Modified watershed technique and post-processing for segmentation of skin lesions in dermoscopy images

作者：Wang Hanzheng^*; Moss Randy H; Chen Xiaohe; Stanley R Joe; Stoecker William V; Celebi M Emre; Malters Joseph M; Grichnik James M; Marghoob Ashfaq A; Rabinovitz Harold S; Menzies Scolt W; Szalapski Thomas M

来源：Computerized Medical Imaging and Graphics, 2011, 35(2): 116-120.

DOI：10.1016/j.compmedimag.2010.09.006

• Malignant melanoma
• Watershed
• Image processing
• Segmentation
• Neural network

In previous research, a watershed-based algorithm was shown to be useful for automatic lesion segmentation in dermoscopy images, and was tested on a set of 100 benign and malignant melanoma images with the average of three sets of dermatologist-drawn borders used as the ground truth, resulting in an overall error of 15.98%. In this study, to reduce the border detection errors, a neural network classifier was utilized to improve the first-pass watershed segmentation; a novel "edge object value (EOV) threshold" method was used to remove large light blobs near the lesion boundary; and a noise removal procedure was applied to reduce the peninsula-shaped false-positive areas. As a result, an overall error of 11.09% was achieved.

• 出版日期2011-3