| dc.description.abstract | Presence of clutters, occlusions and dark corner artifacts in dermatoscopy images causes unsupervised and intensity-based image analysis systems to erroneously segment lesion boundaries required for accurate and reliable skin cancer diagnosis. Preprocessing algorithms designed to address these challenges increase resources, computational cost and introduce extraneous features, thereby reducing the efficacy of the diagnostic system. We propose a new approach to accurately segment skin lesions without the need for preprocessing steps to eliminate these confounding factors. The proposed method begins by thresholding with a correction factor in a color channel image with optimal discrimination between the target object and background. Next, the output of the preliminary segmentation undergoes angular displacement. Finally, we iterate, a number of times, the set difference between the binarized image and its rotated version, to simultaneously detect lesion borders and eliminate occlusions and clutters. The proposed method outperform selected state-of-the-art segmentation algorithms on 600 images with different types of confounding factors. | en_US |
