3D Nuclei Segmentation through Deep Learning

Abstract

Nowadays, deep-learning has been used successfully to solve difficult problems in fluorescence microscopy field. In this work, we propose a Drosophila 3D Nuclei segmentation based on a pipeline that detects nuclei centers and then segments each detected nucleus individually, using a different 3D U-net for detection and segmentation steps. Our method is among the top-3 performers in the Cell Tracking Challenge segmentation benchmark for Light Sheet Microscopy Drosophila dataset, reaching a final score of 0.827. The proposed methodology: i) allows the utilization of a U-net model to perform a detection task, and ii) requires much fewer training samples than direct segmentation of the entire volume, reducing the manual annotation effort.