Predicting optical coherence tomography-derived diabetic macular edema grades from fundus photographs using deep learning