Cfdd-net: a novel comprehensive featured-dual decoding network for melanoma segmentation in dermoscopic images

Precise dermoscopic image segmentation, which is challenging for complexity of lesion regions, is pivotal in skin melanoma diagnosis. In this paper, we introduce a novel Comprehensive Featured-Dual Decoding Network (CFDD-Net) with a dual-route decoding path to enhance the lesion segmentation performance on dermoscopic images. In the proposed network, the cascaded dilated convolution (CDC), dilated convolution dual attention (DCDA) and dual wavelet dilated convolution (DWDC) blocks are specifically designed in the skip connections in addition to the atrous spatial pyramid pooling (ASPP) block for the exploration and transmission of the multi-scale features. Meanwhile, the X-shape blocks are novelly proposed to thoroughly fuse the multi-scale features across the two decoding routes. The performance of CFDD-Net is thoroughly evaluated and is proved superior to the state-of-the-art methods, with Dice reaching 92.33%, 91.72% and 94.82%, mean Intersection-over-Union (mIoU) reaching 86.06%, 85.78% and 90.33%, and Average Symmetric Surface Distance (ASSD) reaching 0.54 pix, 0.55 pix, 0.26 pix on three public skin melanoma region segmentation datasets, ISIC-2018, ISIC-2017 and ISIC-2016, through experiments, respectively. Compared to the state-of-the-art methods, the CFDD-Net provides improved segmentation performance, and thus, may help dermatologists to provide accurate lesion segmentation in clinical practice.