MFSNet

This repository contains the official implementation of our paper titled "MFSNet: A Multi Focus Segmentation Network for Skin Lesion Segmentation" published in Pattern Recognition, Elsevier.
[Preprint]

Preprocessing

To run the script for inpainting, run the following using the command prompt:

python inpaint.py --root "D:/inputs/" --destination "D:/images/"

Training the Network

Follow the directory structure as follows:

+-- data
|   +-- .
|   +-- train
|   |   +-- images
|   |   +-- masks
|   +-- test
|   |   +-- images
|   |   +-- masks
+-- train.py
+-- test.py

Run the following to train the MFSNet network:

python train.py --train_path "data/train"

Other available hyperparameters for training are as follows:

• --epoch: Number of epochs of training. Default = 100
• --lr: Learning Rate. Default = 1e-4
• --batchsize: Batch Size. Default = 20
• --trainsize: Size of Training images (to be resized). Default = 352
• --clip: Gradient Clipping Margin. Default = 0.5
• --decay_rate: Learning rate decay. Default = 0.05
• --decay_epoch: Number of epochs after which Learning Rate needs to decay. Default = 25

After the training is complete, run the following to generate the predictions on the test images:

python test.py --test_path "data/test"

Evaluating Performance

Run eval/main.m using MATLAB on the ground truth images and the predicted masks, to get the evaluation measures.

Citation

If you find this repository useful, please cite our work:

@article{basak2022mfsnet,
  title={MFSNet: A Multi Focus Segmentation Network for Skin Lesion Segmentation},
  author={Basak, Hritam and Kundu, Rohit and Sarkar, Ram},
  journal={Pattern Recognition},
  pages={108673},
  year={2022},
  publisher={Elsevier}
}