Low Effectiveness of Non-Geometric-Operation Data Augmentations for Lesion Segmentation with Fully Convolution Networks

Data augmentation is a prevalent strategy to enlarge the training data in order to enhance the generalization of the model of deep convolutional neural networks (DCNNs). However, not all of augmentation schemes are always effective for all types of DCNNs models, especially for fully convolutional networks (FCNs) which greatly improved semantic segmentation by employing a skip architecture that fuses the feature hierarchy to combine deep, coarse, semantic information and shallow, fine, appearance information. In order to make the effectiveness of data augmentation clear, in this work, we propose to divide the augmentation schemes into two groups, geometric operations and non-geometric operations. Through analyzing the performance of them for lesion segmentation with FCNs, it is found that non-geometric-operation data augmentations are less effective in two dermoscopy datasets. Moreover, we further theoretically revealed that the skip architecture in FCNs is the main reason behind this finding. This work is of value on guiding the practice of data augmentation while using FCNs, and enlightening significance for analyzing other skip architecture deep neural networks.