Fig. 2: Visual diagram of the DeepDR system.
From: A deep learning system for detecting diabetic retinopathy across the disease spectrum
DeepDR system consisted of three sub-networks: image quality assessment sub-network, lesion-aware sub-network, and DR grading sub-network. We first pre-trained the ResNet to form the DR base network (top row). The trained weights of the pre-trained DR base network were then shared in the three different sub-networks of the system, indicated by the red arrow. These three sub-networks took retinal images as input and performed different tasks one-by-one. Furthermore, the lesion features extracted by the segmentation module of the lesion-aware sub-network (indicated by the green arrow) were concatenated with the features extracted by the DR grading sub-network (indicated by the blue arrow). DR, diabetic retinopathy; NPDR, non-proliferative diabetic retinopathy; PDR, proliferative diabetic retinopathy.
