Fig. 2: Residual in residual transformer generative adversarial network (RRTGAN) restores the pixel resolution of sparsely sampled images to match dense sampling. | npj Artificial Intelligence

Fig. 2: Residual in residual transformer generative adversarial network (RRTGAN) restores the pixel resolution of sparsely sampled images to match dense sampling.

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

Fig. 2: Residual in residual transformer generative adversarial network (RRTGAN) restores the pixel resolution of sparsely sampled images to match dense sampling.

a, f Sparsely sampled images of the cone photoreceptors from two participants (P1 and P2) having a pixelated appearance due to reduced pixel resolution. Pixel resolution of the cones enhanced by AI using (b, g) enhanced super-resolution generative adversarial network (ESRGAN), c, h SwinIR, and d, i RRTGAN (ours). e, j Ground truth densely sampled images for visual comparison. The bright dots in the images represent the individual cone cells. The yellow arrows show cells that are more distinguishable in RRTGAN compared to the sparsely sampled, ESRGAN, and SwinIR images. Scale bar: 50 µm.

Back to article page