MICROSCOPY RESEARCH AND TECHNIQUE 64:103112 (2004)

Colocalization Analysis Yields Superior Results After Image Restoration

LUKAS LANDMANN*
AND

PERMSIN MARBET

Department of Anatomy, University of Basel, CH-4056 Basel, Switzerland

KEY WORDS

3D-imaging; confocal microscopy; image processing; median ltering; deconvolution; signal-to-noise ratio; background; resolution; synthetic objects

ABSTRACT Colocalization analysis is a powerful tool for the demonstration of spatial and temporal overlap in the distribution patterns of uorescent probes. In unprocessed images, background affects image quality by impairing resolution and obscuring image detail in the lowintensity range. Because confocal images suffer from background levels up to 30% maximum intensity, colocalization analysis, which is a typical segmentation process, is limited to highintensity signal. In addition, noise-induced, false-positive events (dust) may skew the results. Therefore, suppression of background is crucial for this type of image analysis. Analysis of synthetic and biological objects demonstrates that median ltering is able to eliminate noise-induced colocalization events successfully. Its disadvantages include the occasional generation of false-positive and false-negative results as well as the inherent impairment of resolution. In contrast, image restoration by deconvolution suppresses background to very low levels (10% maximum intensity), which makes additional objects in the low-intensity but high-frequency range available for analysis. The improved resolution makes this technique extremely suitable for examination of objects of near resolution size as demonstrated by correlation coefcients. Deconvolution is, however, sensitive to overestimation of the background level. Conclusions for practical application are: (1) In raw images, colocalization analysis is limited to the intensity range above the background level. This means the higher the RS/N the better. Unfortunately, images of most biological specimens have a low RS/N. (2) Filtering improves the result substantially. The reduction of background levels and the concomitant increase of the RS/N are generated at the expense of resolution. This is a quick and simple method in cases where resolution is not a major concern. (3) If colocalization in the low-intensity range and/or maximum resolution play a role, deconvolution should be used. Microsc. Res. Tech. 64: 103112, 2004. 2004 Wiley-Liss, Inc. INTRODUCTION Multichannel uorescence microscopy makes use of multiple uorochromes of different excitation and emission wavelengths for localizing different probes in a single specimen. The technique allows the demonstration of spatial and temporal relationships in the distribution pattern of various probes and has found widespread applications, including the tracing of dynamic cellular processes. Colocalization of signal from different channels, i.e., the locations in an image where two or more uorochromes are found simultaneously, is of major interest in this type of work. Background, which is present in variable degrees in any image, is a major factor restraining the power of colocalization analysis. Therefore, it is highly desirable to increase the signal-to-noise ratio (RS/N), which, in a given image, can be achieved by ltering or image restoration. This review shows that ltering and image restoration by deconvolution substantially improve performance, accuracy, and resolution of colocalization analysis. Filter techniques enhance the RS/N at the expense of resolution by amalgamation of signal with background and noise. In contrast, removal of background by image restoration allows for colocalization at very low intensities and at very high resolution.

Due to its spatial resolution power, confocal microscopy is extremely well suited for colocalization analysis. However, wide-eld microscopy in combination with appropriate deconvolution procedures is also suitable and even superior to confocal microscopy in cases where optical properties are not the overriding criterion, e.g., in work with living cells that are susceptible to light damage. WHAT IS COLOCALIZATION ANALYSIS? Colocalization analysis in digital images is a typical dual-image pixel point process (Shotton, 1993), between a pair of input images from two different channels that generates a single output image, the colocalization map (two images in, one image out). This data set selects and segments all voxels where both channels have common events dened as signal intensity above a chosen threshold or within a certain range.

*Correspondence to: Lukas Landmann, Ph.D., Dept. of Anatomy, University of Basel, Pestalozzistr. 20, CH-4056 Basel. E-mail: Lukas.Landmann@unibas.ch Received 13 April 2004; accepted in revised form 14 May 2004 Abbreviations used: MIP, maximum intensity projection; PSF, point spread function; RS/N, signal-to-noise ratio DOI 10.1002/jemt.20066 Published online in Wiley InterScience (www.interscience.wiley.com).

