NoiseResilientHistopathology

Official implementation for the paper "Contrastive-Based Deep Embeddings for Noise-Resilient Histopathology Image Classification". (https://arxiv.org/abs/2404.07605)

Requirements

pip install -r requirements.txt

Please note: For embedding extractions, CTranspath and Pathoduet require two different versions of timm. Make sure you have the appropriate version for each of them.

If you want to use PathoDuet or RetCCL backbones, you need to import their repositories to the root.

How To Run

Datasets

Image Datasets

We use pythorch ImageFolder to load dataset so organize your datasets as:

dataset
└──train
    └──class1
        └──img1.png
        └──img2.png
        ...
    └──class2
    ...
└──val
└──test

val and test folders are optional and can be obtained by splitting train.

Embedding Datasets

If you have manually extracted embeddings, place them in ./deep_features/{dataset_name}. Respect nomenclature : {train/val/test}_df_{backbone_name}.npy and {train/val/test}_labels_{backbone_name}.npy.

Else, the code will extract embeddings using image dataset and place them in the right folder before training.

Configs for the experiment settings

Add '*.yaml' file in the config folder for each experiment.

Example: for CTransPath runs on NCT-CRC-HE-100k dataset, write config in configs/crc/ctranspath.yml

epochs: 300
patience: 20

dataset:
  name: 'crc'
  type: 'deep_features'
  backbone: 'ctranspath'
  batch_size: 512
  sigma: 0.4
  num_workers: 2
  val_size: 0.2
  
optimizer:
  name: SGD
  lr: 0.025
  weight_decay: 1.e-4
  momentum: 0.9

scheduler:
  name: CosineAnnealingLR
  T_max: $epochs

criterion:
  name: CrossEntropyLoss

warmup:
    active: False

Scripts

Arguments:

• noise_rate: noise rate
• config_path: path to the configs folder
• version: the config file name
• exp_name: name of the experiments (as note)
• seed: random seed
• resize: resize size (for image-based method)

Example for CTransPath runs on all noise rates for NCT-CRC-HE-100k dataset

#!/bin/bash
python_script="main.py"
seeds=(42 123 1999 78)
runs=("seed=42" "seed=123" "seed=1999" "seed=78")
noise_rates=("0." "0.1" "0.2" "0.3" "0.4" "0.5" "0.6" "0.7" "0.8" "0.9")

for noise_rate in "${noise_rates[@]}"; do
    for i in $(seq 4 $END); do 
        echo "Running $python_script with noise_rate=$noise_rate and seed=${seeds[$i-1]}"
        python3 $python_script --config_path configs/crc/ --version ctranspath --exp_name ${runs[$i-1]} --noise_rate $noise_rate --seed ${seeds[$i-1]}
    done
done

Results and Logs

After running, results and logs should be in experiments/{dataset_name}/{version_name}/{noise_rate}/{exp_name}

Citing this work

If you use this code in your work, please cite the accompanying paper:

@misc{dedieu2024contrastivebased,
      title={Contrastive-Based Deep Embeddings for Label Noise-Resilient Histopathology Image Classification}, 
      author={Lucas Dedieu and Nicolas Nerrienet and Adrien Nivaggioli and Clara Simmat and Marceau Clavel and Arnaud Gauthier and Stéphane Sockeel and Rémy Peyret},
      year={2024},
      eprint={2404.07605},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}