Image analyzers fall into two major classes. Turnkey systems are delivered with very complete software, designed specifically for your applications. No programming is required. In contrast, macro-based systems require some user programming. A simplified programming environment is supplied, to generate mini-programs called scripts or macros. Users can purchase or create complex macros to provide some approximation of turnkey function.

In general, macro-based software is well suited to non-quantitative tasks, while turnkey software should be preferred in quantitative image analysis. The reason for this preference lies in the requirement for replicable measurement. The finest turnkey software has been validated by rigorous testing, has generated an extensive publication record, and has evolved to the point that its use supports the validity of a body of data. Editors, colleagues, and regulatory bodies can be confident that data will be read identically in Toronto and Tokyo, and this standardization must be recognized as a major benefit of the turnkey system.

Given the advantages of turnkey software (no user programming, validated and consistent modes of operation), why are so many commercial image analyzers macro-based? The answer is that the computer programmers and engineers who create most image analyzers lack a deep understanding of end user requirements. Creation of a dedicated biological image analysis system is beyond them, but they can easily market general-purpose image processors with a simplified programming language. These devices are sold as "flexible" (user programmable) and "easy to use" (the programming process is simplified). With such a system, ultimate responsibility for software design and system performance is taken away from the manufacturer and placed upon the researcher.

For simple imaging applications (such as image enhancement in video microscopy), or if the types of image analysis being performed are unique, the macro-based system can be very successful. It can also be lower in cost than a turnkey system, as it is easier for the manufacturer to create and support. However, macro-based systems can be very dangerous in quantitative applications, because macro software may not have been rigorously validated. Macro functions could yield biased or inaccurate data, that are not comparable from laboratory to laboratory.

To avoid problems with accuracy and replicability, an image analyzer should be selected using the same criteria applied to any other complex measurement device. That is, it should be well proven, and function in the same way in every laboratory. A high quality turnkey system is well proven, in that it has been used to generate many publications. It functions in the same way from lab to lab in that it is supplied with standard software that does not require user modification.

The major drawback of the turnkey system is that it is less flexible than a macro-based system, and may be unable to adapt to unusual demands. Therefore, the system repertoire should include turnkey software spanning the most common bioscience applications, and programmable functions that can be adapted to unusual applications. MCID™ supplies both turnkey software, and user-programmable functions (Figure 1).

Figure 1: A simple macro constructed using MCID's object-oriented macro editor. This macro illustrates the principle of chaining a series of operations into a sequence, which can be executed with one keystroke. Three MCID functions are represented by objects, which are executed from left to right. The first object turns on density measurement. The second object digitizes an image. The third object places a sample window onto the image, ready to read data. The macro is easily elaborated. For example, we could attach a camera object to the input of the density object (at the far left of the macro). We could make the sample window circular or rectangular. We could feed the density value from the sample object to an equation, attached to the output at the far right of the macro.

We will refer to an image analyzer equipped with bioscience software as a bioscience image analyzer (BIA). When image analysis was new, many laboratories were forced to spend a great deal of time and money constructing their own BIAs (e.g. Goochee, Rasband and Sokoloff, 1980; Grynkiewicz, Poenie, and Tsien, 1985; Ramm et al., 1984). However, this is no longer the case and a bewildering variety of image analysis systems are now available from dozens of manufacturers. Two developments have been particularly important in widening the BIA user base.

Low-cost imaging systems have become sufficiently capable to handle most image analysis applications.

Image analysis software has been under development for more than 15 years. Some of it works now.

The reduction in hardware costs allows a BIA to be built for the price of a quality microscope. The maturation of software makes the instrument easy to use, applicable to a broad range of tasks, and reliable.

As always in the rapidly changing computer world, software lags behind hardware. However much it has matured, image analysis software remains the limiting factor in this subtle and complex technology. Some image analyzers yield valid, replicable data. Others, which may use similar hardware, do not. How is the researcher to determine which of the commercial systems can be trusted? The problem is compounded if the instrument is new to the market, or uses macros which can vary from site to site. Our next topic details some issues to be considered in evaluating a BIA.

Because imaging systems are visually interesting, many BIAs are designed more for appearance and easy marketing than for real-world data gathering. In contrast, the acquisition of valid data requires a BIA designed with a deep knowledge of image analysis and of bioscience. That is, building an efficient and accurate BIA requires a close-knit group of engineers, programmers, and research scientists. Few companies have access to this level of expertise and, therefore, many BIAs have major problems. These problems include deficiencies in both the accuracy and efficiency with which data are gathered.

