yangsenwxy/Dataset_Merger

Merging nucleus datasets by correlation-based cross training

A framework formulating the nucleus classification as a unified multi-label classification problem with missing labels. For the nucleus from one specific dataset, besides its own label, all the other unknown labels are regarded as missing labels. Correlation between the known labels and unknown labels helps improve the performance. The label completion module in this framework propagates the supervision across labels, and we generate pseudo labels to mutually supervise the base module and label correlation learning module, respectively.

This repository can be used for training and inferring on H&E images. As part of this repository, we supply model weights trained:

• checkpoint

Webpage

More details（e.g. data, visual results) can be found at our project webpage

Set Up Environment

conda install pytorch torchvision cudatoolkit=10.2
conda install tensorboard
conda install scikit-learn
conda install opencv

Above, we install PyTorch with CUDA 10.2, tensorboard for monitoring loss, and scikit-learn for f1 score calculation

Repository Structure

Below are the main directories in the repository:

• dataset/: the dataset loader and preprocessor
• model/: the label completion modules
• training/: scripts for training different modules
• inference/: script for inference

Running the Code

Training

Download Data:

• panNuke: https://jgamper.github.io/PanNukeDataset/
• CoNSeP: https://warwick.ac.uk/fac/cross_fac/tia/data/hovernet/
• NuCLS: https://sites.google.com/view/nucls/home
• MoNuSAC: https://monusac-2020.grand-challenge.org/Home/

Convert Data to the same format

  python dataset/preparation/panNuke.py
  python dataset/preparation/CoNSeP.py
  python dataset/preparation/NuCLS.py
  python dataset/preparation/MoNuSAC.py

Data Format

For training, the ground truth images, masks, and nucleus types are stored separately in png, npy, and npy files in the data folder.

Usage and Options

Usage:

  python training/modules/sparse_training.py [--name=<training_name>] [--root_dir=<rood_dir_of_dataset>]
  python training/modules/label_correlation.py [--name=<training_name>] [--root_dir=<rood_dir_of_dataset>]
  python training/modules/label_completion.py [--name=<training_name>] [--root_dir=<rood_dir_of_dataset>]

All three modules should use the same name, or you could just use the default name.

Inference

Data Format

Input:

• Standard images files and mask, including png, jpg for the image, and npy for the masks.

Output:

• A npy file of array Nx2, where N is the number of instances, and the first column is the instance index, the second column is the nucleus type.
• A png overlay of nuclear boundaries on top of original RGB image

Usage and Options

Usage:

  python inference/infer.py [--root_dir=<data_path>] [--img_name=<image_name>]

Overlaid results of five different sets of labels. The colour of the nuclear boundary denotes the type of nucleus.

Datasets

The re-organized merged dataset is too large to be uploaded as supplementary materials. We will release it soon.