Towards Exploring Deep Features for Efficient Melanoma Diagnosis in Dermoscopic Images

Melanoma is the most dangerous and lethal kind of skin cancer, and its high mortality rate is a major public health concern worldwide. Hence, detecting melanoma at the early stages is essential to ensure a high survival rate. In recent years, convolution neural networks (CNNs) based automated methods have gained attractions for early melanoma diagnosis. However, deriving meaningful features from dermoscopic images is challenging due to several reasons, such as inter-class consistency, intra-class variability, and insufficient training data. To tackle this problem, we propose an automated method for melanoma detection from dermoscopic images using high-level features derived from an effective CNN architecture. The proposed method first preprocesses the images using boundary localization and cropping to generate images containing only essential parts, i.e., lesion regions. Then, the preprocessed images are input to a pre-trained CNN architecture to extract the high-level representations, followed by a multi-layer perceptron (MLP) for the classification of skin images. The PH
²
benchmark dataset is used to evaluate our proposed method. The experimental findings show that the method (DenseNet201+MLP) achieves promising performance compared to other existing approaches.