⭐ The highlight of this work:
- Visualization of explanation methods via saliency maps (different explanation methods).
- Evaluation of explanation methods via perturbation-based method.
- Proposed adatped Area over Perturbation Curve (AOPC) metrics (Samek et al. 2017 ) for regression.
- Comparison of regression models through qualitative (saliency maps) and quantitative (AOPC score) analysis.
- Saliency maps for correct vs incorrect prediction.
💻 About the code:
The code is implimented with Python 3.* and Deep learning library Tensorflow (Keras 2.*) and public library iNNvestigate.
The work is finished together with Caroline Petitjean and Samia Ainouz in LITIS lab and Florian Yger.
Please consider citing this paper when you use it:
@incollection{zhang2020explainability,
title={Explainability for regression CNN in fetal head circumference estimation from ultrasound images},
author={Zhang, Jing and Petitjean, Caroline and Yger, Florian and Ainouz, Samia},
booktitle={Interpretable and Annotation-Efficient Learning for Medical Image Computing},
pages={73--82},
year={2020},
publisher={Springer}
}