Table 1.

Adaptation of used methods for benchmarking

Method name	Adaptation	Advantages	Disadvantages
PXPermute	none	+ open source + model agnostic + simple concept + keeps the data distribution + no parameter tuning	− higher runtime
Channel-wise occlusion	occluding channels instead of pixels	+ rapid runtime + model agnostic + simple concept + no parameter tuning	− changes the data distribution
Guided GradCAM (Selvaraju et al.25)	using the median of the values per channel	+ rapid runtime + open source	− model specific − requires parameter tuning − needs adaptation
Integrated gradients (Sundararajan et al.38)	using the median of the values per channel	+ rapid runtime + open source	− model specific − requires parameter tuning − needs adaptation
DeepLift (Shrikumar et al.37)	using the median of the values per channel	+ rapid runtime + open source	− model specific − requires parameter tuning − needs adaptation
LRP (Bach et al.39)	using the median of the values per channel	+ rapid runtime + open source	− model-specific − requires parameter tuning − needs adaptation

Apart from PXPermute, designed for channel importance, all other methods were adapted from their original design. Their other advantages and disadvantages are based on model specificity, design complexity, runtime, and the need for tuning.