RER-Net: Refractive Error Regression Network for Multi-Angle Eccentric Photorefraction Pupil Images

Refractive error can lead to permanent vision impairment and has become a critical factor to various ocular diseases. Multi-angle near-infrared (NIR) eccentric photorefraction pupil images provide rich refractive information, offering a feasible and convenient solution for large-scale vision screening, which is an important way for early detection of refractive error. In this paper, we propose a refractive error regression network (RER-Net) which can realize the effective regression of spherical component and cylindrical component according to the image characteristics, including a semantic-enhanced noise reduction (SENR) module to improve perceptive ability and a coordinated fusion (CF) module to extract multi-scale shallow features. Our study construct a high-quality ocular visual acuity image dataset comprising 2241 pupil samples and evaluate the proposed RER-Net on this dataset, achieving a mean absolute error (MAE) of 0.4223D and 0.3241D for spherical and cylindrical component, respectively. Moreover, the accuracy of spherical and cylindrical component within the ±0.5D range reach 80.18% and 88.64%, respectively.