Cyst classification using bridge transfer learning with sparse labelled medical images

Abstract

Medical imaging has become an important technique to visualize the internal body for clinical analysis and medical intervention. A broad study of diseases diagnosis using medical imaging has been conducted in various medical fields. As an automatic diseases diagnosis of medical image, many researchers have suggested learning-based algorithms. Since deep learning approaches can achieve impressive performance on most computer vision fields, many methods applying convolutional neural network (CNN) have also been proposed in medical imaging. However, medical images are insufficient due to the scarcity of disease and the protection of personal information. It is also difficult to obtain images with labels because the professional annotation is expensive. Although many studies utilize transfer learning to deal with these problems, the domain gap between natural images used for the source database and medical images has been regarded as an another problem. In this thesis, we propose cyst classification system with bridge transfer learning with sparse labelled medical images. For bridge transfer learning, we construct a bridge database with the images acquired from the same medical domain but different purposes with the target database. This bridge database is used to learn the projection function once trained by the source database and apply the function to the target database for feature representation. For the experiments of cyst classification using dental panoramic X-ray database, we utilize JSRT database (containing chest X-ray images) as the bridge database. The results show that the proposed method using bridge transfer learning helps to achieve higher cyst classification rate by reducing the domain difference between the source database and the target database. Our approach casts a light on medical imaging commonly deemed as a barrier which could not find the breakthrough in sparse labelled images.