Self-supervised Learning and Data Diversity Based Prediction of Spherical Equivalent

Abstract

This report presents the technical details of the approach of Team AIfuture for the Myopic Maculopathy Analysis Challenge Task 3. The approach focuses on the following two aspects: the shift in data distribution between pre-trained and competition datasets, and the diversity of data sample. The ResNet-50 backbone is used to establish a strong baseline, and the first two-stage blocks are frozen. To alleviate the problem of data distribution shift, publicly available medical data is used for self-supervised learning, utilizing the well-known DINO algorithm. Various data augmentation techniques are employed to increase the diversity of data samples. Additionally, it has been observed that using a portion of the training data can significantly improve performance. Finally, test-time data augmentation is used for ensemble prediction, which greatly enhances model performance. The achieved $R^2$ score of 0.8636 and MAE score of 0.7326 on the test data result in the final rank of 2.