Design and validation of contrast enhancement decolourization and segmentation algorithms for macroscopic images of skin lesions

Abstract

Dermatological photography (macroscopy) is an emerging imaging modality newlineextensively used for visualizing the skin lesions. As the macroscopy uses commercial newlinecameras that are easily available, the macroscopic images are widely used in screening newlinedermatology instead of the dermoscopy images. The potential of the wide-field newlinedermatological photo-macrographs to be used as a tool for identifying the type of the newlinesuspicious skin lesions at the newlinepre-screening level is a proven one. newlineUsually, aggressiveness of skin lesions that points to malignancy is newlinecharacterized by geometric/shape features like area, solidity, eccentricity, etc. newlineConsequently, the diagnosis of the skin lesions greatly depends on the accurate newlinesegmentation of the lesions. Automated segmentation algorithms are necessary to newlineeliminate the inter-operator variability inherent in the subjective contouring of skin newlinelesions. Automated segmentation is highly challenging in the presence of uneven newlinebackground illumination and when the input images have relatively low contrast newlinebetween skin lesions and normal skin regions, on macroscopic images. Majority of the newlineavailable image segmentation and feature extraction algorithms are designed for newlinegrayscale images. Hence, conversion of dermatological colour images to grayscale newlinespace is an important step in their automated analysis. newlineAs the macroscopic images are closely-focused views, and contain only newlinelesions as well as the background skin, the segmentation is a thresholding problem. newlineHowever, existing threshold estimation algorithms often fail to predict the threshold newlinevalue that facilitates precise distinction between the lesion and the surrounding skin. newlinePerformance of the majority of the threshold prediction algorithms is image-dependent newlineand inconsistent. Existing illumination correction and contrast enhancement techniques newlinethat produce visually-appealing output images may not really improve the newlinesegmentation. newline