Fusion of deep and wavelet feature representation for improved melanoma classification

Melanoma is an acute and chronic skin disease, and if left untreated, it increases morbidity in the patient. The high degree of similarity between the images of different classes and the complex structure of lesion images makes automated skin lesion diagnosis very challenging. Dermoscopic images are widely used for the diagnosis of skin lesions. This study proposes a novel method by analyzing deep features and wavelet features. In this regard, a standard pre-trained ResNet50 model is used for extracting deep features. Lesion images are transformed to wavelet domain using lifting wavelet transform (LWT). The level-2 approximation component of LWT is taken as wavelet feature. Deep features and wavelet features are fused, and the neighborhood component analysis (NCA) algorithm is subsequently used to select a subset of fused features with reduced dimensions. The NCA-reduced feature set is classified by a multilayer perceptron (MLP). The suggested method is validated on the publicly available ISIC 2016 challenge dataset and PH2 dataset. An accuracy of 98%, auc of 99.62% on PH2 dataset while 71% accuracy, 80.61% auc on ISIC 2016 dataset is obtained. The experimental results are comparable to the existing state-of-the-art methods. This study demonstrates that the integration of LWT features improves discriminative information. Feature reduction by using NCA is able to provide robust details with reduced noise and redundancy.