Artificial intelligence-assisted retinal imaging enables dense pixel sampling from sparse measurements

High resolution in vivo adaptive optics (AO) imaging has facilitated cellular level assessment of microscopic cone photoreceptors. However, the necessity for dense pixel sampling for good pixel resolution imposes a tradeoff with acquisition speed, leading to motion artifacts and extensive data generation. We introduce an artificial intelligence (AI) assisted imaging framework utilizing residual in residual transformer generative adversarial network (RRTGAN), an AI method that works alongside AO imaging to restore the pixel resolution of sparsely sampled images, circumventing the need for dense sampling. Our results show that RRTGAN can enable data-efficient imaging, restoring high-quality images from just one-fourth of the data and closely matching ground truth images. Cone spacing estimates across four participants aligned well with histology at various retinal locations. These results demonstrate AI assisted imaging's potential to overcome pixel sampling and imaging speed tradeoff, an important step toward improving the efficiency of routine AO imaging in the clinic.