Deep learning wavefront retrieval in a double-pass setup

Wavefront retrieval is well established in single-pass systems such as telescopes but remains largely unexplored in double-pass configurations like the human eye. Here, we develop a deep learning wavefront retrieval method and an accompanying experimental double-pass setup to validate it. In double-pass imaging, light is projected onto the retina through the eye’s optics and the resulting retinal image is observed again through the same optics, requiring novel approaches for wavefront characterization. We simulated extended-source spread functions in double-pass configuration using random combinations of the first 12 Zernike coefficients after piston and tilt, sampled from ranges characteristic of healthy human eyes. The extended source was a uniform circular disk subtending 0.5 ◦ at the retina, producing corresponding spread functions which we refer to as disk spread functions (DSFs). For each wavefront, we computed a through-focus DSF volume consisting of five images spanning −1D to +1D in 0.5D steps. A DenseNet-based model predicted Zernike coefficients from the input DSF volumes. Using the known forward physical model of light propagation, we computed the wavefront from these coefficients and reconstructed the DSF volume, enabling self-supervised training by minimizing the Zero-Mean Normalized CrossCorrelation (ZMNCC) loss between input and reconstructed volumes. Experimental validation was performed using a spatial light modulator (SLM) to introduce known aberrations in a model eye. Across 100 experimental DSF volumes, the model achieved a mean ZMNCC loss of 0.040 ± 0.006 on normalized DSF volume and a mean RMSE of 0.067 ± 0.018µm for Zernike coefficient prediction. This work demonstrates, for the first time, deep learning wavefront retrieval in a double-pass system and establishes a foundation for in-vivo applications in human eyes.