Search

The automatic classification of skin lesions in dermoscopy images remains challenging due to the morphological diversity of skin lesions, the existence of intrinsic cutaneous features and artefacts, the lack of training data, and the insufficient recognition abilities of current methods. To meet these challenges, we construct a new densely connected convolutional network termed DenseSFNet-45, which is obtained by integrating our proposed novel architectural unit (an SE-Fire (SF) block) into the dense block of a dense convolutional network (DenseNet). The SF block consists of a cascade of a Fire module and a squeeze-and-excitation (SE) block, enhancing the representational power of DenseNet by exploiting both spatial and channel-wise information. Based on DenseSFNet, we propose a novel two-stage framework consisting of skin lesion segmentation followed by lesion classification to accurately classify skin lesions. The classification step is performed on the segmented lesion rather than the whole dermoscopy image, enabling the classification network to extract more specific and discriminative features. The proposed method is extensively evaluated on three public databases: ISBI 2017 Skin Lesion Analysis Towards Melanoma Detection Challenge dataset (ISBI-skin-2017), ISBI 2018 Skin Lesion Analysis Towards Melanoma Detection Challenge dataset (ISBI-skin-2018), and PH2 dataset. The experimental results demonstrate the superior performance of our method relative to that of the traditional machine learning algorithms, the existing classical classification models, baselines, and state-of-the-art methods.

Accurate segmentation of skin lesion is an important step in computer aided diagnosis of skin cancer. Recently, deep-learning-based image segmentation methods have drawn much attention and shown exacting results. Unfortunately, existing skin disease datasets can hardly satisfy the requirement of massive training samples for deep neural networks. To meet this challenge, we introduce a data augmentation method using deep convolutional generative adversarial network (DCGAN), which can generate realistic samples with lesion features learned from the existing dataset, increasing both the quantity and diversity of training samples. The architecture of our proposed segmentation algorithm is built upon deep fully connected networks (FCN) and the DenseNet is employed as feature extractor. Extensive experiments are carried out on “ISBI 2018: Skin Lesion Analysis Towards Melanoma Detection Challenge dataset”, and the results demonstrate that the proposed algorithm significantly improves the accuracy of skin lesions segmentation without requiring new actual training samples.