Registration of microscopic images and its uses in segmentation and annotation ] 3-D Registration of Biological Images and Models © istock photo.com/beano5 IEEE SIGNAL PROCESSING MAGAZINE [71] JANuARy 2015semantic properties such as the identities or categories to objects or patterns. Segmentation and annotation are critical to address important biological questions (e.g., quantification of gene expression patterns, generation of the ontology databases, and digital atlases of model animals).From Image-To-Image regIsTraTIon To model-To-Image regIsTraTIon Registration is often needed to compare, fuse, or quantify objects or patterns in images. In many cases, registration is also required to map images to models and vice versa. In these latter situations, a model often consists of geometric shape description of the anatomy or spatial layout of biological objects in the respective images.
Image-to-Image RegIstRatIonMany system biology studies rely on aligning images of gene expressions in different cell populations [6]- [8] or specimens that correspond to different developmental times [9]. In several recent brain mapping projects of the Drosophila (fruit fly), it became critical to align a number of 3-D confocal images of the insect's brains. Each fly had been genetically engineered to express fluorescent proteins in a specific population of neurons, which were aligned to a standard space so that they could be compared with each other [ Figure 1(a)]. The FlyCircuit project in Taiwan [10] and the FlyLight project at the Janelia Research Campus of the Howard Hughes Medical Institute [11] each generated tens of thousands of 3-D fruit fly brain image stacks represented some of the biggest neuroscience efforts to date to understand the brain's structure. In each of these brains, some neuron populations are labeled using genetic methods. In both projects, registration of brain images is crucial. Registering images that correspond to the same population is useful to quantify the intrapopulation variability of neurons, which can further help define the meaningful neuron types. Registering images that correspond to different populations is useful to quantify the spatial proximity of neurons and thus helps estimate the putative connectivity of neurons. Similarly interesting results for the zebrafish (Danio rerio) were also reported recently [3], [4], [12]. Sophisticated volumetric image registration methods have been developed in the biomedical imaging field. Many methods, such as mutual information registration [13], spline-based elastic registration [14], invariant moment feature-based registration [15], and congealing registration [16], [17], have been widely used and extended to align molecular and cellular images. However, since many of them were originally designed for magnetic resonance imaging and computer tomography data, in many cases it remains challenging to use them easily and effectively in aligning the microscopy images that have larger-scale and fuzzier contents.Two major challenges in biological image registra...