Figure 4.

Attribution images

The attribution images show which pixels the classifier weights when “deciding” how to categorize. As shown, the classifier uses pixels involved in the skin finding but may also use other pixels as well, some of which may represent confounders.²⁸ Our research team examined these attribution methods during stages of classifier training and testing to determine how to improve performance, such as by incorporating other datasets for training, or when adjusting the neural network hyperparameters. Clockwise, from top left: NF1, TSC, ML, and MA. Images sources (all are used with appropriate permission) are listed in the Web resources.