2004 WILEY-LISS, INC.

104

L. LANDMANN AND P. MARBET

Therefore, colocalization analysis shares the following steps with other image analysis procedures: (1) (Optional) denition of a region of interest in order to restrict analysis to a spatially dened region of the image or to reduce requirements on processing hardware. (2) Denition of a threshold intensity value to separate features of interest from the rest of the image in each channel. This crucial operation has great inuence on the outcome and is normally done interactively by adjusting the chosen intensity to a value separating signal from background. (3) Segmentation of the input intensity images to generate an output binary image in which all voxels that fail to meet both threshold requirements are set to zero intensity and all voxels that pass are color-coded by a look-up table. In the gures of this review, voxels with a colocalization event are consistently pseudeocolored at maximum intensity in the additive color generated by the two original channels (e.g., green and red threshold yellow 255). (4) Computation of quantitative data such as co localizing voxel number or intensity in each channel. Additional computations can be performed by any tool provided by an image analysis programs. (5) Finally, the colocalization map can then be added to the original data set as separate channel(s) highlighting voxels meeting the conditions chosen for colocalization. There are various colocalization software modules available that help the user to analyze multichannel images by displaying a 2D histogram of pixel intensities for every singular image or for whole image stacks (Demandolx and Davoust, 1995). The histogram arranges corresponding pixels from two images of identical size but from different channels according to their red and green intensities along a green x- and a red y-axis. Thus, every singular pixel of an image is characterized by a pair of intensities that can be regarded as coordinates in a Cartesian system (Fig. 1). Analysis of the distribution pattern of the intensity pairs allows for the identication of colocalization as well as discrimination of background, crosstalk between channels, uorescence attenuation, and channel misregistration (Demandolx and Davoust, 1995, 1997). It is then the task of the user to dene in the histogram the area reecting colocalization by picking an adequate intensity threshold or range. The aim of colocalization analysis is, by denition, to include only those events generated by signal from the specimen and to exclude events contributed by unspecic background and noise. It cannot be overemphasized that this is the single most critical step of the procedure. Choosing too low a threshold includes high background intensities and results in the generation of false-positive colocalization events, whereas a conservative estimate, i.e., thresholding at too high an intensity, misses colocalization displayed in the specimen. LIMITATION BY BACKGROUND Colocalization analysis discriminates between signal- and background-generated events by thresholding at a chosen intensity level. Unfortunately, the image is not a perfect copy of the distribution of uorescent probes in the specimen (Fig. 2). Additional intensities originating from the effect of a non-spherical point spread function (PSF) or blur, from single photon hits or stray light, and from photomultiplier offset or base-

Fig. 1. Colocalization softwares analyze topographically corresponding voxels in images or image stacks from two channels. Every singular voxel is characterized by two (e.g., a green and a red) intensities that are used for arranging it in a 2D histogram along a green x- and a red y-axis. Close ups (top) display voxels with their green and red intensities. The highlighted voxel with an intensity of green168 and red135 occupies the encircled position in the histogram (bottom), which gives intensities as percentages. The white area reects thresholding at 25% maximum intensity in both channels. [Color gure can be viewed in the online issue, which is available at www.interscience. wiley.com.]

line, all of which are subject to random uctuations or noise, add up to background. They impair image quality by obscuring image detail in the low-intensity range and by affecting resolution. Because no information can be extracted from intensities below the background

COLOC ANALYSIS IMPROVED BY DECONVOLUTION

105

Fig. 2. Comparison of imaging modes and their intensity proles in the same optical section from a data set showing microtubules in a RHCC. The image is presented in unprocessed form (left), after median ltering (center), and after deconvolution (right). The graph (bottom) plots intensity against x-position (white line) for all three imaging modes. Images are in false colors to show the power of image processing. Structures that in raw images are barely recognizable become quite distinct after ltering and perfectly smooth with wellbalanced internal intensity gradients after deconvolution. Structural

improvement is paralleled by reduction of background and noise. Note scattered appearance of the raw image prole (open circles) with single photon hits (arrows) occurring not only in the low-intensity range but also distorting signal. Median ltering (blue dotted line) smoothes the prole at the expense of resolution. This is shown by the highest peak, which has a smaller half-width after deconvolution (red line) than after ltering. Note almost complete suppression of background by deconvolution.

