Auto-Generate-WLs

Code repository supporting the paper "Auto-Generating Weak Labels for Real & Synthetic Data to Improve Label-Scarce Medical Image Segmentation," which was accepted at MIDL 2024. For reproducibility, we have included all datasets + generated weak labels used in the paper.

Paper

Using the code:

1. Prepare codebase and data

• Clone this repository:

git clone https://github.com/stanfordmlgroup/Auto-Generate-WLs
cd Auto-Generate-WLs

• Create a conda environment:

  conda env create -f environment.yml
  conda activate auto_wl

• Place the gold-standard (GS) dataset and unlabeled dataset under pytorch-nested-unet/inputs in the following format:

  inputs
  └── <dataset name>
  ├── images
  |   ├── 001.png
  │   ├── 002.png
  │   ├── 003.png
  │   ├── ...
  └── masks
      ├── 0
      |   ├── 001.png
      |   ├── 002.png
      |   ├── 003.png
      |   ├── ...
      └── 

  We also provide the gold-standard dataset and weak labels used in this paper for reproducibility.

Datasets:

• BUSI
• ISIC 2016 Part 3B
• CXR-COVID

2. Train a model on the gold-standard dataset

Next, train an initial model on the gold-standard dataset. This will be used to generate prompts for MedSAM and generate the weak labels:

cd pytorch-nested-unet
python train.py --dataset <gs dataset> --arch NestedUNet 
    --img_ext .png --mask_ext .png --batch_size 4 
    --input_w 256 --input_h 256

For example:

python train.py --dataset busi-25-small --arch NestedUNet 
        --img_ext .png --mask_ext .png --batch_size 4 
        --input_w 256 --input_h 256

Then, generate predictions on the unlabeled dataset using this model:

python eval.py --name <gs dataset>_NestedUNet_256_woDS 
    --dataset <unlabeled dataset>

For example:

python eval.py --name busi-25-small_NestedUNet_256_woDS 
        --dataset busi-25-aug-25

This will output the predictions to a outputs/<model name>/<dataset name>/0 (the prediction path).

3. Generate predictions on the unlabeled dataset

Download the MedSAM bounding-box checkpoint and/or the MedSAM point-prompt checkpoint here and place them in the folder generate-weak-labels/MedSAM/work_dir/MedSAM/.Run the notebook generate-weak-labels/generate-masks.ipynb, which will generate the weak labels and provide a visualization of the coarse labels, prompts, and weak labels.

4. Train a model on the augmented dataset

Now, we are ready to train a model on the augmented dataset, as follows:

python train.py --dataset <augmented dataset> --arch NestedUNet 
    --img_ext .png --mask_ext .png --batch_size 4 
    --input_w 256 --input_h 256

For example:

python train.py --dataset busi-box-aug-50 --arch NestedUNet 
        --img_ext .png --mask_ext .png --batch_size 4 
        --input_w 256 --input_h 256

Then, we can evaluate both our base model trained on the gold-standard dataset as well as the model trained on the augmented dataset, as follows:

python eval.py --name <gs dataset>_NestedUNet_256_woDS 
    --dataset <test dataset>
python eval.py --name <augmented dataset>_NestedUNet_256_woDS
    -- dataset <test dataset>

and compare the resulting DICE and IOU. You can also compare the predictions in the respective output folders of the models.

Acknowledgements

This codebase uses code segments from UNet++ and MedSAM.