ASLERNet: an automated skin lesion extraction and refinement deep learning system

In recent years, melanoma skin cancer has become a major cause of death. The survival percentage of this chronic condition may be raised by prompt diagnosis of melanoma skin cancer lesions. However, it can be difficult to detect these lesions, particularly when there are challenges including blood vessels within the lesions, clinical artifacts, color variation, and poor contrast. Various traditional and CNN models have been developed in the past decades for accurate identification of skin lesions that help in the further diagnosis process. However, the capacity to represent complex features, their sensitivity to noise, and the requirement for manually extracting the lesions are the limitations of traditional approaches. Also, the pre-trained networks suffer from domain mismatch, are computationally expensive, and lack interpretability, even if they have the benefit of utilizing information from large datasets. So, to overcome these issues, we have developed a feature extraction-refinement[1]based deep framework for skin lesion extraction ASLERNet. The developed ASLERNet comprises two different networks a feature extraction network (FEN) and a feature refinement network (FRN). The designed FEN is capable of retaining structural and contextual information from the dermoscopic image effectively. Further, these extracted details are enhanced with reduced noise using the developed FRN. The designed ASLERNet network is trained using the images of ISIC 2016 data and tested by considering the images from ISIC 2016 and also for unseen images taken from ISIC 2017 and ISIC 2018 datasets. Without any pre-processing technique, the network achieves an accuracy of 96.1% on ISIC 2016. Also, for unseen images considered from ISIC 2017 and ISIC 2018 datasets, it gives its effectiveness by providing an accuracy of 96.4% and 96.2% respectively. The outcomes acquired from the proposed ASLERNet are compared with the 20 existing recent approaches for validation of the network.