level, the discriminating intensity threshold, ideally, should separate signal from background. In low-contrast biological specimens, background and noise can reach up to 30% maximum intensity even in confocal images (Fig. 2). Inclusion of such intensities results in the generation of unspecic, mainly single-voxel colocalization events. Therefore, colocalization analysis is restricted to the high-intensity range in raw images. In order to extend the intensity range available for analysis, image processing routines that improve RS/Nare highly recommended. SUPPRESSION OF BACKGROUND Filtering High-intensity background is successfully suppressed by low pass, median, and Gaussian lters (Demandolx et al., 1997). Filtering techniques recover a voxel value from the local neighborhood and compute a weighted average (see Fig. 2, intensity proles). This, inherently, results in loss of resolution because the procedure convolves the image with a smoothing ker-

nel. Therefore, background is merged with and distributed over signal, which, in extreme cases, can result in artifacts. False-negative (dissipation of signal intensity below threshold value) as well as false-positive (merger of background with low-intensity signal resulting in above threshold intensities) results can be generated. A great advantage of ltering techniques is their ease of handling and their speed. Experience demonstrates that median ltering can improve RS/N at least 3-fold (see Fig. 5). Deconvolution Deconvolution (Shaw and Rawlins, 1991; Van der Voort and Strasters, 1995) uses the imaging properties of the optical system in the form of the point spread function (PSF) for putting the light back where it is coming from. Imaging of an entire 3D object may be described as the convolution of this object with the PSF (Castleman, 1993). Therefore, the PSF can be used for calculation of a likely model of the object from the recorded data set in an iterative process. This proce-

106

L. LANDMANN AND P. MARBET

Fig. 3. Selection of threshold levels for colocalization. Histograms of 2-channel recordings of calibration specimens stained for the green (left) or red (center) channel only. The width of the pixel clouds, which are associated with the green or red axis, respectively, is a

guideline for threshold selection in the nal 2-channel specimen (right). Note that this test does not consider background caused by unspecic binding of primary antibodies. [Color gure can be viewed in the online issue, which is available at www.interscience.wiley.com.]

dure eliminates effectively blur caused by distortion. In addition, by assuming a Poisson distribution of stray light (Sheppard et al., 1995), it suppresses background to very low levels (Fig. 2). The technique yields images of appreciably increased contrast. Background levels at less than 10% maximum intensity increase the RS/N by a factor of at least 5 as compared to raw images. This makes additional objects in the low-intensity (but highfrequency!) range available to analysis. In addition, removal of noise and distortions induced by the optical system results in improved resolution, which is of critical importance in cases involving objects of near resolution size (see cytoskeleton below). With the incorporation of deconvolution algorithms into commercially available software modules (we used here the Huygens module, Scientic Imaging B.V., Hilversum, The Netherlands), a powerful tool for image restoration has become available. Disadvantages of the technique include the requirement of computing power, the exact estimation of background and RS/N (see below) and time. HOW TO CHOOSE A THRESHOLD? Discrimination of signal from background is not always easy and intensity distribution plots provided by many softwares do not help a great deal in most cases. With experience, the thresholds of two or more channels can be chosen interactively with great precision by picking a threshold for volume or surface rendering that is then applied to colocalization analysis. A more methodical way includes the adjustment of all photomultipliers to their respective channel intensities followed by the recording of control specimens probed with one uorochrome only as a multichannel data set. The histograms of this data set (Fig. 3) show a cloud of pixels associated with the corresponding axis. The thickness of the cloud yields the background intercept or threshold on the complementary axis. Processing of this data set results in threshold values suitable for processed images. Care must be taken not to miss the seemingly few pixels lying high above the cloud in the histogram.

EXAMPLES Rat Hepatocyte Couplets Isolated rat hepatocyte couplets (RHCC, Fig. 4) were prepared according to standard procedures (Coleman and Roma, 2000) and exposed for 15 minutes to TxR-conjugated dextran (red), a marker for uidphase endocytosis, andasialoglycoprotein linked to Cy5 (blue) that labels specically the lysosomal endocytic route. The green channel shows microlaments probed with phalloidin Alexa488. Like most biological specimens, this sample yields noisy images with an intermediate RS/N (Fig. 5, left group of bars) that was less than 20 in all channels. Because deconvolution requires the recording of images without truncation of the dynamic range, thus yielding images with less contrast than usual, the difference between raw and processed images seems greater than usual. Treatment with a 3 3 3 median lter resulted in an improvement by roughly factor 3, whereas the same data set after deconvolution displayed a RS/N improved 6-fold as compared to raw images. These data quantitatively support the impression that background levels are decreased considerably by ltering and image restoration. Therefore, intensity thresholds can be chosen at lower levels after image processing. In the raw image, thresholds were selected at 20% maximum intensity, whereas a value of 15% was chosen for the ltered image and 10% for the restored data set. Colocalization events, indicated by the pseudo-colors yellow, cyan, and magenta, respectively, are illustrated in a single optical section (Fig. 4, top row). Comparison of the resulting colocalization events shows that raw data display channel overlap in large structures of high intensity, the number and size of which are increased considerably by median ltering and drastically by deconvolution. Whereas an increased number of colocalization events is caused by lower threshold settings (20 vs. 15 vs. 10% maximum intensity), enhanced image detail is a benet of image processing techniques and the suppression of noise. This is demonstrated by

