Two-Stage Cross-Domain Ocular Disease Recognition With Data Augmentation

Ophthalmic diseases afflict many people, and can even lead to irreversible blindness. Therefore, the search for effective early diagnosis methods has attracted the attention of many researchers and clinicians. At present, although there are some ways for the early screening of ophthalmic diseases, the early screening of fundus images based on deep learning is generally favored by the medical community due to its non-contact characteristic, non-invasive characteristic and high recognition accuracy. However, the generalization performance of a common model and cross-domain identification is usually weak due to different collection equipment, race, and patient conditions. Although the existing fundus image recognition technology has achieved some results, the effect is still in the cross-domain problem and is not satisfactory. This paper proposes a cross-domain retinal image recognition framework based on data augmentation and deep neural networks. Firstly, the ResNeXt101 model pretrained on the ImageNet dataset is selected as the base framework. The one-stage model is then trained in the source domain using this framework. Secondly, the model obtained from the first stage is further fine-tuned in the target domain to obtain the two-stage final model. During this process, various data augmentation techniques and focal loss are employed to improve the recognition performance. Experimental results demonstrate that by incorporating common data augmentation techniques and focal loss, the proposed framework achieved the following performance metrics in cross-domain experiments from train-site to on-site: a kappa score of 0.845, an F1 score of 0.923, and an AUC (Area Under the Curve) value of 0.974. In conclusion, the proposed method effectively addresses the issue of poor generalization in cross-domain early retinal screening and provides insights and directions for future related work.