Large Aberration Wavefront Reconstruction for Human Eyes Based on Deep Learning and Shack-Hartmann Wavefront Sensors

Objective: Human eyes exhibit various high-order aberrations with significant individual variability, directly deteriorating retinal image quality and visual system performance. Keratoconus, a progressive corneal pathology, causes large-amplitude irregular aberrations that significantly compromise patients' visual functions. While Shack-Hartmann wavefront sensor (SHWFS) has become the mainstream technology for measuring human eye aberrations due to its non-contact and real-time capabilities, it faces significant challenges with large-amplitude aberrations. The limited dynamic range of SHWFS—primarily due to sub-aperture displacement thresholds—introduces error in wavefront reconstruction when the wavefront aberration is significant. Although existing deep learning approaches have been applied to SHWFS wavefront reconstruction, they fail to consider the fundamental differences between SHWFS spot images and traditional computer vision datasets. These approaches do not account for the unique spatial characteristics of SHWFS data. This study aims to develop an innovative approach that effectively utilizes SHWFS spatial structural characteristics to improve reconstruction accuracy for large aberrations and extend the dynamic range of SHWFS. Methods: A spatial prior attention residual network (SPResNet) incorporating SHWFS spatial structural characteristics was proposed. The network employed ResNet-50 (a residual network) as the backbone architecture with four residual layers. The core innovation was the prior aperture attention (PAA) module, which leveraged the spatial arrangement of microlens arrays to guide network attention toward physically meaningful regions of spot images. For dataset construction, we established simulation datasets based on measured human eye aberration statistics. Using real aberration data, we calculated mean vectors and covariance matrices for normal and highly aberrated eyes' 36-order Zernike coefficients, constructing multidimensional Gaussian distributions. Through random sampling from these distributions, we generated Zernike coefficient labels with a 1:1 ratio of normal to aberrated samples, and subsequently simulated the corresponding spot images using SHWFS. The PAA module performed channel-wise pooling and concatenation with normalized coordinate matrices. The aperture bias component divided feature maps into N × N sub-aperture regions based on grid parameters, assigning independent learnable bias parameters. Through convolutions and sigmoid activation, attention weight matrices corresponding to microlens structures were generated. SPResNet employed depthwise separable convolutions for computational efficiency while maintaining feature extraction capabilities. Training utilized mean squared error loss with the AdamW optimizer and cosine annealing scheduling. Results and Discussions: The dataset contained 14000 samples with 10000 for training and 4000 for testing, covering a range from normal to highly aberrated eyes with peak-to-valley (PV) mean value of 20.5854λ and root mean square (RMS) mean value of 3.8948λ. SPResNet achieved optimal performance with a validation loss of 0.00000757, representing reductions of 51.9 and 64.6 compared to ConvNeXt and ResNet-50, respectively. For wavefront reconstruction accuracy, SPResNet achieved a wavefront reconstruction root mean square error (RMSE) of 0.0137λ, outperforming ConvNeXt and ResNet-50 by 28.6 and 43.2, respectively. Distribution across the 4000 test samples showed consistent performance. Evaluation across the 4000-sample test set demonstrated consistent superiority in wavefront reconstruction. For a high-aberration sample (RMS of 9.9940λ), the wavefront reconstruction RMSE of SPResNet was merely 0.0231λ, achieving 68.1 and 58.2 reductions compared to ResNet and ConvNeXt, respectively. Statistical analysis of 623 high-aberration samples (RMS of >8.5) revealed SPResNet's mean wavefront reconstruction RMSE of 0.0151λ, outperforming ConvNeXt (0.0205λ) and ResNet (0.0257λ) by 26.3 and 41.2, respectively. Ablation experiments demonstrated the effectiveness of PAA, showing 43.2 and 38.6 lower wavefront reconstruction RMSE compared to the basic ResNet-50 and ResNet-50 with traditional spatial attention, respectively. Transfer learning experiments highlighted SPResNet's excellent generalization capability across diverse microlens configurations (16 × 16, 24 × 24, 32 × 32, 40 × 40), achieving 39.5, 52.0, 30.3, and 58.4 accuracy improvements with 50% less training time. An improved transfer strategy adjusting Grid parameters and unfreezing PAA modules achieved additional accuracy improvements of 78.6, 59.6, 76.8, and 66.6. Conclusions: This study develops SPResNet, a novel architecture that overcomes the dynamic range limitation of SHWFS in reconstructing large-amplitude aberrations. The innovative PAA module utilizes the spatial structure of SHWFS, significantly enhancing feature extraction capabilities and wavefront reconstruction accuracy. The method achieves an RMSE of 0.0137λ in wavefront reconstruction, demonstrating 28.6 and 43.2% lower error than ConvNeXt and ResNet-50, respectively. Comprehensive studies validate PAA effectiveness over traditional attention mechanisms, while transfer learning demonstrates excellent generalization across various configurations. The improved transfer strategy achieves consistently stable performance across diverse SHWFS configurations. This research provides a robust method for high-precision large aberration reconstruction, offering significant application value in fundus imaging, ophthalmic diagnosis, and visual health management. The integration of optical physics priors with deep learning frameworks opens new possibilities for promoting adaptive optics in clinical ophthalmology.