ScaleDependantCNN

This repository contains all the necessary to reproduce the work "Scale dependant layer for self-supervised nuclei encoding" by Peter Naylor, Yao-Hung Hubert Tsai, Marick Laé and Makoto Yamada. You can find the paper here.

Requirements

You will need to have an installation of conda and of nextflow. For pytorch and torchvision, please be sure to take compatible versions.

Nextflow specifics

Please modify the nextflow.config to meet the specificities of your informatics setup. We used a SGE cluster provided by RIKEN AIP with CPU and GPU queues. Processing the PanNuke dataset requires more memory and computational power then the others, if you are limited in this regard, please remove this dataset from the benchmark.nf file.

Reproduce results in paper

Please run the commands in this order:

# setup the conda environnement 
make setup_conda

# download the data
make download_data

# run the main experiment
make experiment

# run the supplementary experiment of MoCo
make moco_experiment

# run the supplementary experiment to see the effect of different augmentations on the test set
make augmentation_experiment:

Hyper parameter-settings

To reproduce the main experiment in the paper please use the following hyper-parameters:

methods_selection = ["ascending", "descending"]
LAMBDA = [0.00078125, 0.0078125, 0.078125]
LR = [0.0001, 0.001, 0.01, 0.1]
WD = [0, 1e-6, 1e-4, 1e-2, 1.]
FEATURE_DIM = [32, 64, 128]
MB = [32768, 65536]
KS = [3, 5]
epochs = [ "tnbc": 100, "consep": 100, "pannuke": 50, "tnbcpadded": 100, "conseppadded": 100, "pannukepaddded": 50]
bs = [ "tnbc": [128], "consep": [128], "pannuke": [512], "tnbcpadded": [64], "conseppadded": [64], "pannukepadded": [64]]
number_bs = [0] // This parameter is to iterate over the batch size, please modify if you specify many batch size

models = ["ModelSDRN", "ModelSRN"]
opt = ["--inject_size", "--no_size"]
repetition = 20

Cite

Please cite our paper if you use our dataset or code.

@article{naylor2022scaledependant,
  title={Scale dependant layer for self-supervised nuclei encoding},
  author={Naylor, Peter and Tsai, Yao-Hung Hubert and Laé, Marick and Yamada, Makoto.},
  journal={arXiv preprint arXiv:},
  year={2022}
}