Evaluation of different segmentation-based approaches for skin disorders from dermoscopic images

Abstract

Skin disorders are the most common type of cancer in the world and the incident has been lately increasing over the past decades. Even with the most complex and advanced technologies, current image acquisition systems do not permit a reliable identification of the skin lesion by visual examination due to the challenging structure of the malignancy. This promotes the need for the implementation of automatic skin lesion segmentation methods in order to assist in physicians’ diagnostic when determining the lesion's region and to serve as a preliminary step for the classification of the skin lesion. Accurate and precise segmentation is crucial for a rigorous screening and monitoring of the disease's progression. For the purpose of the commented concern, the present project aims to accomplish a state-of-the-art review about the most predominant conventional segmentation models for skin lesion segmentation, alongside with a market analysis examination. With the rise of automatic segmentation tools, a wide number of algorithms are currently being used, but many are the drawbacks when employing them for dermatological disorders due to the high-level presence of artefacts in the image acquired. In light of the above, three segmentation techniques have been selected for the completion of the work: level set method, an algorithm combining GrabCut and k-means methods and an intensity automatic algorithm developed by Hospital Sant Joan de Déu de Barcelona research group. In addition, a validation of their performance is conducted for a further implementation of them in clinical training. The proposals, together with the got outcomes, have been accomplished by means of a publicly available skin lesion image database.