Employing texture loss to denoise OCT images using generative adversarial networks.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Speckle noise in optical coherence tomography (OCT) images obscures pathological features and hinders accurate segmentation.
• Previous deep learning denoising methods (CNNs) using pixel-wise losses often resulted in blurred images.
• Perceptual losses improved OCT denoising but textural fidelity remained a challenge.

Purpose of the Study:

• To develop a novel generative adversarial network (GAN) based method for denoising OCT images.
• To enhance the textural similarity of denoised OCT images to gold-standard averaged OCT images.
• To evaluate the performance of different GAN architectures and loss functions for OCT image reconstruction.

Main Methods:

• A conditional GAN (cGAN) framework was designed for OCT image reconstruction.
• PatchGAN discriminator was employed as a texture loss to improve image quality.
• UNet and ResNet were compared as generators, and PatchGAN was compared with Wasserstein GAN.

Main Results:

• The combination of PatchGAN and UNet achieved superior denoising performance (PSNR = 32.50) compared to SiameseGAN (PSNR = 31.02).
• Qualitative assessment by masked clinical ophthalmologists favored the reconstructed OCT images using PatchGAN texture loss.
• The proposed method demonstrated improved image quality and preservation of pathological features.

Conclusions:

• The novel GAN-based method with texture loss effectively denoises OCT images, preserving essential details.
• This approach offers significant advantages over existing denoising techniques for OCT imaging.
• The method shows promise for improving clinical interpretation and diagnosis of retinal diseases.