Accuracy
Measurement procedures within the BIA software are rarely documented, and there are no agreed standards for BIA operation. These factors can lead to serious, hidden deficiencies in measurement accuracy. Problems are most likely to arise when poorly trained sales people install "custom" macros at customer sites. Insist that any quantitative functions be validated by the BIA manufacturer.

The chord, for example, measures the maximum internal distance that does not cross a boundary. Chords can be straight or curved. A machine-calculated curved chord should correspond to a human observer's tracing of a maximum internal length (Figure 2). The diameter should correspond to a width taken across the chord. Therefore, an accurate curved chord is both a measure of internal length, and a critical first step to valid diameter measurements. Although the curved chord is a very difficult measurement for a machine, MCID automatically generates chords that are very close to those of a human observer (Figures 2-3).

Figure 2: Calculating the chord of an irregular structure. The target at left contains a straight chord. This is the longest straight-line internal distance. The target at right contains a curved chord, which is the measurement of internal distance required in most biological applications. Our systems generate either or both of these chords.

Figure 3: Accuracy of MCID's curved chord measurements. The targets were irregularly-shaped, thionin-stained cells. The curved chords were scanned automatically, or were traced by hand. The correlation between automatic and hand-traced measurements is 0.99.

If the image analysis system you are evaluating does not have an extensive literature validating its use, we recommend you check each type of measurement. Do not simply assume that the numbers are accurate. As examples of the types of check that could be made, refer to Figures 4-5.

Figure 4: Comparisons of cell areas derived from automated scanning and manual tracing. Area is one of the simplest measurements for an image analyzer. MCID users Robert and Barbara Hitzemann (Hitzemann, Qian and Hitzemann, 1993) report a correlation of 0.98 between fully automated MCID area measurements of tyrosine hydroxylase positive cells, and area measurements derived from manual tracing of those cells. Paraformaldehyde-fixed brains were postfixed and cryoprotected in 20% sucrose. Sections were cut at 30u, exposed to rabbit anti-rat TH antibody for 48 - 72 hours at 5' C. Staining was developed using an avidin peroxidase complex with DAB as the capturing agent. The stain was enhanced with osmium.

Figure 5: MCID/AIS performance in grain counting. A radiolabelled in situ probe was measured within counterstained cells. The left graph shows that the number of grains counted by the image analyzer correlates very highly with counts made by a human observer. The right graph shows that there is also a high correlation between visual grain counts and proportional grain area within the cell, as measured by the image analyzer. Proportional area is a simpler measurement for a machine than counting.

Efficiency
Efficiency is critical if we are to gather large quantities of data in minimal time. Efficiency is directly dependent upon the skill with which the user interface is implemented. As examples of how efficiency affects image analysis, consider the following.

A competitor's BIA took about 5 minutes to accomplish a simple density calibration that would be handled by MCID or AIS in about 30 seconds. Following calibration, data throughput (the number of specimens analyzed / unit time) was very poor.

Users of phosphor plate imagers report that data throughput is much greater when the images are analyzed on MCID or AIS, than when the software supplied with the plate imager is used. In one typical case, a body of data that took 16 days to analyze with the plate imager software took 4 days to analyze with MCID (Potchoiba et al., 1995).

Our software for the analysis of high density arrays (dot blots, microwells, etc.) processes many thousands of discrete targets within minutes.

Some rules for evaluation
BIA accuracy and efficiency are difficult to determine without extensive testing. An adequate evaluation requires more extensive trial than can be had during a brief demonstration. Fortunately, it is possible to obtain enough information to make an educated choice, if the following recommendations are kept in mind.

Do not be impressed by aggressive marketing presentations. A sales demonstration may not relate to real-world performance. In particular, functions that appear easy to perform during a demonstration may be inefficient for actual data gathering. Insist on analyzing a quantity of your own data during any demonstration.

Evaluate the BIA's acceptance by the research community. How many users of the product are there? A large and satisfied user base is the best assurance of performance, and of the manufacturer's permanence. It is also an indication that the BIA data will be accepted as valid by reviewers, government agencies, etc. Contact some of the experienced users and obtain their impressions.

