Improving melanoma classification using transfer learning based wavelet features

Abstract

The most deadly and serious type of skin cancer is melanoma, and due to its high death rate, it is a significant threat to global health. Therefore, identifying melanoma is necessary to guarantee an elevated survival rate. Automated techniques based on convolutional neural network (CNNs) have become more popular recently for the early diagnosis of melanoma. However, there are several issues that make it difficult to extract useful features from dermoscopic images, including insufficient training data, intra-class variability, and inter-class consistency. In order to address this issue, we provide an automated technique that uses highlevel features obtained from the efficient ResNet50 CNN architecture to diagnose melanoma from dermoscopic images. The suggested approach begins with preprocessing, which is lesion area extraction, followed by resizing to fit the lesion image to the deep learning model. The resized images are converted to the wavelet domain using the lifting wavelet transform (LWT). A support vector machine (SVM) is used to classify skin images after preprocessed and LWT-transformed images are fed into the ResNet50 pre-trained CNN architecture to retrieve the high-level representations. We use the International Skin Imaging Collaboration 2016 (ISIC 2016) benchmark dataset to evaluate the proposed method. The experimental findings show that the inclusion of wavelet features provides promising performance.