Automatic Segmentation of the Golgi Apparatus in Volumetric Data with Approximate Labels

Abstract

The Golgi apparatus (GA) is a cellular organelle involved in the processing and sorting of proteins in eukaryotic cells. Due to its numerous functions, structural complexity, and organizational dynamics, the role of the GA in normal and pathological processes is still under intensive research. In this work, we present an approach to automatic segmentation of the GA in electron microscopy volumetric data, consisting of i) a neural network trained on approximately labelled data, ii) active contours for refining the segmentation, and iii) filtering of the segmented regions. Evaluation on 3D volumes of a urinary bladder epithelial cell shows that the proposed algorithm is able to segment the GA with 89% sensitivity and 99% specificity. Using approximate labels reduced the time needed for manual annotation of the ground truth by a factor of five.