Accurate segmentation of cells and tissues in 5D (3D + time + multiple channels) microscopy data is critical in quantitative developmental biology. This is a challenging task due to the large volumes of data, e.g. tens of gigabytes for single time lapse 3D sequence, and the inherent characteristics of microscopy image acquisition. In this research we develop new supervised and unsupervised segmentation methods with a focus on application to morphogenesis.

First, we briefly explore the approach of correcting an over-segmented volume by using a trained model. The proposed method automatically initializes with seeds according to the local density of cells in the volume. Next, this algorithm merges pairs of super-pixels based on a learned model using a feature representation which effectively discriminates between spurious and correct boundaries.

Second, we develop a principled approach to unsupervised segmentation and fusion of multiple segmentations. A linear optimization framework is proposed for the joint correction of multiple over-segmentations obtained from different methods. The main idea motivating this approach is that over-segmentations, from a pool of methods with various parameters, are likely to agree on the correct segment boundaries, while spurious boundaries are likely to be method or parameter-dependent.

Third, we introduce an interactive segmentation and analysis tool for 5D microscopy data, called CellECT. An adaptive confidence measure, called cellness metric is used to highlight regions of uncertainty in the segmentation. This metric quantifies how much a segment deviates from a typical correct segment. This metric adapts to the dataset and learns from the user interactions. The proposed methods are validated on an ascidian time-lapse 3D volume data. CellECT is distributed as an open source software and is used in other quantitative biology applications.