Saliency Induced Fusion for Skin Lesion Detection
Abstract
Skin lesion segmentation from dermoscopic images is one of the most important and fundamental step in computer-aided diagnosis (CAD). It is more challenging due to the presence of hairs, ruler marks, gels, dark corners, colour inconsistency, shape, size etc. in the images. It is highly essential by ignoring the aforementioned challenging factsto extract the accurate skin lesions. We have proposed a saliency-based approach that initially detects the different saliency maps i.e., frequency, color, location, covariance and mean. The fusion of different saliency maps is implemented in the proposed method for enhancing the lesion regions so that it will help in the segmentation process to extract the lesions precisely. One of the simplest algorithm i.e., Otsu algorithm is used for segmentation of skin lesions from dermoscopic images. The ffnal lesion mask is obtained by further processing the segmentation output with the post processing techniques. For evaluating the proposed method, the images from ISIC datasets are considered. A large number of images having aforementioned challenging factors are considered to compute the performance metrics. The proposed method is able to extract the lesion masks and lesion regions more accurately. The method archives an average accuracy of 96.29%, 96.48% and 95.89% for ISIC 2016, ISIC 2017 and ISIC 2018 benchmarked datasets.