The novel deep learning method is a computer based system that differentiates objects by morphological parameters such as colors, shape and other abstract features. The augmented accuracy of deep convolutional neural networks and the accessibility of clinical pictures, have brought about the increasing use of this method for visual recognision of images in medicine. In melanoma, deep convolutional nets may be implicated for early identification of suspicious lesions, therefore may reduce melanoma mortality and improve survival. This study examines the capability of a computerized system to visually diagnose melanoma, and to compare these results to other diagnostic measures. To address these goals, we trained a deep convolutional neural network on the International Skin Imaging Collaboration (ISIC) Archive dataset of melanoma and benign lesion images. The system was than further fine- tuned. Using InceptionV3 network architecture, the system achieves high scores in diagnosing melanoma (sensitivity and accuracy values of 84.8% and 84.5%, respectively). These values are comparable to published accuracy of dermatologists. In conclusion, computerized systems may have an accuracy level similar to dermatologists in diagnosing melanoma. However, for the system to become applicable, larger cohorts of higher-quality, clinically annotated skin lesion imaging as well as dermatoscopic images should be collected.

Classification of pigmentary lesions: Can a Computerized System Replace the Dermatologist?