MPB-CNN: a multi-scale parallel branch CNN for choroidal neovascularization segmentation in SD-OCT images

Abstract

Choroidal neovascularization (CNV) generally appears in the advanced stage of age-related macular degeneration (AMD) and has more complex pathological characteristics comparing with most other retinal diseases. So far, few literatures focused on the CNV segmentation from a machine learning and computer vision perspective. Therefore, accurate CNV segmentation is a challenging problem in SD-OCT images. In this paper, a multi-scale parallel branch CNN (MPB-CNN) is proposed for CNV segmentation. First, three parallel branch networks are used for multi-scale feature extraction. In each branch, standard convolution is replaced with atrous convolution, in which wider and more powerful multi-scale information can be extracted due to the sparsity of convolutional kernels. To further improve the segmentation results, intra-branch connections are introduced to preserve signal transmission and inter-branch connections are introduced to interact multi-scale information. Then, feature maps from different branches are cascaded with low-level features computing by several convolutional layers. The combined feature is used as the input to acquire the final prediction map by stacking three convolutional layers. Besides, extra branch supervisions are applied at the end of each branch to guarantee the discrimination of feature representations from each branch and benefit the network optimization. Finally, gradient constraint is added to loss function to preserve the boundary of the CNV lesion. An effective cross validation is performed on 202 cubes from 12 patients based on patient independence with the mean dice value being 0.757 and the mean overlap ratio being 60.8%. Experiment results indicated that the proposed MPB-CNN can provide reliable segmentations for CNV from SD-OCT images.