Linux Ubuntu 20.04.6
- 2 NVIDIA A100 GPUs (40GB each)
- CUDA version: 12.2
- Python version: 3.9.18
Install Anaconda
Create a new environment and activate it
conda create --name mil_feat_ext python=3.9.18
conda activate mil_feat_extInstall all required packages
pip install -r requirements.txt
pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 --extra-index-url https://download.pytorch.org/whl/cu117
pip install timm-0.5.4.tarA. TCGA-NSCLC (Lung Cancer) → Provided by DSMIL, it can be download directly in both 2.5x and 10x magnifications from the following link
B. Camelyon16 → The HS2P project, build on CLAM is utilized for standardized tissue patch extraction. Following default configuration file config/extraction/default.yaml from the HS2P, below are the most important configurations to note:
- spacing: 0.5 µm/px (~20x magnification)
- tissue_thresh: 0.1
- patch_size: 256
- save_patches_to_disk: True
We follow the official train/test split on the Camelyon16 dataset and take the first 836 entries in datasets_csv/tcga-nsclc.csv for training and the remaining 210 entries for testing
Below is the example of the TCGA-NSCLC dataset structure after splitting into classes and train/test:
Folder structure
<datasets>/
├── data_tcga_lung_tree/
├── train/
├── LUAD/
├── TCGA-4B-A93V-01Z-00-DX1/
├── x1_y1
├── x11_y11.png
├── x12_y12.png
├── x2_y2.png
├── x3_y3.png
├── ...
├── ...
├── LUSC/
├── TCGA-18-3408-01Z-00-DX1/
├── x4_y4
├── x41_y41.png
├── x42_y42.png
├── x5_y5.png
├── x6_y6.png
├── ...
├── ...
├── test/
├── ...
Create a config file under feature_extraction/configs/ with the name as follows: [name of backbone]-[name of pre-training method]-[name of pre-training dataset].yaml. Below are the examples of config file names:
- resnet50-supervised-imagenet1k (Timm)
- swin_base-supervised-imagenet21k (Timm)
- vit_small16-ssl-dino-imagenet1k (PyTorch Hub)
- resnet50-lunit-ssl-barlow_twins-tcga-tulip (Benchmark SSL Pathology)
A good starting point is to follow the config file below (e.g., resnet50-supervised-imagenet1k.yaml), making changes only to the sections marked with '#CHANGE' comments (see feature_extraction/configs/ for more details)
Config File
gpu_id: 0 # CHANGE
# Output settings
output_dir_path: 'outputs'
feature_extractor: 'resnet50-supervised-imagenet1k-transform' # CHANGE
resume: False # CHANGE
# Dataset settings
dataset:
name: 'tcga-nsclc' # CHANGE: ['tcga-nsclc', 'camelyon16']
patch_size: 224 # CHANGE
base_folder_path: '../../feature_extractor_MIL_study/datasets/data_tcga_lung_tree' # CHANGE
slide_data_path: 'slide_data/${dataset.name}.csv'
extracted_summary_path: '../datasets_csv/${dataset.name}/${feature_extractor}.csv'
slide_missing_path: 'slide_missing/${dataset.name}.csv'
subsets:
- train
- test
classes: # CHANGE
- LUAD
- LUSC
# Model settings
model:
backbone: 'resnet50' # CHANGE
feats_dim: 1024 # CHANGE
kernel_size:
trained_path:
# Feature Extraction settings
feature_extraction:
save_patch_features: False # CHANGE
normalization: 'imagenet' # CHANGE: ['imagenet', 'lunit']
stain_norm_macenko: False
To kick off the feature extraction process using resnet50-supervised-imagenet1k features
cd feature_extraction/
python extract_features.py --config-name resnet50-supervised-imagenet1kThis will result in the slide features being stored under feature_extraction/outputs/tcga-nsclc/resnet50-supervised-imagenet1k/
Below are the useful arguments to train MIL models
seed # Seed to reproduce experiments
device # Which GPU for training
num_classes # Depends on the number of positive classes (e.g., camelyon16: 1, tcga-nsclc: 2)
feature_extractor # Name of the feature extractor used
feats_size # Dimension of the feature vector
model # Which MIL model to use ['abmil', 'dsmil', 'transmil', 'DTFD']
distill # Only Used for DTFD-MIL ['MaxMinS', 'MaxS', 'AFS']To train all four SOTA MIL models with 3 times running (e.g., seed 1, 17, 2000) using the previously extracted features (e.g., resnet50-supervised-pretrained-imagenet-1k) is as below:
cd ..
bash train.shOur code is mainly built from these amazing works IBMIL, ABMIL, DSMIL, TransMIL, DTFD-MIL, HS2P, Re-implementation HIPT
If you find our work useful in your research or if you use parts of this code, please consider citing our paper
@inproceedings{wong2025rethinking,
title={Rethinking pre-trained feature extractor selection in multiple instance learning for whole slide image classification},
author={Wong, Bryan and Hong, Sungrae and Yi, Mun Yong},
booktitle={2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI)},
pages={1--5},
year={2025},
organization={IEEE}
}