Multi-label detection of ophthalmic disorders using InceptionResNetV2 on multiple datasets

Recently, AI-based methods have been extensively used to help in the process of diagnosing eye diseases due to their prevalence. But since these methods can’t be generalized well, they can’t be used in the real world. In this paper, we compared the two fundamental approaches for improving the model’s performance on the eye disease detection task. The idea is that, for real-world applications, using multiple datasets for robustness is more beneficial than enhancing the architecture just to increase the accuracy. To demonstrate this, we chose three state-of-the-art architectures as our baseline and changed them slightly so that the overfitting wouldn’t happen. For the first approach, we change the classification head to XGB and SVM, and for the second approach, we combine the two datasets for the training stage. The results show that high-quality data with representative distribution can have a better effect than sophisticated architecture for real-world applications. This approach performed 3% better than the last state-of-the-art model. The implementation is available at https://gitlab.com/asiyeh.bahaloo/eye-disease