COLOC ANALYSIS IMPROVED BY DECONVOLUTION

107

Fig. 4. Isolated rat hepatocyte couplets (RHCC) were incubated for 15 minutes in medium containing 7.5 mg/ml dextran conjugated to TxR (red), a marker for uid-phase endocytosis, and 0.01 mg/ml asialoglycoprotein linked to Cy5 (blue) that labels specically the lysosomal endocytic route. After xation and permeabilisation, cells were probed with phalloidin-Alexa488 (green) for microlaments. A singular optical section (top) is shown as raw image (left), after ltering (center), and after deconvolution (right). Colocalization thresholds were selected at 20%

maximum intensity in the raw image, at 15% for the ltered, and at 10% for the restored data set. Colocalization maps are color-coded in additive colors at maximum intensity. Bottom row: Close-ups from a different image of the same stack. Unconnected colocalizing voxels in the raw image (overlap of green and red yellow) become patches after processing with a median lter. Deconvolution clearly displays overlap of green and blue (cyan) and red and blue (magenta), which is not detected in the ltered image.

Fig. 5. RS/N was calculated as ratio of maximum intensity divided by average background as estimated by the Huygens software in a non-object volume (radius 0.5 m, axial size 0.3 m) of the data set, because the more accurate method of Sheppard (Sheppard et al., 1995), which is more adequate in low-contrast confocal images, cannot be applied to processed images. The groups of bars at the left-hand side are from the RHCC image presented in Figure 4, those to the right from the broblast in Figure 6. Note differential increase of RS/N after image processing (see text). Green, red, and blue channels are indicated as G, R, and B, respectively.

Although a conservative threshold above the noise level apparent in the histograms has been selected, the occurrence of noise-generated colocalization events is evident. This makes discrimination of artifact from real overlap difcult. After processing with a median lter and application of a lower intensity threshold, most isolated colocalizing voxels disappeared and channel overlap was clustered into distinct patches. After deconvolution, the data displayed enhanced image detail and were virtually free of singular voxel events, although thresholds were chosen at even lower intensity. This results in additional colocalization events of low intensity that were not detected by ltering (Fig. 4, cyan and magenta colocalization signal). On the other hand, a major part of green-red overlap (yellow), which is indicated by singular voxels in raw images and patches after ltering, can no longer be detected after image restoration. This demonstrates that the colocalization pattern may differ positively as well as negatively in a data set before and after image processing, although the general trend shows an increase of colocalization events in ltered and even more so in deconvolved images as compared to unprocessed data. Fibroblasts Rat SM2 broblasts (Fig. 6) were grown on coverslips to half-conuence and were probed after xation for actin with phalloidin-Alexa488 (green), for microtubules by in-

close-up views of a different optical section from the same data set (Fig. 4, bottom row). Areas of intermediate intensity in raw images display colocalization events consisting predominantly of isolated (unconnected) voxels.

108

L. LANDMANN AND P. MARBET

Fig. 6. Rat SM2 broblasts were probed after xation for actin with phalloidin-Alexa488 (green), for microtubules by indirect immunouorescence (Cy5, blue), and for intermediate laments with an antibody against vimentin that has been conjugated to Cy3 (red). Unprocessed data (left), ltered (center), and deconvolved (right) images are presented with their colocalization maps in the simulated

uorescence process (top row). Singular optical sections (bottom row) show that irregular overlap of raw images is clustered by ltering but not always detected by deconvolution. This is caused by an object size that is smaller than the resolution limit of the instrument (see text).

direct immunouorescence (Cy5, blue), and for intermediate laments with an antibody against vimentin that has been conjugated to Cy3 (red). This specimen has a similar contrast as the former one with a RS/N between 10 and 20 in raw images (Fig. 5, right group of bars). Filtering improves contrast approximately 6-fold to levels twice as high than those in RHCC, whereas deconvolution results in a 2-fold increase as compared to ltered images. Volume rendering by simulated uorescence process (Fig. 6, top row) demonstrates that image processing results in increased RS/N. Similar to RHCC, in unprocessed images colocalization events (Fig. 6, bottom row, close-ups from a single optical section) are made up to a large proportion of singular voxels the number of which is not altered substantially after ltering but decreased drastically after deconvolution. In this specimen, however, colocalization analysis of deconvolved images yields results different from those in RHCC although thresholds were lowered to a similar degree (30 vs. 20 vs. 10% maximum intensity): Whereas channel overlap is increased after ltering, deconvolution results in a marked decrease. This is shown by the close-up views that display a seemingly unambiguous colocalization in raw images becoming coarser after ltering and disappearing over wide areas after deconvolution. Quantitative Differences Voxel Number, Sum of Intensities, and Number of Objects. Quantitative data from cytoskeletal elements in broblasts differ markedly from those obtained