Examine the manufacturer's support policies closely. Can callers easily contact software developers and scientists, or must they filter through layers of marketing staff? The unfortunate experience of many BIA purchasers is that many sales representatives do not provide adequate after-sales support.

The producer of the BIA should be directly accessible to end users, even if the BIA is sold by a dealer. Few dealers have a deep understanding of image analysis in general, or BIA function in particular. Therefore, you will appreciate technical support from the BIA producer's scientific staff.

Make sure that the BIA exists as an independent entity, even if a particular dealer ceases to carry it. There is a long history of BIAs being discontinued at short notice by major dealers, and this can lead to a complete loss of support for a costly instrument.

Evaluate your willingness to program the BIA. Turnkey BIAs (e.g. MCID, AIS) do not require programming. Rather, they can be used by any lab personnel without modification. Macro-based systems (e.g. the Quantimet, Image1, Bioscan) require that at least one person become an expert in programming the system, and in image analysis. You must learn to program a macro-based system, or rely on the instrument supplier's staff to program it for you. Remember that relying on the supplier is very dangerous, without proof of a highly developed competence and careful software validation.

Inquire as to how the manufacturer's R&D staff is structured. Are scientists closely involved with BIA development? For example, our own R&D staff includes both life scientists and software professionals. We maintain an active research lab performing quantitative autoradiography, cell biology (including tissue culture, optical physiology, and ratio fluorescence imaging), molecular biology, and biomolecular screening. Finally, we constantly poll our user base (more than 500 systems) to ensure that everything is as it should be. Try to ensure that the manufacturer of your BIA takes similar care with their product.

Attention to the above general points should assist in your evaluation of competitive BIAs. In addition, a purchase decision is easier if you have some knowledge of how BIAs operate. In the following sections of this report we will discuss the operating principles of BIAs. We will also provide some characteristics and specific performance figures for our own MCID systems. We hope that this information will serve as a guide, and as an index of comparison for competitive systems.

A definition of image processing
Digital image processing is a body of techniques used to uncover information in images. The information is brought out by transforming the digital image (e.g. increasing visual contrast), or by directly extracting data values (e.g. density, edges) from it. An initial step in digital image processing is the transformation of image features (density, color, position) into discrete digital values. This transformation process is known as digitization. During the digitization process, the continuous spatial extent of the image is broken into discrete spatial elements which are stored in a memory bank (image memory) within the image processor. A single "picture element" in image memory is a pixel. With the location and density or color of each pixel digitally coded, image processing becomes a matter of "number crunching".

Each pixel in image memory has values associated with it (Figure 6). There is a memory address corresponding to a specific X and Y location in a Cartesian coordinate system, and a density value in the form of number of "gray levels". In most imaging systems, monochrome densities are resolved into 256 gray levels (8 bit density resolution). Higher density precision (e.g. 4096 gray levels, 12 bit resolution) is available from specialized scanners and advanced imaging systems (such as MCID/AIS).

Figure 6: A rectangular image, represented as an array of pixels. The pixel at top left is at position (0,0) in a Cartesian coordinate system ascending from left to right and from top to bottom. The pixel at bottom right is at position (3,3). In representing pictorial data, each of the pixels would contain a gray level value.

In true color systems, a set of values from one of the standard color spaces (RGB, HSI, CMYK) replaces the gray level density of a pixel. Most commonly, colors are assigned a total of 24 bits, 8 bits for each of red, green, and blue. From these RGB values, any combination of hue, saturation, and intensity can be created.

Imaging boards
Most image analyzers include both a host computer and a dedicated peripheral device, the imaging board. There are two common types of imaging board.

A complex board is a combination of acquisition (frame grabber), memory, and processing circuitry. This type of board (actually often a set of boards) is specialized for massive image data throughput. It performs image acquisition, transformation, and display functions, under control of the host computer.

Simple boards function as camera interface controllers, and transfer data from analog and digital cameras to the computer memory. Simple boards lack the speed and processing power of complex boards, but can be very cost-effective. Fast CPUs (e.g. a 300 mHz Pentium II) can use software to process image data quite rapidly. Because image processing work is done in the host CPU, this type of system is referred to as a host-based system.

We are often asked whether a host-based system is sufficient. With the advances in Computer processing technology over the past few years host based imaging systems are now the norm, personal computers have become fast and with the introduction of digital cameras using proprietry interfaces such as firewire or USB the need for adding expensive image processing boards and framegrabbers is now a thing of the past.

