Skin cancer is a deadly disease and is on the rise in the world. Computerizeddiagnosis of skin cancer can accelerate the detection of this type of cancerthat is a key point in increasing the survival rate of patients. Lesionsegmentation in skin images is an important step in computerized detection ofthe skin cancer. Existing methods for this aim usually lack accuracy especiallyin fuzzy borders of lesions. In this paper, we propose a new class of fullyconvolutional networks with novel dense pooling layers for segmentation oflesion regions in non-dermoscopic images. Unlike other existing convolutionalnetworks, the proposed dense pooling layers are designed to preserve all of theinput features. This has led to highly accurate segmentation of lesions. Ourproposed method produces dice score of 91.6% which outperforms allstate-of-the-art algorithms in segmentation of skin lesions based on theDermquest dataset.
展开▼
