The Effects of Masking in Melanoma Image Classification with CNNs Towards International Standards for Image Preprocessing

The classification of skin lesion images is known to be biased by artifacts of the surrounding skin, but it is still not clear to what extent masking out healthy skin pixels influences classification performances, and why. To better understand this phenomenon, we apply different strategies of image masking (rectangular masks, circular masks, full masking, and image cropping) to three datasets of skin lesion images (ISIC2016, ISIC2018, and MedNode). We train CNN-based classifiers, provide performance metrics through a 10-fold cross-validation, and analyse the behaviour of Grad-CAM saliency maps through an automated visual inspection. Our experiments show that cropping is the best strategy to maintain classification performance and to significantly reduce training times as well. Our analysis through visual inspection shows that CNNs have the tendency to focus on pixels of healthy skin when no malignant features can be identified. This suggests that CNNs have the tendency of “eagerly” looking for pixel areas to justify a classification choice, potentially leading to biased discriminators. To mitigate this effect, and to standardize image preprocessing, we suggest to crop images during dataset construction or before the learning step.