A hybrid deep learning approach for skin cancer diagnosis using subband fusion of 3D wavelets

An automated system for skin cancer diagnosis is needed for early diagnosis to reduce the mortality rate of skin cancer. Noninvasive clinical routines are essential for diagnosis, but they are largely subjective. In this study, a hybrid deep learning (HDL) approach that uses subband fusion of 3D wavelets is proposed. It is a noninvasive and objective method for inspecting skin images. In the first stage of the HDL approach, simple median filtering is used to remove unwanted information such as hair and noise. In the second stage, the 3D wavelet transform is applied to obtain textural information from the dermoscopic image via a subband fusion approach. In the final stage, multiclass classification is performed by the HDL approach using the fused subband. The performance results of the HDL approach on PH 2 database images indicate that it can discriminate normal, benign, and malignant skin images effectively with 99.33% average accuracy and more than 90% sensitivity and specificity. This study confirms the observation that the HDL approach can realize improved classification results.