Automatic diagnosis of melanoma from dermoscopic image using real-time object detection

Among all types of skin cancer, a melanoma is the deadliest. Melanoma is typically a small, usually black or brown colored mole, which can develop anywhere on the skin. Detecting melanoma in its earliest stages is among the most important factors in improving the outcome of a melanoma diagnosis. In this paper, a real-time object detection technique is used to automatically detect melanoma in dermoscopic images. For detecting melanoma in real-time a state-of-the art detection model named YOLOv2 (You Only Look Once: version 2) is used. YOLOv2 uses a single neural network to the full image, enabling real-time performance. It is capable of processing images at 40-60 fps using a Titan X GPU. Our proposed model predicts the diagnosis of a mole with an accuracy of 86.00%, sensitivity = 86.35% and specificity = 85.90%. In addition, the proposed method is shown to be invariant to the presence of hair in the image.