Exploring LRP and Grad-CAM visualization to interpret multi-label-multi-class pathology prediction using chest radiography
Authors:
- 1 July 2022
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 258-263
Abstract
The area of interpretable deep neural networks has received increased attention in recent years due to the need for transparency in various fields, including medicine, healthcare, stock market analysis, compliance with legislation, and law. Layer-wise Relevance Propagation (LRP) and Gradient-weighted Class Activation Mapping (Grad-CAM) are two widely used algorithms to interpret deep neural networks. In this work, we investigated the applicability of these two algorithms in the sensitive application area of interpreting chest radiography images. In order to get a more nuanced and balanced outcome, we use a multi-label classification-based dataset and analyze the model prediction by visualizing the outcome of LRP and Grad-CAM on the chest radiography images. The results show that LRP provides more granular heatmaps than Grad-CAM when applied to the CheXpert dataset classification model. We posit that this is due to the inherent construction difference of these algorithms (LRP is layer-wise accumulation, whereas Grad-CAM focuses primarily on the final sections in the model's architecture). Both can be useful for understanding the classification from a micro or macro level to get a superior and interpretable clinical decision support system.Keywords
This publication has 18 references indexed in Scilit:
- Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologistsPLOS Medicine, 2018
- Opportunities and obstacles for deep learning in biology and medicineJournal of The Royal Society Interface, 2018
- Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization2017
- Densely Connected Convolutional Networks2017
- Explaining nonlinear classification decisions with deep Taylor decompositionPattern Recognition, 2017
- Learning Deep Features for Discriminative Localization2016
- On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance PropagationPLOS ONE, 2015
- Deep learningNature, 2015
- ImageNet: A large-scale hierarchical image database2009
- Learning representations by back-propagating errorsNature, 1986