Boosting Contrastive Learning for Myopia Prediction: Which Data Augmentation Works Best?

Myopia is an increasingly prevalent global health issue, underscoring the need for accurate and early-stage prediction methods. Given the inherent scarcity of labeled medical data, contrastive learning offers a promising solution for myopia prediction. However, it still remains unexplored. In this paper, we focus on investigating the impact of data augmentation strategies on contrastive learning for fundus-based myopia prediction and propose CImPred, a contrastive learning-based framework for myopia prediction using retinal fundus images. To support this task, we construct a fundus image dataset, RFMPred. We systematically evaluate 36 combinations of common augmentation techniques and demonstrates that the combination of color distortion and random rotation yields the best performance. CImPred achieves an RMSE of 1.630, PCC of 0.672, and MAE of 1.321 on the RFMPred dataset. We also provide preliminary insights into fundus regions associated with early-stage myopia, which may support early intervention and risk reduction efforts. The code will be made publicly available upon publication.