in RHCC. Whereas the number of colocalizing voxels shows only slight alterations after ltering and a strong decrease after deconvolution, total intensities are decreased slightly after ltering and strongly after deconvolution. In contrast, object numbers show a decrease similar to that in RHCC. This discrepancy is explained by the size of objects displaying colocalization: Structures in this specimen are smaller than the resolving power of the instrument. Because F-actin (green, 6 nm), vimentin (red, 10 nm), and microtubules (blue, 24 nm) are smaller than the sampling unit (voxel size 50 50 100 nm), the probability is high that cytoskeletal elements occur in the same voxel accidentally without necessarily being associated. This occurrence generates false-positive colocalization events that even deconvolution cannot suppress. If, however, objects are present in closely associated voxels, the probability is high that straylight generates falsepositive events in raw and ltered but not in deconvolved images. Channel Correlation. Channel correlation functions compute a number of coefcients that characterize the degree of overlap between channels in an image. Pearsons linear correlation coefcient (Gonzalez and Wintz, 1987; Manders et al., 1993) can be used to estimate the overlap of pixels. Its value is between 1 and 1. A value of 1 represents perfect correlation, 0 no correlation, and 1 perfect inverse correlation. Because average intensity in both channels is a factor, it is independent of image background. In contrast to the 2D histogram, the coefcient does not take into account

COLOC ANALYSIS IMPROVED BY DECONVOLUTION

109

Fig. 8. Channel correlation coefcients in colocalized volume of RHCC (left) and broblasts (right). In RHCC, correlation of the 3 channel pairs decreases after ltering. After deconvolution, an increase is displayed only by the red-blue pair. This demonstrates a genuine overlap of dextran (red) and ASGP (blue) both of which are endocytic markers. Association of these probes with actin (green) is caused predominantly by image defects as demonstrated by coefcients decreasing with increasing image quality. Cytoskeletal elements in broblasts show coefcients that in raw images are lower than in RHCC. The consistent increase after ltering is caused by dissipation of signal over more voxels. All channel pairs show a decrease after deconvolution to values lower than those in raw images, which indicates that colocalization effects are due predominantly to image defects.

Fig. 7. Effect of ltering and deconvolution on number (top), sum of intensities (center) of colocalization events (voxels), as well as on number of objects (bottom) in each channel pair (GR green and red; GB green and blue; RB red and blue) of the data sets shown in Figure 4 (RHCC) and Figure 6 (broblasts). Values are normalized to raw images as 100%. Numbers of colocalizing voxels and their intensities are increased moderately after median ltering and strongly after deconvolution in RHCC whereas in broblasts intensities and voxel numbers show little change after ltering, but are decreased after deconvolution. This discrepancy is explained by the size of colocalization events, which in RHCC are larger and in broblasts smaller than the resolution limit. Consequently, image processing, which allows the recruitment of voxels in the low-intensity but high-frequency range, results in an increase in RHCC. In contrast, overlap of cytoskeletal elements in broblasts, which is due to diffraction-caused distortion, stray light, and blur, is apparent in raw and ltered images. Deconvolution avoids this artifact by its superior resolution power and its ability to suppress background. Consequently, voxel and object numbers as well as intensities are lowest in this imaging mode. Analysis of object numbers, with values normalized to objects 1 voxel (connected lled bars) in raw images, shows that colocalization maps of raw images contain many singular voxels (open bars) displaying an event that reects high-intensity noise although thresholds were chosen conservatively. This artifact is suppressed to a satisfactory degree by ltering and almost completely by deconvolution. In RHCC, object numbers are decreased by ltering to a level that is not decreased further by deconvolution. In contrast, object numbers in broblasts that contain predominantly colocalization events of subresolution size are decreased additionally by deconvolution.