The best aspect of simple imaging boards and host-based processing is that some imaging systems can now offer unprecedented levels of analytical power at relatively low cost. However, there is also a dark side to this force. Many of the commercial BIAs built around the most basic hardware appear to be educational toys, rushed to market by unskilled developers. Some of the boards distort image data during acquisition, and the software is poorly constructed and inaccurate. Our advice is to evaluate the functions and the accuracy of low cost BIAs very carefully. Select a low cost image analyzer with the same care as you would a state-of-the-art instrument. Lower cost will mean sacrifices in speed, and in the number of advanced functions available. There should be no sacrifices in data accuracy or the quality of software.

Imaging system resolution
Most modern image analysis systems will accept medium (e.g. about 700 x 500 pixels) and high resolution data (e.g. 1280 x 1024 pixels). Medium resolution is adequate for many bioscience applications. At 640 x 480 pixels, for example, an entire digitized rat brain section can be displayed with a spatial resolution of about 30 um/pixel. A 1280 x 1024 image offers four times the display area of 640 x 480 pixels. Therefore, high resolution BIAs (such as our MCID Core system) can more easily work with large specimens, and will see finer details through the microscope. High resolution systems also allow many images to be displayed on one screen. This can be an advantage in showing how conditions change across time or across anatomical location. For example, the high resolution display might contain a series of many smaller ratio images, showing changes in intracellular calcium ion concentration over the course of an experiment.

Even the pretty standard PC resolution of 1600 x 1200 pixels may be insufficient for some purposes. For example, imaging an entire human brain section at a practical resolution of about 100 um would require about 4000 x 4000 pixels. Each image would be more than 8 MB in size. Certainly, there are no commercial image analysis systems that can display such large images at full resolution. Therefore, a common strategy is to work with small areas at a time. That is, we view a small portion of the specimen, analyze that and move on to the next portion. The disadvantage of this is that it is difficult to reconstruct a view of the entire large image from the smaller components. We have a solution to this problem, which we call large image format.

Large image format
When an MCID or AIS system encounters an image that is too large to be contained within its display, it creates a compressed view with a roaming window, and a full resolution view. The roaming window is moved around on the compressed view, to select part of the large image for high resolution display (Figure 7).

Figure 7: How large images are handled by MCID. A large image can be displayed in compressed form. A roaming view window is moved about on the compressed image, to select a portion for high resolution display.

The advantage of large image format is that we analyze data from the high resolution view. This gives the best possible accuracy. At the same time, we can position ourselves very easily using the compressed view. Because large images are handled so well, MCID and AIS are ideal for images from ultra-high resolution cameras, drum scanners, scanning densitometers, and phosphor plate imagers. These all can create large images.

Image Memory
Image memory (also called a frame buffer) is a bank of random access memory that lies within the imaging system. Standard imaging systems incorporate large amounts of fast image memory. However, there is a growing tendency towards the use of fast host memory as virtual image memory. This is only efficient if high speed bus architectures (particularly PCI) are used. MCID and AIS make extensive use of host memory as virtual image memory.

Image Transformation and Display
It is possible to display a digital image which looks the same as the original. Very often, however, the image must be transformed to alter its appearance or to discriminate features of interest (targets) from background. There are four major classes of transformation.

A lookup table (LUT) is a means of mapping gray level values onto new values. Input LUTs transform data after they are digitized, but before they are stored in image memory. Thus, input LUTs change the data in image memory and are used in many types of target discrimination. In contrast, output LUTs do not change the data in image memory. The output LUTs take data out of image memory and assign display characteristics to the data values. Output LUT processing produces the familiar contrast enhancement, pseudocolor and other visual effects of imaging systems.

Multiple image transformations, including image averaging and subtraction, require arithmetic operations to be performed on discrete pixels in two image channels.

Spatial transformations process pixels in groups. Registration functions alter image shape and position, and spatial convolutions include high and low pass filters and edge enhancement routines. Morphology is the transformation of images in terms of shape and size, using elementary patterns called structuring elements.

