SSTEM Cell Image Segmentation Based on Top-Down Selective Attention Model

Abstract

We propose an automatic method for segmenting neurons in the TEM cell images based on a top-down attention model, which is efficient to solve the discontinuity problems in TEM cell image caused by loss of section or branching of cell. At first, the proposed model enhances cell boundaries using a partial differential equation based on hessian matrix, which can improve the contrast and continuity of cell membranes in the TEM images. Then, a top-down attention model trains the shape characteristics of the desired target neurons through the reinforcement and inhibition learning process. The top-down attention model localizes a candidate neuronal region in subsequent TEM image, which was implemented by a growing fuzzy topology adaptation resonance theory network (GFTART) model. It is efficient to resolve the discontinuity problem of TEM cell image. The localized candidate target neurons are finally indicated whether they are correct ones by an active appearance model (AAM). Experimental results show that the proposed method is efficient to segment the TEM images.