High-Accuracy Retinal Layer Segmentation for Optical Coherence Tomography Using Tracking Kernels Based on Gaussian Mixture Model

Ophthalmology requires automated segmentation of retinal layers in optical coherence tomography images to provide valuable disease information. Sensitive extraction of accurate layer boundaries stable against local image quality degradation is necessary. We propose and demonstrate a powerful, accurate segmentation method with high stability and sensitivity. The method uses an intelligent tracking kernel and a clustering mask based on the Gaussian mixture model (GMM). Combining these concepts yields robust, degradation-free tracking with highly sensitive pixel classification. The kernel extracts boundaries by moving and matching its double faces with locally clustered images generated by GMM clustering. The cluster-guided motion of the kernel enables sensitive classification of structures on a single-pixel scale. This system targets seven major retinal boundaries. Then, using peak detection, additional two simple boundaries are easily grabbed in regions where their distinct features emerge sufficiently in the limited space remaining after the previous segmentation. Using these hybrid modes, successful segmentation of nine boundaries of eight retinal layers in foveal areas is demonstrated. A 0.909 fraction of a pixel difference appears between boundaries segmented manually and using our algorithm. Our method was developed for use with low-quality data, allowing its application in various morphological segmentation technologies.