Departmental Bulletin Paper 転移学習を用いたメラノーマ自動診断システム

吉田, 琢也

In recent years, a new machine learning schema called deep learning has gained many promising achievements in a wide range of industries and thus attracted more than just researchers. A convolutional neural network (CNN) is a principal aspect of deep learning techniques specialized for machine learningand/or computer vision. Through the influence of these promising achievements, replacement of existing methods has been progressed rapidly. The automated melanoma diagnosis system is no exception, and anautomated melanoma diagnosis system using CNN has already been proposed. However, CNN’s learning requires a large amount of training data, and since melanoma images can only be obtained from limitedorganizations such as medical institutions, it is difficult to collect a large amount of data. In this study, weovercome these problems using transfer learning and applying effective pre-processing to the data set. Themelanoma classifiers constructed using the proposed method achieved AUC: 0.873 on the ROC curve.Key Words : melanoma, skin lesion, dermoscopy, convolutional neural network, transfer learning

Number of accesses :  

Other information