Prediction of Spherical Equivalent with Vanilla ResNet

Huayu Li, Wenhui Zhu, Xiwen Chen, Yalin Wang

Abstract: Recently, an increasing of deep learning models has been introduced to address various fundus image tasks, e.g. segmentation, classification, and enhancement. Concurrently, this emergence has been accompanied by the development of increasingly intricate model architectures and a parameter surge. However, these methods have overlooked the crucial significance of the intrinsic characteristics of the data. In this regard, we posit that a vanilla ResNet, when coupled with appropriate data augmentation, can also achieve a favorable performance. This study assesses the performance of the foundational deep learning architecture, the standard Residual Network (ResNet), in predicting Spherical Equivalent (SE) values. Spherical Equivalent (SE) is a crucial parameter in optometry employed for prescribing eyeglasses to correct vision. Our study intends to demonstrate the effectiveness of ResNet with proper data augmentations in achieving precise SE predictions without additional enhancements. Utilizing a vanilla ResNet50 model, we secured the third position in the Myopic Maculopathy Analysis Challenge 2023, task 3. Remarkably, our findings reveal that the unaltered deployment of ResNet yields exceptional predictive performance for estimating SE, highlighting the architectural prowess of simple models in a vital yet frequently demanding optometric context. Our software package is available at https://github.com/HuayuLiArizona/Prediction-of-Spherical-Equivalent-With-Vanilla-ResNet.