Frequency domain transformations include the familiar Fourier transform. Here, the pixel spatial function f(x,y) yields a two-dimensional frequency function F(u,v), where u and v are the spatial frequency coordinates. The high and low frequency components of the Fourier data contain different aspects of image information. The low spatial frequencies contain luminance and gross feature information, while the higher spatial frequencies contain granularity, sharp edges, etc. Image properties (e.g. blur) may be selectively processed in the spatial domain, before the modified frequency function G(u,v) is transformed back into a modified spatial function g(x,y).

Lookup tables
The output LUTs do not change the data stored in image memory. Rather, output LUTs accept data output from image memory, and use a programmed transformation function to dictate how the image memory data values appear on the display.

There are many types of transforms which can be programmed into the output LUTs. A 1:1 transformation (known as a linear ramp function) results in what appears to be an exact duplication of the original specimen. A reverse linear ramp transforms black to white and produces a negative image. A pseudocolor display is generated by a LUT function that feeds different image data values to the various color guns in the display monitor. For example, high density data could feed the red gun only. The result would be an image in which "hot spots" are bright red, and less dense regions do not show up at all. Usually, however, the output LUTs are programmed with color mixtures that span the entire range of gray levels. Each gray level would be represented by a unique combination of red, green and blue. Densities from dark to light are commonly mapped onto a spectral pattern from red to blue. Our visual systems are comfortable with this type of mapping.

Because any given LUT can be reprogrammed at will, its mapping function can be altered to produce contrast enhancement and other visual effects. For example, a spectral LUT maps image data values of 0-255 levels onto colors from dark red to light blue. If the image actually contained the full range of data values (0-255), all colors would be present on the display. However, we may be viewing an image containing only values from 50-200 levels. Therefore, the display does not contain the full range of colors and appears somewhat washed out. To improve image contrast, we could change the LUT function so that the full range of colors maps onto our limited set of 50-200 density levels. This contrast enhancement operation makes the display more visually striking.

Multiple image operations
Combination functions include image addition and subtraction. For example, an image of total binding is overlaid upon an image of nonspecific binding. Following the overlay, a pixel at position X,Y in the total binding image lies within a brain region. A pixel at the same X,Y position lies within the same brain region in the nonspecific binding image. Then the pixel in the nonspecific binding image is subtracted from its spatial counterpart in the total binding image. When this process is repeated across all pixels, the result is an image of specific binding. Complex imaging boards include high speed processors that perform combination functions in real time. However, host-based imaging systems implement combination functions in software, and are slower.

Spatial processing

Registration
Registration procedures are used to alter the dimensions or orientation of an image. For example, a section could be stretched vertically by 10% to remove the effects of cutting compression from the display. Registration procedures are useful in applications such as 3-D visualization (for aligning and editing images), morphometry, and for matching of experimental and control 2-D gels. Unfortunately, registration procedures (such as image warping) are quite computationally intensive and are performed slowly on most imaging systems.

Convolutions
Convolutions are a class of functions which modify an image by basing the intensity value of each pixel on its nearest neighbors. In practice, a set of pixel weights (convolution kernel) is passed over the image, and the output of the weighting process is written into each pixel location. High pass and low pass filtering, and edge detection or enhancement are usually performed by different types of spatial convolution. Spatial convolutions can be used to enhance image appearance, or as a first step in image segmentation for automated morphometric or counting tasks. Most of today's better image analyzers offer hardware that can perform some convolutions in real time. Convolutions can be very useful in grain counting and morphometric applications.

Mathematical morphology
Mathematical morphology (Serra, 1982) alters an image by using a set of geometric operators to enhance or suppress image components based on size, shape, spatial arrangement and density. Instead of convolution kernels, morphological processing uses geometric operators known as structuring elements. The core of morphology is a set of primitive structuring elements used for erosion, dilation, opening, and closing. These operators function with either binary or gray level image data. Erosion and dilation are the primary operations. Combinations of erosion and dilation result in opening (n erosions followed by n dilations) and closing (n dilations followed by n erosions). Opening eliminates small objects, disconnects contiguous objects and smooths boundaries from the inside. Closing fills holes within objects, connects close objects and smooths boundaries from the outside.

a. Morphological operators are useful in various forms of automated image segmentation.

b. Morphology is used to deconvolve features that touch, allowing a computer to recognize multiple components within a larger unit. For example, a series of morphological operations (iterative opening and closing) can be used to separate clumped cells.

c. Often, an image contains valid targets that share a common density value with other features. Using an operator that selects for spatial properties of the valid targets can aid in their discrimination.

d. Many images contain a significant amount of background gray level variability. Thus, a gray level threshold which is appropriate to detect a given target at center may not be appropriate at an edge. Another good example is the detection of grains lying over stained cells. In this case, the cell density is often as high as background grain density and grains cannot be readily discriminated on the basis of a simple gray level threshold. Morphology is used to create a locally relevant threshold.

