Assessing the impact of segmentation on skin lesion classification accuracy

Skin lesion segmentation is a possible pre-processing step in a skin lesion classification processing chain. Previous studies have shown that segmentation can increase or decrease classification accuracy, depending on several factors. In this paper, we extend those studies, by considering additional segmentation and classification models and by employing a different, wider data set. Our results show that when the segmentation mask is fed to the network as an additional channel, in parallel with the lesion image, it has a beneficial impact on performance. In contrast, when the segmentation mask is exploited to produce and extract a bounding box containing the lesion, the classification performance is reduced.