Abstract
Mobile teledermoscopy is an area that enables patients to get an early detection for suspicious lesions using their mobile phones and hence creating a patient-centric health management. Due to the inherent variability in the appearance of skin lesions, detection of skin cancer from dermoscopic images is a difficult task even for medical experts. Recent advancements in image processing using Deep Convolutional Neural Networks (CNN) have led numerous researchers to use them for skin lesion classification which concluded that CNN performed on par with expert dermatologists. In this paper, we used a dataset of 48,373 dermoscopic images collected from three different archives labelled and validated by expert dermatologists. In our work, we manually trained a resource constrained CNN model called MobileNetV2 using transfer learning for binary classification of skin lesions into benign or malignant classes. Using batch size of 32, the trained model resulted in an overall accuracy of 91.33%. The trained model was then used to develop a mobile application for iOS devices using the Core ML library. The mobile application was then tested on a new dataset to assess its performance on an unseen library of images.