There are many other potential uses for morphology. It can be used to remove inclusions from targets, making the measurement of area more accurate. It can resolve proportions of targets of a given shape or size characteristic, even if they touch and overlap. However, there are also disadvantages.

The quality of morphological segmentation depends upon the selection of appropriate structuring elements and the appropriate chaining of operators for a specific class of targets. Once the structuring elements and operators are defined, the morphological operation is straightforward. However, it is highly tuned for one or more specific geometric entities. Thus, morphological processing is most efficient in highly automated scanning for a fixed set of target categories. If targets are not classifiable into a few spatially homogeneous categories, definition of operators may be more tedious than the interactive image editing required by less sophisticated segmentation techniques.

Morphology is usually applied to categorization and counting. There are few data regarding the effects of morphological segmentation upon the accuracy of automated morphometry. In other words, does morphological processing alter the shapes of targets in undesirable ways? We are conducting our own evaluations of this issue.

Morphological processes are all computationally intensive. For adequate speed of performance, they must be implemented in hardware. If morphology is an essential aspect of your routine data analyses, a system equipped with specialized processors would be advisable.

We have developed image analysis systems on both laboratory microcomputers (Ramm et al., 1984; Ramm and Kulick, 1985) and personal computers (Ramm, 1985, 1990, 1994). In most bioscience applications, an advanced personal computer is an ideal host. This opinion contradicts the common practice of some institutional imaging laboratories. Laboratories heavily involved in unique software development and/or 3-D visualization have traditionally used the scientific workstation as host. Therefore, researchers may be faced with conflicting advice from computer scientists who tend to favor workstations, and end users who appreciate the low cost, standardization, and ease of use of PCs.

There is no advantage in using a workstation for most forms of 2-D image analysis that we have encountered in bioscience. There are significant disadvantages, including higher computer cost, restricted selection and higher cost of software, requirement for expert management of the UNIX operating system, and non-portability of software from one workstation to another. We recommend a PC-based system for most forms of image analysis.

For pure 3D visualization applications, workstation-based systems do offer the best software selection at present. The choice is more difficult if a combination of image analysis and 3D visualization is required. Demanding graphics applications, such as surface and volume rendering, are performed better on workstations than on PCs. However, there is a very limited choice of image analysis software available for workstations. Many owners of 3D visualization systems complain that they can communicate data using 3D renderings, but they can't quantify those data in the original 2D images. Our own priority is for image quantification. Given precise image quantification, we are willing to accept a slightly slower and less powerful visualization system. In this context, the PC is a suitable host for a system that must serve for both measurement and visualization.

PC Operating Systems
Like any demanding application, image analysis benefits from a multitasking, 32-bit operating system. Microsoft recommends Windows XP Home for most single-user (home) computers, while the Windows XP Professional & Windows 2000 operating system is targeted at demanding applications in networked environments. Our software runs under Windows 2000 & Windows XP professional.

For quantitative densitometry, stability of illumination is absolutely essential. The illumination produced by an unstabilized illuminator varies markedly with changes in line voltage (Table 1). With a fluorescent illuminator, line voltage variations of +- 5 volts, which are not at all uncommon, would lead to variability of about +- 3% in optical density readings. Although +- 3% variation may not appear overly large, the overall variability in light transmission is about 6% from minimum to maximum, and this can have very large effects on readings of tissue isotope concentrations taken from dense regions of autoradiographs.

Table 1: Illumination (percent transmission) from a custom-built, DC-powered, halogen illuminator, and from two fluorescent illuminators (standard X-ray illuminator and Imaging Research Northern Light). Line voltage ranged from 95 - 130 V, controlled with a variable transformer. All illuminators warmed up for 20 minutes prior to test.

For many qualitative applications (e.g. gel documentation), in which density measurements are absent or uncalibrated, illuminator variation is unimportant. In contrast, an illuminator used in any form of quantitative (calibrated) analysis must provide stable output over the normal range of line voltage variability. Minor spatial variations in light intensity can be corrected by software. Unpredictable temporal variations in light intensity cannot be corrected. We manufacture a unique fluorescent illuminator (the Northern Light) that provides spatial homogeneity, and excellent temporal stability.

