Retinal pigment epithelial (RPE) cells play an integral role in maintaining visual function and retinal health. Adaptive optics-optical coherence tomography (AO-OCT) has enabled the in vivo visualization of the hexagonal structure of RPE at cellular scale resolution. However, it is difficult to clearly visualize the RPE mosaic in single AO-OCT volumes due to inherent speckle noise, which can be overcome by averaging a large number of AO-OCT volumes acquired at sufficiently spaced time intervals to allow speckle decorrelation for improved cell contrast. Here, we present a deep learning based siamese discriminator (DS) generative adversarial network (GAN) to recover the hexagonal RPE mosaic from only a single AO-OCT volume. The DS provides additional feedback to the generator that resulted in improved visualization of RPE morphology compared to traditional GAN. Experimental results from five healthy subjects suggest that the RPE images generated using DS-GAN were comparable to ground truth images obtained by averaging multiple AO-OCT volumes. Quantitative comparison of cell-to-cell spacing, density, and image quality assessment metrics further confirmed the accuracy of recovered RPE mosaics relative to ground truth. These results establish a potential strategy in which deep learning can be leveraged to eliminate the need for volume averaging and speckle decorrelation for more efficient RPE imaging.
|