![]() In addition, most of these previous studies concentrated on the textures and landmarks of retinal images by maintaining the vascular structures. Nevertheless, for most of these studies, the spatial resolutions of the synthesized images were poor, and the small numbers of trained images resulted in blurred and low-contrast optic discs, vessels, and retina boundaries. ![]() Several studies on synthesizing retinal images via GAN have already been conducted 14– 20. The discriminator feedbacks its answer (i.e., degree of realism defined with similarity metric) to the generator via backpropagation, to enable the generator to modify its weight and synthesize more realistic images. GAN is composed of two deep-learning networks: a generator, which attempts to synthesize realistic images, and a discriminator, which learns image characteristics from training data and attempts to discriminate whether the synthesized images are real or synthesized. Generative adversarial Network (GAN) is a powerful deep learning algorithm that synthesizes high-resolution images in an unsupervised manner 13. On the other hand, the AI-based generation model demonstrated the synthesized image has efficacy for balancing imbalanced datasets for diagnostic model development with avoidance from patient privacy issues 12. Recently, increasing demand for medical care and advances in medical technology gradually enable to collect images covering those features 11. The best approach to learn the various features of retinal images is to collect big data that are sufficiently large to cover all disease patterns and patient-dependent variations, including distributions of race, ethnicity, age, and sex 9, 10. Recently, retinal images are also being actively studied for the diagnosis of various ocular diseases such as diabetic retinopathy (DR) 3, 4, age-related macular disease (AMD) 5, 6, and glaucoma 7, 8.īecause of the various features in retinal images that need to be learned, developing a robust interpretation tool for retinal diseases remains challenging. AI-based CAD systems show significant potential to increase the accuracy of diagnoses and enable appropriate treatment plans based on the predicted disease progression. Owing to the rapid development of computer vision with the help of deep learning, many researchers have attempted to develop artificial-intelligence (AI)-based computer-aided diagnosis (CAD) systems for medical fields 1, 2. Our study shows the realistic retina images were successfully generated with insignificant differences between the real and synthesized images and shows great potential for practical applications. The classification performance after augmenting synthesized images outperforms every ratio of imbalanced real datasets. Vessel amounts and average SNR comparisons show 0.43% and 1.5% difference between real and synthesized images. Here, sensitivity represents correctness to find real images among real datasets. ![]() The efficacy of synthesized images was verified by deep learning-based classification performance. ![]() The synthesized images were validated via image Turing tests, qualitative analysis by retinal specialists, and quantitative analyses on amounts and signal-to-noise ratios of vessels. This study aims to synthesize retinal images indistinguishable from real images and evaluate the efficacy of the synthesized images having a specific disease for augmenting class imbalanced datasets. However, training a generative adversarial network (GAN) for image synthesis remains challenging because of the large amounts of data required for training various kinds of image features. Realistic image synthesis based on deep learning is an invaluable technique for developing high-performance computer aided diagnosis systems while protecting patient privacy.
0 Comments
Leave a Reply.AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |