Abstract: Spectral domain optical coherence tomography (SD-OCT) imaging technique is widely used in the field of ophthalmology.The segmentation of retinal tissue layers plays a vital role in the diagnosis of retinal disease.The traditional 3-D graph search method is able to segment k surfaces simultaneously.Problems of this algorithm associated with high time complexity and weak robustness of segmenting pathological retinal images have limited its utility.Multi-scale theory was introduced to the traditional 3-D graph search modal,and a 3-D graph search algorithm based on multi-scale for segmenting retinal images was proposed in this paper.Firstly,reasonable cost functions are constructed for each surface according to the characteristics of the layer.Then the minimum and maximum differences in height of adjacent columns are utilized to construct inter-column constraints to improve smoothness constraints of surfaces.Finally,3-D graph search method is used for coarse segmentation of lower-scale images and then a search region of higher-scale images is redefined for a more accurate result.The improved algorithm was evaluated on 3 groups of normal eyes and 1 group of abnormal eye with age-related macular degeneration.The experimental results were compared with manual segmentation and traditional 3-D graph search method.The results demonstrate that the improved method can effectively detect 3 layer surfaces (the mean absolute boundary positioning difference is 3.86±2.50μm) more closely to manual segmentation (3.78±2.76μm),and it is better than traditional 3-D graph search method (7.92±3.31μm).

Key words: Spectral domain optical coherence tomography,Retinal image segmentation,Multi-scale,3-D graph search