An Improved MPB-CNN Segmentation Method for Edema Area and Neurosensory Retinal Detachment in SD-OCT Images

In the past, when facing with spectral-domain optical coherence tomography (SD-OCT) images of various types of age-related macular degeneration, such as neurosensory retinal detachment (NRD), pigment epithelial detachment (PED), hard exudate (HE), cystic edema (CE), and diffuse edema (DE), it was difficult to obtain satisfied segmentation results using traditional methods, because the DE and CE easily disturb the accuracy of NRD segmentation. In this paper, an improved multi-scale parallel branch convolutional neural network (MPB-CNN) network is proposed to perform the edema area (EA) segmentation and the NRD segmentation, where the supervised loss function is modified by adding area perimeter ratio constraint. The experiments on 98 cubes from 54 patients indicates that our method can achieve a mean overlap ratio 72.48% (NRD) and 75.93% (EA), respectively.