Skin lesion detection based on deep neural networks

The skin lesion classification is a difficult challenge as a result of the unique features of the skin lesions and their range of forms, especially because of the identification problem of early melanoma. Thus, to resolve the limitation in the previous approach, a neural network-based approach was proposed to distinguish dermoscopic images enclosing two skin lesion types. In this initiative, the proposed method consists of four phases: (1) processing of the initial image; (2) segmentation of skin lesions; (3) extraction of features; and (4) DNN-based classification. Image processing is the first step to remove any unnecessary noise using a median filter, Then, by using Otsu's image segmentation; the particular areas of skin lesions are segmented. The skin lesion Features are extracted further in the third stage. Features that are extracted using GLCM, 2D DWT, and RGB color model. The fourth stage is to categorize the different forms of skin disease based on the backpropagation deep neural network with the Levenberg Marquardt (LM) generalization method to minimize mean square error. The model was trained using ISIC 2017 dataset to evaluate the proposed deep learning model. We have applied DNN on and achieved the highest accuracy i.e. 84.45% compared to the other states of art machine learning Classifiers. •In order to detect skin lesions for cancer imaging, deep neural networks are crucial.•Dermoscopic image classification can be performed using a deep neural network model.•ISIC dataset was used to train the model.•A deep learning model is used to achieve an accuracy of 84.45%.