Almost all BIAs are capable of pseudocolor image display. That is, the image is digitized in monochrome and the computer assigns artificial colors to the gray levels displayed on the monitor. Many BIAs can also show the true colors in the original object. The ability to detect image components on the basis of color can be very useful with stained histological material, or with multiple fluorescent probes (see Fluorescence Imaging ).

True color imaging systems have traditionally defined colors as combinations of red, green and blue values with 8-bits (256 levels) for each primary. Thus, each pixel contains 24 bits of information, as opposed to the 8 bits/pixel found in most monochrome systems. This sounds wonderful, but there are drawbacks.

A pixel's color or density is described in terms of all of the primary color components (three 8-bit values instead of one). As all three primaries are used, working a true color image takes longer to process than a monochrome image.

Three times as much storage is required to hold the 24 bit/pixel images. A 640 x 480 x 24-bit image would require 1 MB. We have large images, made with our tiled field mapping system, that can be more than 100 MB in size.

There are some situations in which true color provides the only means of adequately separating targets from background. Those dealing with stained histological material or multispectral fluorescence should give true color imaging serious consideration. In contrast, true color is rarely missed in densitometric applications, where its disadvantages usually outweigh its potential benefits. Weigh your needs carefully. Balance sacrifices in cost, resolution, and speed against the detection of targets by color.

Photography
Photographic prints or slides provide the best reproduction of images. Photographs can be taken directly from the image monitor, though the image quality is far lower than if a film recorder is used (see below). Instructions for photographing monitors are available in Kodak Publication No. AC-10, Photographing Television and Computer Screen Images. This pamphlet ($1.00 last time we checked) may be ordered from any camera store. The advice comes down to using an absolutely dark room, SLR camera, speed no faster than 1/8 sec, tripod mounting, cable release, and settings of 1/8 sec at f/5.6 for Ektachrome 100 (Daylight) or Kodacolor VR-G 100. It can also help to use a moderate telephoto lens, which flattens the otherwise curved face of the monitor screen.

A better option is to use a film recorder, which is convenient and which takes much better pictures. However, film recorders are costly, with prices ranging upwards from about $4,000.

Both analog and digital film recorders are available. An analog recorder accepts analog image data, and contains a small monitor with a color camera mounted in front of it. Analog film recorders are quick (picture times of 5 sec are typical), and do not require special software drivers. The major disadvantage of the analog film recorder is that it is limited to the resolution of the display monitor. You cannot print large images (e.g. 3000 x 2000 pixels) at full resolution, or use drivers that create high resolution text fonts or graphics.

Digital film recorders are like other digital printers, in that they accept data in digital form. The data stream is fed from a serial, parallel, SCSI or other digital interface, under control of the operating system or imaging system software. MCID™and AIS™support the AGFA ProSlide 35 and PCRII+, and the Polaroid digital film recorders. Digital film recorders provide flexible resolution (up to about 8000 x 8000 pixels on some models), and can print the largest images. This flexible resolution is very useful with the MCID/AIS large image format. Text and graphics will also have a noticeably smoother appearance when photographed at higher resolution.

Digital recorders can take anywhere from about 10 seconds to a minute to complete a picture, and very complex graphics can take longer. However, the flexibility and superior image quality of digital film recorders are major advantages, and we recommend them in almost every case.

Printers
As with film recorders, both analog and digital printers are available. Analog printers connect directly to the image monitor, and capture the image into an internal storage memory from which it is then printed. Analog printers offer reasonable image quality, fast printing, and low page costs. Therefore, they are popular in image documentation applications (e.g. gel documentation). Analog printers cannot print images larger than the display monitor, and cannot take advantage of digital techniques for resolution enhancement.

Digital image printers transfer data via a computer port. Popular printer technologies include ink jet, thermal transfer, laser, dye sublimation, and Pictrography. Most digital printers will print letter size images on standard paper (ink jet, thermal transfer, laser) or special paper (dye sublimation, Pictrography). Dye sublimation and Pictrography printers have higher print costs than other printers (about $1 - $4 /page).

