Hierarchical Features Integration and Attention Iteration Network for Juvenile Refractive Power Prediction

Refraction power has been accredited as one of the significant indicators for the myopia detection in clinical medical practice. Standard refraction power acquirement technique based on cycloplegic autorefraction needs to induce with specific medicine lotions, which may cause side-effects and sequelae for juvenile students. Besides, several fundus lesions and ocular disorders will degenerate the performance of the objective measurement of the refraction power due to equipment limitations. To tackle these problems, we firstly propose a novel hierarchical features integration method and an attention iteration network to automatically obtain the refractive power by reasoning from relevant biomarkers. In our method, hierarchical features integration is used to generate ensembled features of different levels. Then, an end-to-end deep neural network is designed to encode the feature map in parallel and exploit an inter-scale attentive parallel module to enhance the representation through an up-bottom fusion path. The experiment results have demonstrated that the proposed approach is superior to other baselines in the refraction power prediction task, which could further be clinically deployed to assist the ophthalmologists and optometric physicians to infer the related ocular disease progression.