spatial relations. Therefore, it can be calculated in the entire data set volume or in the colocalized volume only. Figure 8 shows that channel correlation differs markedly in the two 3-channel images presented in Figures 4 and 6. In RHCC, correlation of the 3 channel pairs in the colocalizd volume decreases after ltering consistently and, after deconvolution, increases in the red-blue pair only. This demonstrates that dextran (red) and ASGP (blue) are the only channels displaying genuine overlap in endocytic vesicles, whereas association of these probes with actin (green) is coincidental. In contrast, channel correlation of cytoskeletal elements in broblasts is lower in raw images, increased constantly after ltering, and decreased after deconvolution to coefcients lower than those in raw images. Whereas the ltering-induced increase is due to dissipation of signal over more voxels generating more false-positive overlap, the decrease associated with deconvolution reects the higher resolution of this method resulting in less false-positive overlap. Thus, channel correlation, too, demonstrates the superior resolution of image restoration techniques as compared to ltered or raw data. Effects of RS/N and Background on Deconvolution. Although our data clearly demonstrate that deconvolution results in an improved resolution, it should be borne in mind that this technique depends crucially on the settings of RS/N and background (Landmann, 2002). Over- and underestimation of both parameters alters the results seriously. Overestimation of RS/N and background causes distortions that are greater than those caused by underestimation, with overestimation of background resulting in the worst alterations in low-contrast specimens. This means for practical applications that correct estimation of background is more important than that of

110

L. LANDMANN AND P. MARBET

Fig. 9. Virtual experiments with synthetic spheres. Computer-generated spheres (left) with a radius of 800 nm (top row, 2D histograms) and 50 nm (middle row, colocalization highlighted in yellow) were convolved with a synthetic PSF followed by degradation with articial Poisson noise (raw images, center left). Raw images were processed with a median lter (center right) or restored by deconvolution (right). The graph (bottom left) shows the number of voxels (black bars) and sum of intensities (grey bars) of the 800-nm sphere. Colocalizing volume of 50-nm spheres is shown in the graph at bottom right. The larger sphere (top) is characterized by two channels in perfect register that give rise to a diagonal line in the histogram (synthetic image). Articial distortion and noise (raw image) widen the line slightly and give rise to two noise-induced accompanying lines, 10% of maximum intensity (noise

level) apart. These lines are suppressed by ltering (median), a process dissipating not only background but also peak intensities of the object as demonstrated by the lack of signal in the upper right-hand corner of the histogram. Deconvolution, nally, is also able to remove background and yields a thinner and continuous line of overlap. A sphere with a radius of 50 nm, the channels of which were shifted by 100 nm along the x-axis, shows no colocalization (yellow) in the synthetic image (center left) while overlap is present in degraded raw images, ltered and deconvolved data stacks. Colocalizing volume is greatest in ltered and smallest in deconvolved images demonstrating that at near-resolution conditions, deconvolution is not able to restore the original situation completely. The technique, however, has the capacity to improve image quality considerably.

RS/N and that a conservative approach is less likely to introduce artifactual alterations. Synthetic Experiments The conclusion that ltering but not deconvolution impairs signal by the merger of background and noise is not easy to demonstrate. Objects yield distorted images that do not faithfully reect the real specimen. In

contrast, computer-generated objects have known properties and can be compared with their images after appropriate processing. Therefore, virtual experiments were conducted (Huygens software package, SVI B.V., Hilversum, The Netherlands). Synthetic spheres with appropriate excitation and emission wavelengths were generated in a matrix consisting of 50 50 100 nm voxels, i.e., with a sampling density

COLOC ANALYSIS IMPROVED BY DECONVOLUTION

111

corresponding to the RHCC and broblast specimens. The objects were convolved with a theoretical PSF, degraded by articial Poisson noise assuming the realistic values of 100 photons per unit and 10% maximum intensity, and joined to a 2-channel data set. Images with complete channel overlap were ltered and restored as before, and analyzed for colocalization using a threshold of 50% background-corrected maximum intensity, i.e., half-maximum amplitude width.

