Automatic Retinal layers Segmentation from OCT images using Generative Adversarial Networks
Soheib Takhtardeshir1 *, Mohammad Mahdi Moradi1
- Engineering Department, Shahid Beheshti University, Tehran, Iran
Abstract: The Retinal at the back of the eye is essential for all vision. Each layer of cells in this tissue serves a specific purpose. For many retinal diseases, retinal layer thicknesses are measured from optical coherence tomography (OCT) images. We present an automatic Generative Adversarial Networks segmentation algorithm to segment intra-retinal layers in OCT images which suffer from low image contrast, irregularly shaped morphological features and local image noises.
Methods: This study used 700 two-dimensional OCT images as datasets. We scanned these subjects in the Imaging Center of Noor Eye Hospital. Aiming to extract boundaries of retinal layers in each slice, we used Pix2Pix network that uses 600 data for the train and 100 data for test after quality assurance and preprocessing. Pix2Pix is a general-purpose solution to image-to-image translation. This network uses a "U-Net"-based architecture for generator and a convolutional "PatchGAN" classifier which only penalizes structure at the scale of image patches.
Results: After implementations, the algorithm made a prediction for segmentation of the OCT images. The two well-known evaluation quantities in image processing concepts are "PSNR," and "SSIM". We measured the "PSNR" and "SSIM" average of all the test samples which is shown in Table.1. Besides, one sample of network output is represented in figure 1.
Conclusion: In conclusion, a model based on Pix2Pix for segmenting individual retinal layers is proposed. The proposed algorithm was demonstrated to achieve a high speed, high resolution and accurate intra-retinal segmentation on low contrast retinal OCT images. Also reduce human error as much as possible. We can improve the capability of this algorithm by increasing more input samples in the training step.
