Retinal Optical Coherence Tomography Image Classification With Label Smoothing Generative Adversarial Network Publisher

Summary: A study found a new AI model improves eye disease detection despite limited data, boosting diagnosis. #EyeHealth #ArtificialIntelligence

He X¹ ; Fang L¹ ; Rabbani H² ; Chen X³ ; Liu Z³
Source: Neurocomputing Published:2020

Abstract

In this paper, we propose a label smoothing generative adversarial network (LSGAN) for optical coherence tomography (OCT) image classification to identify drusen, i.e., the early stage of age-related macular degeneration (AMD), choroidal neovascularization (CNV), diabetic macular edema (DME) and normal OCT images. The LSGAN can expand the dataset to address the issue of overfitting when only limited OCT training samples are available. Specifically, our LSGAN consists of three components: generator, discriminator, and classifier. The generator generates synthetic unlabeled images that are similar to the real OCT images, while the discriminator distinguishes whether the synthetic images are real or generated. To train the classifier with both real labeled images and synthetic unlabeled images, we design artificial pseudo labels as label smoothing for the synthetic unlabeled images. Thus, the mixing of the synthetic images and real images can be used as training data to improve the classification performance. Experimental results on two real OCT datasets demonstrate the superiority of our LSGAN method over several well-known classifiers, especially under the condition of limited training data. © 2020 Elsevier B.V.