How Important Is Each Dermoscopy Image?

Catarina Barata, Carlos Santiago; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 742-743

Deep neural networks (DNNs) have revolutionized the field of dermoscopy image analysis. Systems based on DNNs are able to achieve impressive diagnostic performances, even outperforming experienced dermatologists. However, DNNs strongly rely on the quantity and quality of the training data. Real world data sets, including those related to dermoscopy, are often severely imbalanced and of reduced dimensions. Thus, models trained on these data sets typically become biased and fail to generalize well to new images. Sample weighting strategies have been proposed to overcome the previous limitations with promising results. Nonetheless, they have been poorly investigated in the context of dermoscopy image analysis. This paper addresses this issue through the extensive comparison of several sample weighting methods, namely class balance and curriculum learning. The results show that each sample weighting strategy influences the performance of the model in different ways, with most finding a compromise between correctly classifying the most common classes or biasing the model towards the less represented classes. Furthermore, the features learned by each model differ significantly, depending on the training strategy.