The data for spheres with a radius of 800 nm (Fig. 9), which is well above the resolution limit of the instrument, demonstrate that synthetic spheres are characterized by a volume that compares favorably to theory (98%). In raw images, voxel numbers that are similar to those of convolved but smaller than those of virtual spheres are obtained. This discrepancy is explained by the convolution procedure that spreads light and affects low-intensity signal more than peak values, thus clipping the spheres at the periphery. Sum of intensities in raw images corresponds fairly well to the added values of convolved spheres and Poisson noise. Voxel numbers as well as sums of intensities are moderately increased after ltering and decreased after deconvolution, an effect of the constant threshold values that prevent the recruitment of low-intensity signal in processed images. In contrast to raw images, ltered data sets show intensity maxima that are 10% lower, a difference that is slightly less than the chosen background level. This demonstrates that removal of background by ltering is achieved by dissipation of onevoxel noise over larger volumes, a process that inherently decreases absolute maximum values. However, this is not true for deconvolution, which is independent of the original intensities and distributes its results over the full dynamic range. Resolution power was tested on spheres of subresolution size with a radius of 50 nm, the two channels of which were shifted by twice the radial distance along the x-axis in order to get apposed but not overlapping spheres. Using a threshold at half-maximum intensity, colocalization directly reects impaired imaging quality. Figure 9 shows that synthetic objects display no overlap, whereas both convolution with a PSF and background sum up to substantial overlap in raw images. Deconvolution improved this defect by more than 50%, whereas ltering resulted in images of distinctly lower resolution power. It should be emphasized that

Fig. 10. Endocytosis of dextran in RHCC. A 5-minute pulse of the uid-phase marker dextran-TxR (red) (7.5 mg/ml medium) was followed by a maker-free chase of various time intervals. After xation, specimens were immunostained for asialoglycoprotein- and immunoglobulin A-receptors (ASGP-R, pIgA-R). Panels show a singular optical section (left) and a surface rendering (right). Endocytosis of dextran for short time intervals (2.5 minutes, top) labels predominantly a tubulo-vesicular network subjacent to the basolateral membrane domain. After prolonged endocytosis (30 min, center and bottom), dextran becomes concentrated in vesicles close to the nucleus and the apical membrane. Quantitative analysis of dextran-positive volume (graph at bottom) demonstrates a rapid uptake during the rst 2.5 minutes, which is followed by a plateau up to 10 minutes and a subsequent decrease. After short time intervals, both receptors (blue) display a similar colocalization pattern with dextran (magenta) at the basolateral membrane and in the tubulo-vesicular network (top, pIgA-R). This pattern becomes weaker after longer time periods and is changed in a receptor dependent mode. Overlap of ASGP-R with dextran is found predominantly in perinuclear vesicles (center), whereas colocalization with pIgA-R is shifted to the apical cell pole (bottom). This is paralleled by quantitative data (graph at bottom) that show that overlap of dextran and both receptors increases during the pulse phase and decreases thereafter. Differences between the receptors include a slightly later peak for pIgA-R and different volume fractions after longer (15 minute) time intervals. Different kinetics reect the quick internalization and recycling of ASGP-R, which mediates uptake of compounds destined for lysosomal degradation, and the vectorial itinerary of pIgA-R that stays associated with its ligand during the transcytotic transport of pIgA, respectively.

112

L. LANDMANN AND P. MARBET

in borderline situations even deconvolution is not able to restore the original situation as shown by the smaller but still persisting overlap. The data demonstrate that in high-resolution work, deconvolution has the capacity to improve image quality to some degree but not completely. Endocytosis in RHCC Colocalization analysis in combination with deconvolution allows not only the elucidation of 3D relationships in static specimens, it is also a powerful tool for the examination of dynamic processes as well. In order to characterize endocytic pathways in RHCC, a 5-minute pulse of the uid-phase marker dextran-TxR was applied followed by a maker-free chase of various time intervals. After xation, specimens were immunostained for asialoglycoprotein- and immunoglobulin A-receptors (ASGP-R, pIgA-R). Internalization of dextran for short time intervals (2.5 minutes) resulted in the labeling of a tubulovesicular network close to the basolateral membrane domain (Fig. 10). After prolonged endocytosis (30 minutes), dextran became concentrated in vesicles close to the nucleus and the apical membrane. Quantitation of dextranpositive volume showed a rapid uptake during the rst 2.5 minutes, which was followed by a plateau up to 10 minutes and a subsequent decrease. ASGP-R mediates uptake of compounds destined for lysosomal degradation (Spiess, 1990; Steer and Ashwell, 1990) and has a steady-state distribution in the Golgi as well as in and beneath the basolateral plasma membrane. Initially, it colocalized with dextran at the basolateral membrane and in the tubulo-vesicular network. A shift into perinuclear vesicles became prominent with progressing time. This is consistent with the view that ASGP-R dissociates from its ligands gradually during its transport from the early sorting to the late endosomal compartment and is recycled back to the cell surface. In contrast, pIgA-R, which mediates the transcytotic delivery of IgA into bile (Brown and Kloppel, 1989) and is expressed in various vesicular compartments distributed throughout the cytoplasm, showed colocalization patterns with dextran similar to ASGP-R at early time points. After longer time intervals, overlap disappeared at the basolateral cell pole and became concentrated in an apical tubulo-vesicular network. Quantitative analysis of overlap of dextran and both receptors showed an increase during the pulse phase followed by a sharp decrease. The two receptors differed in their kinetics with pIgA-R peaking slightly later than ASGP-R and in the volume fraction preserved after longer (15 minute) time intervals. This discrepancy reects differences in the itinerary and rate of the receptors and agrees well with the extremely high afnity of ASGP-R for its ligands. CONCLUSIONS The data show that image processing by median ltering and deconvolution successfully suppresses background and noise induced single voxel colocalization events that clutter raw images. Median ltering allows for the selection of lower thresholds by dissipation of high-intensity background. The inherent attenuation of signal may lead to failures in detecting small and/or low-intensity objects as well as to false-positive responses if small objects of high intensity are examined. Image restoration by deconvolution is able to

