SNE-UNET: An Efficient and Accurate Model for Skin Lesion Segmentation
Anam Khurshid, Department of Software Engineering, University of Engineering and Technology, Taxila, Pakistan.
Ali Javed, Department of Software Engineering, University of Engineering and Technology, Taxila, Pakistan.
Corresponding Author:
Ali Javed (ali.javed@uettaxila.edu.pk)
Abstract:
Melanoma is the deadest type of skin cancer with the highest incidence rate affecting millions of people worldwide and causing almost 75% of deaths. Early detection of melanoma can significantly improve the survival rate by up to 99%. Many segmentation approaches based on traditional machine learning and deep learning techniques are being used to segment skin lesions. However, skin lesion segmentation is a challenging task due to intrinsic visual complexity and the presence of various artifacts in dermoscopic images such as human blood vessels, human hair, clinical ruler, etc. Deep learning-based approaches have shown promising performance in segmenting skin lesions precisely. U-NET is one of the most popular deep-learning models used to segment medical images. The original UNET faces the issue of vanishing gradient and slower network learning by using the RELU activation function. It also faces the issue of generalization of the network and overfitting due to Batch Normalization. It performs simple normalization that ignores key features at the middle and end layers. Therefore, we proposed an enhanced version of U-NET for robust and precise segmentation of skin lesions. We integrated Switchable Normalization and ELU activation to overcome the above mention limitations. The use of SN and ELU not only overcomes the issues of the original UNET but also boosts the performance in terms of accurate segmentation of the skin lesions. Experimental results conducted on ISIC 2018 and ISIC 2017 datasets indicate that our proposed approach performs better than the existing state-of-the-art methods in terms of accuracy.
Keywords:
Deep Learning; Melanoma; Medical Image Segmentation; SNE-UNET