A customized deep learning framework for skin lesion classification using dermoscopic images

Abstract

Automated analysis of skin lesions in dermoscopy images has gained much attention due to its medical importance in the early detection of melanoma. Detection of lesions has become a challenge due to the strong visual similarity between benign and malignant skin lesions. In this research, a customized deep convolutional neural network (CNN) architecture has been designed to discriminate between benign and malignant lesions. The model is designed carefully with lesser convolution layers, fewer filters, and parameters to achieve better classification performance compared to pretrained VGG16, ResNet50, InceptionV3 models and, ensures state‐of‐the‐art performance. The proposed model is composed of nine trainable layers: eight convolution layers and one fully connected layer. The suggested framework is extensively evaluated on the benchmark ISIC 2016 challenge dataset. The effect of different input transformations over the dataset has been studied. For fair comparison, standard deep learning models such as VGG16, ResNet50, and InceptionV3 have been used for lesion classification using transfer learning approach. The memory requirement of the proposed model is reduced by 388, 68, and 63 times and FLOPs needed are lowered by 95%, 85%, and 84% compared to VGG16‐TrL, ResNet50‐TrL, and InceptionV3‐TrL, respectively. Results show that class balancing with external images improves classification performance.