A Multi-Scale Deep Convolutional Neural Network For Joint Segmentation And Prediction Of Geographic Atrophy In SD-OCT Images

Geographic atrophy (GA) generally appears in the advanced stage of age-related macular degeneration (AMD). It is a principle cause of the severe central visual loss for elder adults with non-exudative AMD in developed countries. In this paper, a multi-scale deep convolutional neural network is proposed for the joint segmentation and prediction of GA. First, restricted summed-area projection (RSAP) technique was used to generate GA projection images from the SD-OCT volumetric data. Then, GA projection images were sent to the multi-scale branches to acquire multi-scale feature maps. The final GA segmentation results were obtained by refining the multi-scale feature maps with a voting decision strategy. In the end, those multi-scale feature maps were cascaded with low-level features computed from the original images to predict the growth of the GA lesion. The segmented and predicted GA lesion in the tested scenarios resulted in a satisfying accuracy, comparing with the observed ground truth.