U-Net Based Segmentation and Multiple Feature Extraction of Dermascopic Images for Efficient Diagnosis of Melanoma

Abstract

Skin cancer is found to be one of the most common types of deadly cancers among human beings in recent years. Computational-based techniques are developed to support the dermatologists for the early diagnosis of skin cancer. Computational analysis of the skin lesions in the dermascopic images is a challenging task due to the difficulties such as low-level of contrast between the lesion and surrounding skin regions, irregular and vague lesion borders, artifacts and poor imaging conditions. This paper presents a U-Net based segmentation and multiple feature extraction of the dermascopic images for the efficient diagnosis of skin cancer. The input dermascopic image is preprocessed to remove the noise and hair in the skin image. Fast Independent Component Analysis (FastICA) is applied to the skin images for obtaining the melanin and hemoglobin components. The U-net segmentation is applied to the dermascopic image to separate the cancer region from the background of the skin image. Different features such as vascular features, color features, texture features, RGB features, and depth features are extracted from the segmented image. RVM classification is applied to classify the normal and abnormal images. With the efficient segmentation and extraction of multiple features, our proposed work yields better performance than the existing segmentation and feature extraction techniques.