Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a fundus image processing machine learning models that is configured to process one or more fundus images captured by a fundus camera to generate a predicted label. One of the methods includes generating training data, comprising: receiving sets of one or more training fundus images captured by a fundus camera; receiving, for each of the sets, a ground truth label assigned to a different image of the eye of the patient corresponding to the set that has been captured using a different imaging modality; and generating, for each set of training fundus images, a training example that includes the set of training fundus images in association with the ground truth label assigned to the different image of the patients eye; and training the machine learning model on the training examples in the training data.