The image quality from printers is highly variable. Color laser printers create images that look like faded photographs, in which most of the vibrant colors are lost. These images are fine for routine documentation, but are not publication-quality. Ink jet printers can produce vibrant colors, but are slow, require lots of attention (clog and smear), and the colors are less smooth than in a photograph. Dye sublimation prints are similar to a slightly faded and fuzzy 8 x 10 photographic print. Bright colors are handled well, but fine lines and lettering may be blurred. The best output we have seen is from the Fuji Pictrography printers, which can produce true publication quality. MCID™and AIS™support any of the hundreds of printers supported by Windows NT (e.g. monochrome and color lasers, Postscript dye sublimation). They also support the Sony UP-D family of digital thermal and dye sublimation printers, and the Fuji Pictrography printers.

Carothers, A.D. Counting, measuring, and mapping in FISH-labeled cells: Sample size considerations and implications for automation, Cytometry 16:298-304 (1994).

Elias, H. and Hyde, D.M. A guide to practical stereology. Karger, New York, (1983).

Feingold, E., Sashadri, S.B. and Tretiak, O.J. Hardware and software design considerations in engineering an image processing workstation: Autoradiographic analysis with DUMAS and the BRAIN autoradiograph analysis software package, in D.L. McEachron (ed.,), Experimental Biology and Medicine Vol. 11; Functional Mapping in Biology and Medicine, pp. 175-201, Karger, (1986).

Goochee, C., Rasband, W. and Sokoloff, L. Computerized densitometry and color coding of (14C)deoxyglucose autoradiographs, Annals of Neurology, 7 (1980) 359-370.

Grynkiewicz, G., Poenie, M. and Tsien, R.Y. A new generation of Ca2+ indicators with greatly improved fluorescence properties, Journal of Biological Chemistry 260:3440-3450 (1985).

Gundersen H.J.G.: Stereology of Arbitrary Particles, Journal of Microscopy 143:3-45 (1986).

Hitzemann, R., Qian, Y. and Hitzemann, B. Dopamine and acetylcholine cell density in the neuroleptic responsive (NR) and neuroleptic nonresponsive (NNR) lines of mice, The Journal of Pharmacology and Experimental Therapeutics 266:431-438 (1993).

Huijsmans, D.P., Lamers, W.H., Los, J.A. and Strackee, J. Toward computerized morphometric facilities: a review of 58 software packages for computer-aided three-dimensional reconstruction, quantification, and picture generation from parallel serial sections, Anatomical Record, 216 (1986) 449-470.

Lear, J.L, Mido, K., Plotnick, J. and Muth, R. High-performance digital image analyzer for quantitative autoradiography, Journal of Cerebral Blood Flow and Metabolism, 6 (1986) 625-629.

Lear, J.L., Plotnick, J. and Rumley, S. Digital autoradiography: Design, development and evaluation of a solid-state image analyzer, Journal of Nuclear Medicine, 28, 218-222 (1987).

Potchoiba, M.J., Tensfeldt, T.G., Nocerini, M.R. and Silber, M.B. A novel quantitative method for determining the biodistribution of radiolabeled xenobiotics using whole-body cryosectioning and autoradiography, The Journal of Pharmacology and Experimental Therapeutics, 272:953-962 (1995).

Ramm, P. Low-cost, computer assisted densitometry for autoradiography: A practical guide to system components costing less than $10,000, Society for Neuroscience Abstracts, 11 (1985) 1087.

Ramm, P. Image analysis for bioscience, Computerized Medical Imaging and Graphics, 14 (1990) 287-306.

Ramm, P. Advanced image analysis systems in cell, molecular and neurobiology applications, Journal of Neuroscience Methods 54:131-149 (1994).

Ramm, P., Kulick, J.H., Stryker, M.P. and Frost, B.J. Video and scanning microdensitometer-based imaging systems in autoradiographic densitometry, Journal of Neuroscience Methods,11 (1984) 89-100.

Ramm, P. and Kulick, J.H. Principles of computer-assisted imaging in autoradiographic densitometry. In R.T. Mize (Ed.), The Microcomputer in Cell and Neurobiology Research, Elsevier, Amsterdam, (1985), pp. 311-334.

Serra, J. Image Analysis and Mathematical Morphology, Academic Press, London, (1982).

Weibel, E.R. Stereological Methods, Vol. 1: Practical Methods for Biological Morphometry. Academic Press, New York, (1979).