In a recent work, we analyzed how convolutional neural network architectures influence the resulting relevance maps with respect to homogeneity and noise. We compared four model layouts, which actually didn’t differ much with respect to model performance, i.e. accuracy and AUC. The DenseNet models provided the most focused and coherent information.
Checkout our conference paper “Comparison of CNN Architectures for Detecting Alzheimer’s Disease using Relevance Maps” for further details, DOI: 10.1007/978-3-658-41657-7_51.
You can also find the code on GitHub.
