SP_guided_Noisy_Label_Seg

This repository contains a PyTorch implementation of the MICCAI 2021 paper "Superpixel-guided Iterative Learning from Noisy Labels for Medical Image Segmentation". [Video].

If you use our code in your publications, please consider citing our paper:

@inproceedings{li2021superpixel,
  title={Superpixel-Guided Iterative Learning from Noisy Labels for Medical Image Segmentation},
  author={Li, Shuailin and Gao, Zhitong and He, Xuming},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year={2021},
}

Get started

0. Environment

Install the necessary requirements on your environment of choice:

Key Requirements

• python==3.6.10

# install other necessary packages
pip install -r requirements.txt

1.Download Original Datasets

1. ISIC 2017 datasets: Download training data and test data at this link.
2. JSRT datasets: Download image and segmentation masks.

Note: we suggest you to try the code on ISIC dataset first, then apply on JSRT dataset due to its complicated preprocess procedure.

2. Simulate Noise Patterns

Material:

1. data_process/pipeline.py
2. ISIC & JSRT clean data
3. paramters: change params to your directory in main()

Function:

1. train/val split (here val set is used as test set and remain untouched)
2. noisy label generation
3. convert .png to .json, used as unified input labels.

# change to your parameters first (Params, roots, save_dir)
python pipeline.py

3. Generate Superpixels

Generate superpixels for input images.

# change to your parameters first (subdir, Params, ...)
python superpixel_generation.py

4. Noise ratio statistics

**We assume noise ratios are given following previous works, thus obtain them here.

# change to your parameters first (subdir, cls, Params, ...)
python stat_noise_ratio.py

Training

Please replace the <data_dir>, <em_save_pseudo_dir> and <sp_name> in the experiment config files.

<data_dir> is the root dir for the dataset. Make sure it looks like:

• data_dir
  • train
    • image
    • label
    • label_noise_xx_xx
    • <sp_name> # superpixel dir
  • val
    • ...

For the detailed information of directory settings, please refer to lib/datasets/batch_dataset.py.

After that, start training with following codes.

python main.py --cfg 'exp/proposed/skin_a3b5_spemj01.yaml' --id skin_a3b5_spemj01 --parallel 

Inference

python main.py --cfg 'exp/proposed/skin_a3b5_spemj01.yaml' --id skin_a3b5_spemj01 --demo 'val' \
--weight_path './checkpoints/skin_a3b5_spemj01/model_epoch_200.pth' --parallel --save_preds True

** Feel free to contact us if confused with above pipeline, and we will make "readme" more clear.

---

*. [Optional] Crop

**Clavicles are particularly small, to facilitate fine-grained segmentation and reduce consuming time, we crop their region of interest by statistics on the training set.

**After 2. Simulate Noise Patterns

# change to your parameters first (Params, roots, rois)
python crop_pipeline.py

Todo list:

• Upload JSRT Superpixels & Datasets packages.