suppress noise more effectively than ltering, thus making low intensities available for analysis as well as small objects. This results in a greater intensity range and superior resolution. A take-home message can be summarized as follows: In raw Images colocalization analysis requires a high RS/N and yields no information in the low-intensity range. Because most biological specimens do not belong to this category, results will benet from image processing techniques. Median ltering recovers voxel values from the local neighborhood and computes a weighted average. The technique improves the results by decreasing background levels, thus increasing the RS/N. This advantage is associated with loss of resolution. Filtering, therefore, is a quick and simple method for studies not primarily concerned with high resolution. Deconvolution restores the effects of factors resulting in impaired image formation and removes background almost entirely. This method is indicated for colocalization analysis at maximum resolution and/or in the low-intensity range. It involves, however, efforts in image acquisition and processing. ACKNOWLEDGMENTS We gratefully acknowledge the skillful technical assistance of Mireille Toranelli, Jean-Paul Boeglin, and Eva Bryson. REFERENCES
Brown WR, Kloppel TM. 1989. The liver and IgA: Immunological, cell biological and clinical implications. Hepatology 9:763784. Castleman KR. 1993. Resolution and sampling requirements for digital image processing, analysis, and display. In: Shotton DM, editor. Electronic light microscopy. The principles and practice of videoenhanced contrast, digital intensied uorescence, and confocal scanning light microscopy. New York: Wiley-Liss. p 7193. Coleman R, Roma MG. 2000. Hepatocyte couplets. Biochem Soc Trans 28:136 140. Demandolx D, Davoust J. 1995. Multicolor analysis in confocal immunouorescence microscopy. J Trace Microprobe Tech 13:217225. Demandolx D, Davoust J. 1997. Multicolour analysis and local image correlation in confocal microscopy. J Micr 185:2136. Demandolx D, Barois N, Davoust J. 1997. Guidelines for multiuorescence confocal imaging: Acquisition, processing and display. Eur Microsc Anal 7:57. Gonzalez RC, Wintz P. 1987. Digital image processing. Boston, MA: Addison Wesley Publication Company. Landmann L. 2002. Deconvolution improves colocalization analysis of multiple uorochromes in 3D confocal data sets more than ltering techniques. J Microsc 208:134 147. Manders EMM, Verbeek FJ, Aten JA. 1993. Measurement of colocalization of objects in dual-colour confocal images. J Microsc 169:375382. Shaw PJ, Rawlins DJ. 1991. The point spread function of a confocal microscope: its measurement and use in deconvolution. J Microsc 163:151165. Sheppard CJR, Gan X, Gu M, Roy M. 1995. Signal-to-noise in confocal microscopes. In: Pawley JB, editor. Handbook of confocal microscopy, 2nd ed. New York: Plenum Press. p 363371. Shotton DM. 1993. An introduction to digital image processing and image display in electronic light microscopy. In: Shotton DM, editor. Electronic light microscopy. The principles and practice of videoenhanced contrast, digital intensied uorescence, and confocal scanning light microscopy. New York: Wiley-Liss. p 39 70. Spiess M. 1990. The asialoglycoprotein receptor: a model for endocytic transport receptors. Biochemistry 29:10009 10018. Steer CJ, Ashwell G. 1990. Receptor-mediated endocytosis: mechanisms, biologic function, and molecular properties. In: Zakim D, Boyer TD, editors. Hepatology. A textbook of liver disease, vol. 1. Philadelphia: W.B. Saunders Co. p 1:137182. Van der Voort HTM, Strasters KC. 1995. Restoration of confocal images for quantitative image analysis. J Microsc 178:165181.