Automatic skin lesion segmentation based on FC-DPN

Abstract

Automatic skin lesion segmentation in dermoscopy images is challenging due to the diversity of skin lesion characteristics, low contrast between normal skin and lesions, and the existence of many artefacts in the images. To meet these challenges, we propose a novel segmentation topology called FC-DPN, which is built upon a fully convolutional network (FCN) and dual path network (DPN). The DPN inherits the advantages of residual and densely connected paths, enabling effective feature re-usage and re-exploitation. We replace dense blocks in fully convolutional DenseNets (FC-DenseNets) with two kinds of sub-DPN blocks, namely, sub-DPN projection blocks and sub-DPN processing blocks. This framework enables FC-DPN to acquire more representative and discriminative features for more accurate segmentation. Many images in the original ISBI 2017 Skin Lesion Challenge test dataset are given the incorrect or inaccurate ground truths, and these ground truths have been revised. The revised test dataset is called the modified ISBI 2017 Skin Lesion Challenge test dataset. The proposed method achieves an average Dice coefficient of 88.13% and a Jaccard index of 80.02% on the modified ISBI 2017 Skin Lesion Challenge test dataset and 90.26% and 83.51%, respectively, on the PH2 dataset. Extensive experimental results on the two datasets demonstrate that the proposed method exhibits better performance than FC-DenseNets and other well-established segmentation algorithms.