Vol. 51, Issue 4, pp. 541-549 (2021)

Abstract

This work presents an automated segmentation method, based on graph theory, which processes superpixels that exhibit spatially similarities in hue and texture pixel groups, rather than individual pixels. The graph shortest path includes a chain of neighboring superpixels which have minimal intensity changes. This method reduces graphics computational complexity because it provides large decreases in the number of vertices as the superpixel size increases. For the starting vertex prediction, the boundary pixel in first column which is included in this starting vertex is predicted by a trained deep neural network formulated as a regression task. By formulating the problem as a regression scheme, the computational burden is decreased in comparison with classifying each pixel in the entire image. This feasibility approach, when applied as a preliminary study in electron microscopy and optical coherence tomography images, demonstrated high measures of accuracy: 0.9670 for the electron microscopy image and 0.9930 for vitreous/nerve-fiber and inner-segment/outer-segment layer segmentations in the optical coherence tomography image.