Introduction

Quantitative analysis of microscope images forms a growing part of modern cell biology. It is all too easy for such studies to generate vast amounts of data, which can become difficult to keep track of (1,2,3,4,5). The Open Microscopy Environment (OME) software system (1,3,4,5,6) was developed to provide a flexible, powerful, and freely available platform for storing, viewing, and analyzing such images, and has a clear audit trail to link all images with the analyses performed on them. It allows automated mining and analysis of image sets independent of the particular system originally used to capture the images. It facilitates analysis of these image sets using built-in or third-party image-analysis programs. Because images can be arranged in groups as required, and analyses can be performed on entire groups, automated image analysis of such groups of images is possible to an extent not readily achievable by methods that require manual analysis of each image separately.

A particular strength of OME is that it allows for a highly structured description (data model) of the information that it stores (6) and encourages the reuse of these data models in different algorithms and “viewers”—software that can display the images themselves or that can take the results of such algorithms and display quantitative information about them (e.g., a spreadsheet application). The benefit of using common data models is the transparent exchange of structured information between independently developed software.

Common analysis tasks in imaging are finding spots or “blobs” in images based on various criteria and obtaining structured information to describe them (positions of the blobs, their sizes, shapes, intensities, etc.). Although various spot-finding tools have been described, none can fully take advantage of the unique informatics abilities of OME. Here, we describe and validate FindSpots, an analysis package that is fully integrated into OME, version 2.4.

FindSpots can automatically detect and quantify spot-like objects within microscope images, whether they be two-dimensional (2-D) flat images or three-dimensional (3-D) image stacks. It can also be applied to images collected using more than one wavelength and can analyze time-lapse movies, which allows for subsequent tracking of the objects in 3-D. Thus, the algorithm is capable of operating in all five dimensions accessible on a modern digital microscope (4,7). The algorithm underlying FindSpots itself has been used in earlier work (8,9), but the current work represents the first integrated use of FindSpots and OME and the first example in the context of OME of its use to perform complex biological image analysis.

Theoretical Basis

Approaches to Object Detection/Tracking

Object finding (and therefore object tracking) in FindSpots and in many other image-analysis programs, including the commercial package Volocity® (Improvision®, Coventry, UK; www.improvision.com/products/Volocity), is based on a procedure known as global thresholding, whereby in any one image, all pixels above a certain minimum value are potentially considered to be part of an object (subject to a minimum spot-size requirement). An unavoidable problem with defining objects based on a threshold—especially a global one—is that whatever value it is set to, some subthreshold pixels that ideally should form part of the objects can be lost. However, lowering the threshold to include these pixels will cause too many artifactual pixels to be included (10). Where spot contrast or cellular background varies widely within an individual image, such unwanted omission or inclusion of pixels is more likely to be a problem, and it may be necessary to use adaptive thresholding algorithms, which allow the threshold to be varied within any one image (11) or employ advanced background correction methods. Indeed, methods other than thresholding (e.g., edge detection) can sometimes be helpful for image segmentation (i.e., the identification of objects within an image) (11). However, with any segmentation method, it is inevitable that some pixels that should form part of spots will be omitted and vice versa; no segmentation algorithm is perfect and any such quantitative analyses will inevitably be approximations (10,12). Moreover, in many cases, a low level of misclassification will not matter in practice.

Additionally, although we present here a simple but effective method for background correction, more advanced variations on this, as well as other specialized algorithms, also exist and are worth noting (www.moleculardevices.com/technotes_d1/MDC_D1_D50022_MorphologyFilters.pdf; References 10, 13, and 14), but a full discussion of these procedures is beyond the scope of this article. In general, it should be noted that sophisticated segmentation algorithms are no substitute for high-quality image data when performing quantitative image analysis.