Wavelet‐based deep learning for skin lesion classification. Issue 4 (19th February 2020)

• Record Type: Journal Article
• Authors: Serte, Sertan; Demirel, Hasan
• Is Part Of: IET image processing. Volume 14:Issue 4(2020)
• Page Start: 720
• Page End: 726
• Publication Date: 2020-02-19
• Subjects: image classification -- telemedicine -- medical image processing -- wavelet transforms -- learning (artificial intelligence) -- cancer -- skin -- biomedical optical imaging; wavelet‐based deep learning -- skin lesion classification -- skin lesions -- malignant forms -- benign forms -- benign skin lesion types -- malignant types -- skin cancer -- malignant melanoma -- seborrhoeic keratosis lesions -- skin images -- vertical wavelet coefficients -- deep learning models -- approximate coefficients -- sequential wavelet transformation -- approximation coefficients -- transfer learning‐based ResNet‐18 -- model images -- skin lesion detection; Image processing -- Periodicals; 621.36705
• DOI: 10.1049/iet-ipr.2019.0553
• Languages: English
• ISSNs: 1751-9659

Abstract

Abstract : Skin lesions can be in malignant or benign forms. Benign skin lesion types are not deadly; however, malignant types of skin lesions can be fatal. Lethal forms are known as skin cancer. These types require urgent clinical treatment. Fast detection and diagnosis of malignant types of skin lesions might prevent life‐threatening scenarios. This work presents two methods for the automatic classification of malignant melanoma and seborrhoeic keratosis lesions. The first method builds on modelling skin images together with wavelet coefficients. Approximate, horizontal, and vertical wavelet coefficients are obtained using the wavelet transform, and then deep learning (DL) models are generated for each of the representations and skin images. The second method builds on modelling skin images together with three approximate coefficients. This method utilises a sequential wavelet transformation to produce approximation coefficients. Then DL models are generated for each of the representations and skin images. Transfer learning‐based ResNet‐18 and ResNet‐50 DL models provide model images and wavelet coefficients. Then skin lesion detection is achieved by fusing model output probabilities. Both proposed models outperform the methods only based on image data and other previously proposed methods.