README.md

Medical Image Semantic Segmentation

Official pytorch implementation of the paper Histogram of Oriented Gradients Meet Deep Learning: A Novel Multi-task Deep Network for Medical Image Semantic Segmentation
This work presents a novel deep multi-task learning method for medical image segmentation leveraging Histogram of Oriented Gradients (HOGs) to generate pseudo-labels.

Prerequisites

• Python 3.8
• PyTorch 1.8
• CUDA 10.1

Installation

First of all, clone the repo:

    git clone https://github.com/thetna/medical_image_segmentation.git

All the required python packages can be installed with:

    cd medical_image_segmentation
    pip install -r requirements.txt

Training

For training, first put training images and corresponding segmentation maps in separate directories and update the following information in the Config/train.yml file.

config/train.yml
├──  network    - Segmentation achitectures: UNet, UNet+HOG, U2Net, U2Net+HOG
│
├──  dataset    - Specify the datasets to use for training: Cadis, Robinst, Custom
│ 
├──  task       - Task number for Cadis or Robinst datasets: 1, 2, 3
│ 
├──  datasets 
│    └── train  - Training data configurations
│    └── valid  - Validation data configurations
│ 
├──  seg_net 
│    └── in_nc  - Input number of channels
│    └── out_nc - Output number of channels or total number of classes 
│    └── resume_path  - Path for checkpoint to resume from
│ 
├──  hog_decoder 
│    └── out_dim  - Dimension of HOG to use
│    └── resume_path  - Path for checkpoint to resume from
│
├──  train    - All other training parameters

Then start training with the following command:

    python train.py config/train.yml

Inference

Download the pre-trained weights from here. Add the path of images and the models in the config/test.yml. Then run the following command:

    python test.py config/test.yml

Results

Qualitative comparison between the proposed method with its counter-part architecture U2Net on three different tasks. First two rows represent examples from Task 1, the middle two rows, and the last two rows are examples from Task 2 and Task 3 respectively.

Qualitative comparison between before and after applying our method on U2Net in the Task 2 of robotic instrument segmentation challenge held in MICCAI 2017.

Citation

  @article{bhattarai2023histogram,
  title={Histogram of Oriented Gradients meet deep learning: A novel multi-task deep network for 2D surgical image semantic segmentation},
  author={Bhattarai, Binod and Subedi, Ronast and Gaire, Rebati Raman and Vazquez, Eduard and Stoyanov, Danail},
  journal={Medical Image Analysis},
  pages={102747},
  year={2023},
  publisher={Elsevier}
}