3D Level Set Method via Local Structure Similarity Factor for Automatic Neurosensory Retinal Detachment Segmentation in Retinal SD-OCT Images

Qualitative assessment of pathological changes is vital for clinicians to determine the degree of neurosensory retinal detachment (NRD). However, accurate segmentation is challenging due to the diversity of NRD size and location. Spectral domain-optical coherence tomography (SD-OCT) imaging technology can yield high-resolution, three-dimensional images of the retinal histopathological structure without invasion and injury, whereas always accompanied by high-level noise and low-contrast intensity. In this paper, a novel automatic segmentation approach was presented based on two-stage clustering and an improved 3D level set method to quantify NRD regions in SD-OCT images. Considering the difference intensity distribution of NRD regions and the background, an unsupervised two-stage clustering method based on k-means was used to get initial surfaces for subsequence segmentation. In order to reduce the effect of noise and utilize spatial constraint information, 3D local structure similarity factor was proposed by combining the characteristic of the intensity of a voxel can be similar to those of its neighbors, which was introduced into level set model. Comparing with six methods, the experiment of 23 volumes from 12 patients demonstrates that the proposed method can improve the accuracy of location and segmentation for NRD regions in SD-OCT images.