SLSDeep: Skin Lesion Segmentation Based on Dilated Residual and Pyramid Pooling Networks

Abstract

Skin lesion segmentation (SLS) in dermoscopic images is a crucial task for automated diagnosis of melanoma. In this paper, we present a robust deep learning SLS model, so-called SLSDeep, which is represented as an encoder-decoder network. The encoder network is constructed by dilated residual layers, in turn, a pyramid pooling network followed by three convolution layers is used for the decoder. Unlike the traditional methods employing a cross-entropy loss, we investigated a loss function by combining both Negative Log Likelihood (NLL) and End Point Error (EPE) to accurately segment the melanoma regions with sharp boundaries. The robustness of the proposed model was evaluated on two public databases: ISBI 2016 and 2017 for skin lesion analysis towards melanoma detection challenge. The proposed model outperforms the state-of-the-art methods in terms of segmentation accuracy. Moreover, it is capable to segment more than 100 images of size 384 × 384 per second on a recent GPU.

Construction of saliency map and hybrid set of features for efficient segmentation and classification of skin lesion

“…The few best results are given in Table . The maximum achieved segmentation accuracy 98.88% and average accuracy is 95.989% which is superior as compare to recent state‐of‐the‐art techniques (Rahman, Alpaslan, & Bhattacharya, ; Sarker et al, ; Yu, Chen, Dou, Qin, & Heng, ).…”

Non-Local Context Encoder: Robust Biomedical Image Segmentation against Adversarial Attacks

“…SLSDeep, consisting of skipconnections, dilated residual and pyramid pooling, is an efficient skin lesion segmentation model from dermoscopic images. Its loss function, including negative log likelihood and end point error loss, is designed to obtain sharp boundary (Sarker et al 2018).…”

“…We compare the robustness of our proposed NLCEN with that of five state-of-the-art methods for lung segmentation and skin lesion segmentation, including dilated residual and pyramid pooling networks (SLSDeep) (Sarker et al 2018), network-wise training of convolutional networks (NWCN) (Hwang and Park 2017), convolutional networks for biomedical image segmentation (UNet) (Ronneberger, Fischer, and Brox 2015), fully convolutional architectures for multi-class segmentation (InvertNet) (Novikov et al 2018) and segmentation with fully convolutionaldeconvolutional networks (CDNN) (Yuan 2017 Quantitative Evaluation Figures 4 and 5 show evaluation results in terms of DIC and JSC on the JSRT and ISBI 2016 datasets respectively. In these figures, we can find that NLCEN achieves the highest accuracy on clean skin lesion images, and achieves almost the top performance on clean lung images.…”

“…X-ray, Magnetic Resonance Imaging, Computerized Tomography). However, it is challenging to obtain accurate segmentation because of the large shape and size variations of regions of interest, and the diversity of images produced by different biomedical imaging equipments (Hwang Figure 1: Sample adversarial attacks on SLSDeep (Sarker et al 2018), NWCN (Hwang and Park 2017), CDNN (Yuan 2017) and our NLCEN. The input chest radiograph and its ground-truth segmentation are shown in the first row.…”

“…Segmentation results on the adversarial images and the input image are shown in the fourth and fifth rows respectively. and Park 2017; Sarker et al 2018). State-of-the-art biomedical image segmentation methods are based on fully convolutional networks (FCN) (Long, Shelhamer, and Darrell 2015), which is a type of deep convolutional neural networks (CNNs) designed for semantic segmentation in computer vision.…”