Using Semi-supervised Transfer Learning for Classification of Solar Lentigo, Lentigo Maligna, and Lentigo Maligna Melanoma

Abstract

Skin cancer is the most frequent malignancy worldwide, with the number of new cases increasing yearly. Computer-aided diagnosis from skin images has recently become a critical procedure to detect early melanoma stages before becoming metastasis. This study intended to classify three stages of skin cancer: solar lentigo (SL), lentigo maligna (LM), and lentigo maligna melanoma (LMM) using transfer learning and semi-supervised transfer learning of a convolutional neural network algorithm based on VGG-16 and VGG-19. Skin images were obtained from various databases, including labeled and unlabeled data, and were preprocessed using hair removal software and a data balancing technique. The image data were then trained in ten experiments: supervised learning, supervised transfer learning, and semi-supervised transfer learning using VGG-16 and VGG-19 with and without augmentation. The results show that supervised learning gives an accuracy of 0.47. Based on VGG-16 and VGG19 which are comparable in performance, the accuracies increase to 0.72 and 0.72 for supervised transfer learning, and 0.92 and 0.98 for semi-supervised transfer learning, respectively. However, when applying augmentation, the accuracies decrease. Therefore, the use of semi-supervised transfer learning based on VGG-19 gives the best prediction in our study.