Fig. 1: Data split in the local dataset (SIM cohort) for the training and local validation of the three sub-networks of the DeepDR system.

The local dataset was randomly divided into training or validation datasets. All 466,247 images in the training dataset were used for training the image quality assessment sub-network. The lesion detection sub-network was trained using 10,280 gradable images with retinal lesion annotations. Then, 415,139 gradable images in the training set were used for the training of the DR grading sub-network. All 200,136 images in the local validation dataset were used to test the image quality sub-network, and 178,907 gradable images were used to test the DR grading sub-network. Finally, 4621 gradable images labeled with retinal lesions were used to test the lesion detection sub-network. DR, diabetic retinopathy.