Medical Image Super Resolution by Preserving Interpretable and Disentangled Features
Abstract
State of the art image super resolution (ISR) methods use generative networks to produce high resolution (HR) images from their low resolution (LR) counterparts. In this paper we show with the help of interpretable saliency maps that generative approaches to ISR can introduce undesirable artefacts which can adversely affect model performance in downstream tasks such as disease classification. We propose a novel loss term that aims to maximize the similarity of interpretable class activation maps between the HR and LR images. This not only preserves the images’ explainable information but also leads to improved performance in terms of super resolution output and downstream classification accuracy. We also incorporate feature disentanglement that plays an important role in our method’s superior performance for super resolution and downstream classification task.
Cite
Text
Mahapatra et al. "Medical Image Super Resolution by Preserving Interpretable and Disentangled Features." European Conference on Computer Vision Workshops, 2022. doi:10.1007/978-3-031-25082-8_48Markdown
[Mahapatra et al. "Medical Image Super Resolution by Preserving Interpretable and Disentangled Features." European Conference on Computer Vision Workshops, 2022.](https://mlanthology.org/eccvw/2022/mahapatra2022eccvw-medical/) doi:10.1007/978-3-031-25082-8_48BibTeX
@inproceedings{mahapatra2022eccvw-medical,
title = {{Medical Image Super Resolution by Preserving Interpretable and Disentangled Features}},
author = {Mahapatra, Dwarikanath and Bozorgtabar, Behzad and Reyes, Mauricio},
booktitle = {European Conference on Computer Vision Workshops},
year = {2022},
pages = {709-721},
doi = {10.1007/978-3-031-25082-8_48},
url = {https://mlanthology.org/eccvw/2022/mahapatra2022eccvw-medical/}
}