Mobile Application for Early Screening of Skin Cancer Using Dermoscopy Image Data Based on Convolutional Neural Network

The dermoscopy image produced by the dermatoscope represents the skin surface structure in more detail than clinical images because the imaging technique uses optical magnification and liquid immersion lighting or cross-polarization. Detection of skin cancer is a challenge for several reasons. Doctors find it difficult to detect the presence of cancer on the skin only with dermoscopy images. Sometimes further tests are needed to confirm the results of the diagnosis, such as a biopsy or pathology examination, that is relatively time-consuming. Mobile skin screening technology can be used as an early screening tool for the early detection of melanoma, allowing patients to share images of their lesions with their skin specialists quickly. In this study, a novel method for early skin cancer detection based on dermoscopy pictures and the convolution neural network (CNN) algorithm is proposed. Three convolution layers, two average pooling layers, and a batch normalization layer make up the five feature extraction layers of CNN’s proprietary design. The ISIC Archive, which has a collection of skin lesion photos, was the dataset utilized to create the model. Four thousand dermoscopy images in JPG format are used in this study and are divided into two classes: benign and malignant. The dataset is divided randomly into 80% training data and 20% test data. The experiments found that the model with the best performance resulted in a training data accuracy of 99.3% and a test data accuracy of 97.63%. It concludes that the model with batch normalization can distinguish the benign and malignant well. The model requires a small enough space to be easily accessible through mobile devices.