sarmad-m / msrf-net

MSRF-Net: A Multi-Scale Residual Fusion Network for Biomedical Image Segmentation

This repository provides code for our paper "MSRF-Net: A Multi-Scale Residual Fusion Network for Biomedical Image Segmentation" accepted for Publication at IEEE Journal of Biomedical and Health Informatics (arxiv version)(ieeexplore version)

2.) Overview

2.1.)Introduction

In this work, we propose a novel medical imagesegmentation architecture, calledMSRF-Net, which aims toovercome the above limitations. Our proposed MSRF-Netmaintains high-resolution representation throughout the pro-cess which is conducive to potentially achieving high spatialaccuracy. The MSRF-Net utilizes a novel dual-scale dense fusion (DSDF) block that performs dual scale feature exchangeand a sub-network that exchanges multi-scale features usingthe DSDF block. The DSDF block takes two different scaleinputs and employs a residual dense block that exchanges in-formation across different scales after each convolutional layerin their corresponding dense blocks. The densely connectednature of blocks allows relevant high- and low-level featuresto be preserved for the final segmentation map prediction. Themulti-scale information exchange in our network preservesboth high- and low-resolution feature representations, therebyproducing finer, richer, and spatially accurate segmentationmaps. The repeated multi-scale fusion helps in enhancing thehigh-resolution feature representations with the informationpropagated by low-resolution representations. Further, layersof residual networks allow redundant DSDF blocks to die out,and only the most relevant extracted features contribute to thepredicted segmentation maps.

2.2.) DSDF Blocks and MSRF Sub-network

2.3.) Quantitative Results

3.) Training and Testing

3.1)Data Preparation

1.) make directory named "data/kdsb"

2.) make three sub-directories "train" "val" "test"

3.) Put images under directory named "images"

4.) Put masks under directory named "masks"

3.2)Training

Model architecture is defined in model.py Run the script as: python train.py

3.2)Testing

For testing the trained model run: python test.py

4.) Citation

Please cite our paper if you find the work useful:

@article{srivastava2021msrf,
  title={MSRF-Net: A Multi-Scale Residual Fusion Network for Biomedical Image Segmentation},
  author={Srivastava, Abhishek and Jha, Debesh and Chanda, Sukalpa and Pal, Umapada and Johansen, H{\aa}vard D and Johansen, Dag and Riegler, Michael A and Ali, Sharib and Halvorsen, P{\aa}l},
  journal={arXiv preprint arXiv:2105.07451},
  year={2021}
}

5.) FAQ

Please feel free to contact me if you need any advice or guidance in using this work (E-mail)