Retinal OCT Image Segmentation Using Semi-supervised Learning with Generative Adversarial Networks

Abstract

Optical coherence tomography (OCT) is a three-dimensional imaging technique with micrometer resolution by irradiating a laser to optical scattering media such as tissues of living organisms, and can acquire images in depth direction. In particular, it is mainly used in the medical field. This is because, unlike the radiation-based imaging technique or the method of injecting a contrast agent and imaging it, it is a non-invasive method that does not affect the patient and there is little preparation time in imaging. In ophthalmology, OCT can image retinal layers to predict prognosis of vision and other eye diseases. Also, by continuously photographing the OCT, blood flow in the retina may be measured by imaging a fluid flowing in a certain area of ​​the retina, which is associated with various diseases such as diabetes. But analyzing these images is time-consuming and there are differences in the results among doctors. In this paper, we propose 1) CAM generation using Attention-based multiple instance learning (AMIL). After dividing the images into patch units, we design a model that learns the importance of the patches that contribute to the classification, and generate CAMs of the patch units using the trained parameters to calculate the contributions. The CAM thus obtained shows the shape of the object better than the conventional CAM through the whole image-based learning and is similar to the actual label, so it is used when training the segmentation model with a few labelled data. And 2) based on the Generative Adversarial Network (GAN) structure, which learns the generator and discriminator repeatedly as a method for learning the zoning model using a few labelled data and the AMIL-based CAM generated by the technique of 1). We propose a segmentation model. The proposed method was applied to the OCT image provided by Yeungnam University and the OCT dataset published in kaggle. Thus, the proposed weak label generation method showed better performance than the existing CAM method or AMIL method. In addition, the proposed GAN-based method shows that the data using a small amount of accurate label and the weak label created by the proposed method improves the performance by about 2% compared to the learning using only a small amount of accurate label.