In a nutshell
-
For COVID-19 and Thoracic pathology neural network classification/regression models, we first identify the concepts (e.g. pathologies) associated with internal units (feature maps) of the network
We investigate the following questions:
- Does a regression model that is only trained with COVID-19 severity scores implicitly learn visual patterns associated with thoracic pathologies?
-
Does a network that is trained on weakly labeled data (e.g. healthy, unhealthy) implicitly learn pathologies (e.g. Consolidation)?
- We also investigate the effect of pretraining and data imbalance on the interpretability of learned features
- Finally, we propose a semantic attribution (saliency) method to semantically explain each prediction
Resources
View the paper on arXiv (The camera-ready version will appear in the proceedings of MICCAI 2021.)
Check the Code on GitHub
Citation
Please cite the work using the below BibTeX (also available on the Open Access link above)
@misc{khakzar2021semantic,
title={Towards Semantic Interpretation of Thoracic Disease and COVID-19 Diagnosis Models},
author={Ashkan Khakzar and Sabrina Musatian and Jonas Buchberger and Icxel Valeriano Quiroz and Nikolaus Pinger and Soroosh Baselizadeh and Seong Tae Kim and Nassir Navab},
year={2021},
eprint={2104.02481},
archivePrefix={arXiv},
primaryClass={eess.IV}
}
Contact
For inquiries and feedback please contact Ashkan Khakzar (ashkan.khakzar@tum.de). We would be happy to help and we appreciate your feedback.