Certified Medical Image Segmentation

This repository contains the implementation of our MICCAI 2023 submission "Certification of Deep Learning Models for Medical Image Segmentation".

Data and models

We evaluated our method on five medical image segmentation datasets. Three lung datasets comprising both single and multi-class segmentation, one dermatology dataset and one polyp dataset. Follow the instructions below to download the required data:

Skin Lesion

The dataset is obtained from the ISIC 2018 challenge and can be downloaded here.

Download the training and validation input and GT for Task 1 and extract the folders as follows:

datasets/
  lesion/
    ISIC2018_Task1-2_Validation_Input/
    ISIC2018_Task1-2_Training_Input/
    ISIC2018_Task1_Validation_GroundTruth/
    ISIC2018_Task1_Training_GroundTruth/

Then, navigate to datasets/lesion and run python make_dataset.py.

Polyp

The data is obtained from the CVC-ClinicDB challenge. Processed png images can be found here.

Download the dataset and add it as follows:

datasets/
  polyp/
    CVC-ClinicDB/
      Original/
        612.png
        ...
      Ground Truth/
        ...

Then run python datasets/polyp/split_dataset.py.

Lung

Download the three lung datasets by following the links provided below:

• JSRT - Japanese Society of Radiological Technology
• Montgomery
• Shenzhen

Place them in the project as follows:

datasets/
  jsrt/
    train/
      input/
        ...
      label/
        ...
    valid/
      input/
        ...
      label/
        ...
    test/
      input/
        ...
      label/
        ...
  shen/
    ...
  mount/
    ...

Follow the same hierarchy for folder mount and shen which will contain respectively Montgomery and Shenzhen datastets.

Denoising Diffusion Model

The Denoising Diffusion Probabilistic Models used in the paper is based on openai/guided-diffusion. Download the class unconditional pretrained model here and place it in the models directory.

Usage

Setup

First start by installing the requirements of the segmentation by using the environment.yml. If you don't want to use conda, the main requirements for the project are:

• PyTorch 1.7.1
• PyTorch ignite 0.4.3
• segmentation_models_pytorch 0.1.3
• Albumentations 0.5.2
• OpenCV 4.5.1.48

Usage for multi-class segmentation

Launch certification for skin lesions dataset with a Deeplab model

python -u test_certify_jsrt.py --weights weights/jsrt_deeplab/best_model.pt --model deeplab --dataset jsrt --multi --sigma 0.25  --denoise

Launch certification for skin lesions dataset with a Deeplab model and without denoising

python -u test_certify_jsrt.py --weights weights/jsrt_deeplab/best_model.pt --model deeplab --dataset jsrt --multi --sigma 0.25

Usage for single-class segmentation

Launch certification for skin lesions dataset with a UNet model

python -u test_certify.py --weights weights/lesion_unet/best_model.pt --model unet --dataset lesion --sigma 0.25 --denoise

Launch certification for skin lesions dataset with a UNet model and without denoising

python -u test_certify.py --weights weights/lesion_unet/best_model.pt --model unet --dataset lesion --sigma 0.25

Launch certification for polyp dataset with a ResUNet++ model

python -u test_certify.py --weights weights/polyp_resunetpp/best_model.pt --model resunetpp --dataset polyp --sigma 0.25 --denoise

Launch certification for polyp dataset with a ResUNet++ model and without denoising

python -u test_certify.py --weights weights/polyp_resunetpp/best_model.pt --model resunetpp --dataset polyp --sigma 0.25

Similar commands can be used for single-class lung segmentation (montgomery and shenzhen datasets).

Citation

If you find this work useful, please consider citing it:

@InProceedings{laousy23miccai,
author="Laousy, Othmane and Araujo, Alexandre and Chassagnon, Guillaume and Paragios, Nikos and Revel, Marie-Pierre and Vakalopoulou, Maria",
editor="Greenspan, Hayit and Madabhushi, Anant and Mousavi, Parvin and Salcudean, Septimiu and Duncan, James and Syeda-Mahmood, Tanveer and Taylor, Russell",
title="Certification of Deep Learning Models for Medical Image Segmentation",
booktitle="Medical Image Computing and Computer Assisted Intervention -- MICCAI 2023",
year="2023",
publisher="Springer Nature Switzerland",
address="Cham",
pages="611--621",
isbn="978-3-031-43901-8"
}