International Journal of Machine Learning and Computing, Vol. 8, No. 1, February 2018 61 doi: 10.18178/ijmlc.2018.8.1.664 Abstract—Melanoma is a common type of cancer that affects a significant number of people. Recently, deep learning techniques have been shown to be highly accurate in classifying images in various fields. This study uses deep learning to automatically detect melanomas in dermoscopy images. First, we preprocess the images to remove unwanted artifacts, such as hair, and then automatically segment the skin lesion. We then classify the images using a convolutional neural network. To evaluate its effectiveness, we test this classifier using both preprocessed and unprocessed images from the PH 2 dataset. The results show an outstanding performance in terms of sensitivity, specificity, and accuracy. In particular, our approach was 93% accurate in identifying the presence or absence of melanoma, with sensitivities and specificities in the 86%–94% range. Index Terms—Deep learning, dermoscopy image, image processing, melanoma detection I. INTRODUCTION Skin cancer is one of the most common malignancy types. In the US alone, over 5 million cases have been diagnosed every year [1]. Melanoma is one of the most common and fatal types of skin cancer and involves the unrestrained growth of pigment-producing cells. In the US, it is responsible for 4% of all cancer deaths and 6 out of every 7 skin cancer-related deaths [2]. It is estimated that 9,730 people will die from melanoma in 2017 [3]. Melanoma can affect anyone although Caucasians are more likely to suffer from it than other races. However, it is less obvious in people with dark skin, so it is often diagnosed at a later stage when the disease is more advanced. This is because many people have the mistaken impression that people of color cannot get skin cancer [4]. It caused the most cancer-related deaths globally in 2016, with an estimate of 76,380 people having invasive melanomas, of which approximately 46,870 were male and 29,510 were female [5]. Melanoma is highly treatable if it is detected early, but advanced melanoma can spread to the lymph nodes and other organs, which can be fatal. Medical experts and professional equipment are crucial for early and accurate melanoma detection. In contrast, more limited access to such expert opinions makes it a challenge to provide adequate levels of care to the populations that are at the risk of this disease. Usually, patients initially see a skin abnormality. Then medical experts use dermoscopy for diagnosis. This is a Manuscript received July 26, 2001; revised September 30, 2017. The authors are with De La Salle University, Philippines (e-mail: julie_salido@dlsu.edu.ph, conrado.ruiz@dlsu.edu.ph). high-resolution skin-imaging process that reduces skin surface reflections, allowing doctors to examine the deeper underlying structures. They are used to non-invasively evaluate [6] in vivo the colors and microstructures of the epidermis, dermoepidermal junction, and papillary dermis. This has opened up a new avenue for examining pigmented skin lesions and especially identifying the early stages of melanoma [7], [8]. Using this approach, specially trained medical experts have demonstrated diagnostic accuracies as high as 75%–84% [9]. However, the diagnostic performance drops significantly if the doctors have not been adequately trained [10], [11]. To address the issues caused by limited access to specialists, especially in developing countries, there has been considerable research focusing on developing automated image analysis systems that can detect skin diseases based on dermoscopy images. There have been several recent publications reviewing the different techniques used [9], [12] as well as dermoscopy papers developing diagnostic criteria for early melanoma detection [13]-[16]. However, these criteria still involve dermoscopy image characteristics that can only be assessed by dermatologists or medical specialists. In this paper, we proposed a method of detecting and removing hair from dermoscopy images and present a way of classifying skin lesions using deep learning. We then evaluate both the hair detection algorithm and the classifier using images from the PH 2 dataset [1]. The rest of the paper is organized as follows. In Section II, we discuss some related literature. In Section III, we present our proposed methods before implementing and evaluating them in Section IV and analyzing the results numerically. Finally, in Section V, we present our conclusions and plans for future work. II. RELATED WORK Deep learning techniques attempt to enable computers to learn from a large number of examples. Deep learning models automatically categorize input datasets, such as images, audio, or documents, directly. They can yield excellent and up-to-date classifications that can sometimes beat human assessment. Deep learning uses neural network architectures with several layers that are trained with large datasets, with the most popular type [17] being convolutional neural networks (CNNs). Fig. 1 shows a fully connected neural network, a characteristic of the CNN. In this study, we use preprocessed dermoscopy images as input and obtain the classification result (i.e., skin disease type) as output. CNNs have been shown to be very effective for a number Using Deep Learning to Detect Melanoma in Dermoscopy Images Julie Ann A. Salido and Conrado Ruiz Jr.