HNM-WSI

• This repository contains the implementation for our work "Hard Negative Sample Mining for Whole Slide Image Classification", accepted by MICCAI 2024

Environment Setup

Follow these steps to setup the environment

1. Install Conda: Ensure that you have Conda installed on your system. If not, you can download it from Miniconda or Anaconda based on your preference.

2. Create Conda Environment: Navigate to the project's root directory in your terminal or command prompt. Then, use the following command to create a Conda environment from the environment.yml file:

   conda env create -f environment.yml

Dataset

• Camelyon16 whole slide image and TCGA mutation whole slide image were downloaded from Camelyon16 and TCGA Portal, respectively.

Camelyon16 dataset

• We adopt the same method as DSMIL for data preprocessing. For more details about the threshold for filtering the background patches, please refer to the original paper.

TCGA mutation dataset

• Same as the Its2CLR, we preprocessed the data using the DeepPath method.

Pseudo label generation before training

• Before training the model, you need to generate the initial patch-level prediction scores. I use the instance classifier of the DSMIL to generate the patch-level prediction scores.

Training and evaluation

• To train the model on your dataset, you can use the following command:

CUDA_VISIBLE_DEVICES=0 python run.py  \
--expname EXP_NAME \
--threshold 0.3 \
--gpuid 0 \
--augment_transform 0 \
--batch_size 512 \
--pretrain_weight  \ #SSL pretrain weight path
--model_save_root  \ 
--pseudo_label_path /checkpoints/ssl/simclr/checkpoint/ins_pseudo_label_train.p \ #Initial pseudo label path, from the MIL trained on feature extractor using SSL pretrain weight
--MIL_every_n_epochs 5 \
--epoch_to_extract_mil 199 \
--posi_batch_ratio 0.2 \
--ro 0.2 \
--ro_neg 0.2 \
--rT 0.8 \
--warmup 5 \
--init_MIL_training yes \
--negfilter yes \
--neg_ratio 0.2 \
--root_dir xxxx \ # /single
--labelroot xxxx \ # /single
--comet_api_key xxxxx \

Acknowledgements

• This code was developed based on the works of Its2CLR and DSMIL. We would like to express our gratitude to the authors of these projects for their valuable contributions.