DEU-Net: Dual-Encoder U-Net for Automated Skin Lesion Segmentation

This repository contains the official PyTorch implementation of "DEU-Net: Dual-Encoder U-Net for Automated Skin Lesion Segmentation".

Prerequisites

• PyTorch 2.0.0
• Python 3.9
• Additional packages listed in requirements.txt.

Data Preparation

We use four datasets in this project: ISIC2016, ISIC2017, ISIC2018, and PH2.

1. Download the ISIC datasets from https://challenge.isic-archive.com/data/ and place them into the corresponding folders: datasets/ISIC2016, datasets/ISIC2017, and datasets/ISIC2018.

2. Download the PH2 dataset from https://www.fc.up.pt/addi/ph2%20database.html and place it in the datasets/PH2 folder.

Once you have downloaded the data, you need to preprocess it using the following steps:

1. (Optional) To increase network training speed, you can use the preprocess_resize_dataset.py script to resize the dataset to your desired dimensions. Note that this step is not necessary, as the network training script will automatically resize the input data.

2. Use the preprocess_split_dataset.py script to specify training, validation, and test data and store them in the data folder. The files generated will be used for training and testing the network.

Network Training

You can train the network using the train.py script, which takes the following arguments:

• --net-cfg: A configuration file for the desired network (examples can be found in cfg/net folder)
• --train-cfg: A configuration file for the network training settings, such as the number of epochs, optimizer, etc. (example can be found at cfg/train.cfg)
• --dataset: The directory of the dataset (this is the same folder obtained from the second stage of data preparation)

There are several other arguments that can be found in the code.

Network Evaluation

You can evaluate the network using the evaluate.py script, which takes the following arguments:

• --cfg: One or more config files for the desired networks
• --model: One or more files containing the weights of the networks. Multiple models can be used simultaneously for an ensemble output.
• --transforms: A list of transformations to perform on each input image and calculate the final result from the combination of results (possible transformations: vflip, hflip, rotation_90, rotation_180, rotation_270)
• --input-images: A txt file containing the paths of the images to be segmented
• --input-masks: A txt file containing the paths of the ground truth masks corresponding to the input images

If you don't want to apply any transformations and only want to segment the original image, leave the --transforms argument empty.

There are several other arguments that can be found in the code.

Prediction

You can use the predict.py script to produce output for multiple images. The script takes the following arguments:

• --cfg and --model and --transforms: Similar to evaluate.py. Multiple models can be used simultaneously for an ensemble output.
• --input: A list of paths for input images to be segmented

There are several other arguments that can be found in the code.

Citation

If you find this repo helpful, please consider citing our paper:

@ARTICLE{10332179,
  author={Karimi, Ali and Faez, Karim and Nazari, Soheila},
  journal={IEEE Access}, 
  title={DEU-Net: Dual-Encoder U-Net for Automated Skin Lesion Segmentation}, 
  year={2023},
  volume={11},
  number={},
  pages={134804-134821},
  doi={10.1109/ACCESS.2023.3337528}}