2019/07/01 by Md. Mostafa Kamal Sarker, Sarker, Md. Mostafa Kamal, Hatem A. Rashwan +19 · 1 citation
Medicine · #Computer Vision and Pattern Recognition (cs.CV) #Cutaneous Melanoma Detection and Management #FOS: Computer and information sciences #FOS: Electrical engineering #Hair Growth and Disorders #Image and Video Processing (eess.IV) #Nonmelanoma Skin Cancer Studies #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1907.00856
openalex publication_date 2019/07/01 · openalex created_date 2021/07/05 · openalex updated_date 2026/07/29
The determination of precise skin lesion boundaries in dermoscopic images\nusing automated methods faces many challenges, most importantly, the presence\nof hair, inconspicuous lesion edges and low contrast in dermoscopic images, and\nvariability in the color, texture and shapes of skin lesions. Existing deep\nlearning-based skin lesion segmentation algorithms are expensive in terms of\ncomputational time and memory. Consequently, running such segmentation\nalgorithms requires a powerful GPU and high bandwidth memory, which are not\navailable in dermoscopy devices. Thus, this article aims to achieve precise\nskin lesion segmentation with minimum resources: a lightweight, efficient\ngenerative adversarial network (GAN) model called SLSNet, which combines 1-D\nkernel factorized networks, position and channel attention, and multiscale\naggregation mechanisms with a GAN model. The 1-D kernel factorized network\nreduces the computational cost of 2D filtering. The position and channel\nattention modules enhance the discriminative ability between the lesion and\nnon-lesion feature representations in spatial and channel dimensions,\nrespectively. A multiscale block is also used to aggregate the coarse-to-fine\nfeatures of input skin images and reduce the effect of the artifacts. SLSNet is\nevaluated on two publicly available datasets: ISBI 2017 and the ISIC 2018.\nAlthough SLSNet has only 2.35 million parameters, the experimental results\ndemonstrate that it achieves segmentation results on a par with the\nstate-of-the-art skin lesion segmentation methods with an accuracy of 97.61%,\nand Dice and Jaccard similarity coefficients of 90.63% and 81.98%,\nrespectively. SLSNet can run at more than 110 frames per second (FPS) in a\nsingle GTX1080Ti GPU, which is faster than well-known deep learning-based image\nsegmentation models, such as FCN. Therefore, SLSNet can be used for practical\ndermoscopic applications.\n