BackGround Masked Guided Network for Skin Lesion Segmentation in Dermoscopy Image

Skin lesion segmentation in dermoscopy images plays a critical role in the automated diagnosis of melanoma. In this study, to alleviate the impact of the background information in dermoscopy images, a novel framework dubbed Background Masked-Guided Network is proposed to utilize images with background masked to guide the backbone for final segmentation results. Moreover, because traditional CNN will bring inevitable information loss when feeding into activation functions, to alleviate this phenomenon and utilize the advantages of U-Net architecture, we propose a new backbone named U-ConvNext. In U-ConvNext, the architecture of U-Net is retained, and in the Up-Sampling Block, the inverted bottleneck architecture is imported to reduce information loss caused by dimension compression. We evaluate our proposed methods on the ISIC2017 and PH2 datasets; the experimental results demonstrated that our proposed methods could achieve excellent segmentation results on both datasets.