WFC imagery from bead focus pan.
Figure 6 shows an image gallery of 18 beads comprised of two beads selected consecutively from each interval of the focus pan. Darkfield (blue) and fluorescence images (green) are shown for each bead with a gallery of standard imagery on the left and EDF imagery on the right. The EDF imagery clearly maintains a much higher degree of focus over the pan range. Focus blur is present in the standard imagery at ±2 μm of defocus in both darkfield and fluorescent images (objects 701, 702, 1102, 1103). At ±4 μm of defocus blurring is significant (objects 500,501, 1,306, 1,307) and the bead imagery exhibits a marked decrease in peak intensity as well as a large change in apparent area. By ±8 μm of defocus, bead images (objects 100, 101, 1,707, 1,708) become difficult to discriminate from the background.
Figure 6. ImageStream standard (left panel) and EDF (right panel) imagery collected over a 16 μm focus pan. Left and right panels show both darkfield (blue) and fluorescence (green) imagery. Standard image collection exhibits significant blur while EDF imagery maintains size and intensity throughout the entire 16 μm focus pan.
Download figure to PowerPoint
In marked contrast, the WFC-based EDF imagery maintains consistent image characteristics throughout the focus pan with both the bead area and intensity remaining relatively constant throughout the pan. This is particularly evident in the green fluorescent bead imagery in the right hand column. There are some artifacts present at higher levels of defocus in the form of horizontal and vertical lines emanating from the primary bead image and directed toward the top and right hand side of the page. The artifacts largely resemble those generated in the simulation and exhibit much lower intensity than the primary image. These artifacts are a result of slight changes to the PSF with focus and can be minimized with optimizations of the WFC element and deconvolution kernel. Modeling of the nonorthogonal nature of the artifacts has shown that they are also due in part to residual uncorrected spherical aberration in the optical system. The darkfield imagery appears similar in nature to the fluorescence imagery, however, it exhibits stronger deconvolution artifacts especially at high levels of defocus. This may be due in part to the fact that the deconvolution kernel was generated from the fluorescent imagery. Future optimizations will include channel specific kernels, balancing of in focus and out of focus imagery for kernel generation, elimination of residual spherical aberration and optimized WFC waveforms; all of which will reduce artifacts.
Figure 7 provides a quantitative analysis of the entire image set from which the imagery in Figure 6 was selected. In this Figure the peak pixel intensity and area of each object is plotted vs. object number for both the standard and EDF image sets where each dot in a plot represents a single bead. A total of 1,800 objects were imaged, consisting of approximately 200 objects acquired at each of nine focal positions with each focal position separated by 2 μm. Object 1 and object 1,800 are therefore spaced 16 μm apart with the best focus position corresponding to object numbers in the range of 800–1,000. The best focus position was gated using the regions “In Focus pk” or “Infocus Area” with the accompanying mean values for the gated data shown below the plot. In a similar manner, beads from the ±8 μm focus positions were also gated.
Figure 7. Comparison of Peak Intensity and Area feature variation for standard and EDF imaging modes of 2.5 μm beads collected over a 16 μm focus range. Dot plots show feature values (FL1 Peak Intensity or Area) plotted against focus position (object number). EDF image collection demonstrates 7–14 times less variation over the tested focal range.
Download figure to PowerPoint
Referring to the upper plots in Figure 7, the standard image set (left) exhibited close to a 14-fold decrease in average peak intensity (641 counts vs. 44 counts) between the best focus and the “8 μm defocus” positions. In contrast the EDF plot (right) showed only a ∼2-fold decrease in average peak intensity (613 counts vs. 287 counts) over the same focus pan. Allowing for the increased focal range and the larger bead size, these results were consistent with the theoretical models for the behavior of peak intensity.
Referring to the lower plots in Figure 7, the fluorescent EDF imagery exhibited a consistent area of approximately 24 pixels throughout most of the focus range, rising to 32 pixels at the +8 μm defocus position. In contrast, the standard imagery exhibited an increase in area of almost 14× from 32 pixels to 437 pixels at −8 μm of defocus. Using the area and peak intensity as figures of merit, the ImageStream with EDF imaging demonstrates seven to fourteen times better feature consistency through a depth of field covering the majority of most prokaryotic and eukaryotic cell diameters.
A statistical analysis of the noise contained in the imagery was performed by evaluating the standard deviation in background signal outside the bead image for each individual object. An analysis of over 2100 objects for each focus pan indicates the median standard deviation in background signal is 0.97 and 1.30 counts respectively for the standard and EDF focus pans (σ listed in upper plots of Fig. 7). The increase in noise of 0.33 counts will degrade the Signal to Noise ratio and therefore negatively impact the sensitivity of the instrument. However, in comparison to standard imaging, the degradation will be greatest for objects located at the best plane of focus. For objects located away from the best plane of focus, the increased signal maintained via EDF should more than offset the increase in noise. Sensitivity studies of the standard ImageStream demonstrate sensitivity superior to standard flow cytometry for fluorochrome conjugated 3 μm particles and indicate an ability to detect as little as 50 molecules of equivalent soluble fluorescein (17). Future work with EDF imaging will include a detailed sensitivity study of the EDF collection mode.
Comparison of EDF and standard imaging for chromosome enumeration.
Automated chromosome enumeration via FISH is an application for which EDF imaging may confer significant benefits. Defocus causes significant changes in the presentation of probes often blurring one into another or spreading out the signal to such a degree that it is difficult to automatically segment, or visually separate FISH probes from each other or from nonspecific background binding in the nucleus.
To compare the efficacy of chromosome enumeration between the standard and EDF ImageStream configurations, a sample of Jurkat cells was hybridized to a Chromosome Y probe as described in the Materials and Methods section. The results were automatically analyzed to enumerate copies of the Y chromosome in each cell. Simple classifiers using brightfield imagery were developed to exclude cellular debris, doublet events and other artifacts from the analysis. A specialized segmentation routine and connected components analysis was performed on the fluorescence imagery to generate a first pass enumeration of chromosomes on single cells (18, 19). A refinement of the monosomy and disomy classified cells from the first pass enumeration was performed to eliminate false positive events. After the final classification step, the resulting imagery was manually reviewed to qualitatively judge the efficacy of the final classification. This analysis was not intended to be a rigorous examination of the efficacy of the ImageStream with EDF for chromosome enumeration. Rather, this experiment is an extension of the previous work discussed in this paper and was performed to explore an application for which the ImageStream with WFC-EDF imaging may have a beneficial result.
After preparation the sample was loaded into an ImageStream system with a prototype WFC module enabling the EDF collection mode. A file containing 1,000 cells was collected in the standard configuration. The ImageStream was reconfigured for EDF collection and an additional 1,000 cell file was collected. The entire collection time for both files was several minutes (including the time switch from standard to EDF modes). The files were then analyzed using the IDEAS cell analysis program to enumerate chromosomes in each cell in each sample. Visual observation of the resulting monosomy and disomy populations was used to confirm automatic classifier results and to gain insights into the cause of performance differences between the two collection modes.
Figures 8–12 present a brief overview of the analysis of these files. Figure 8 shows a sampling of 20 cells (10 in each collection mode) highlighting some of the challenges in the classification of disomies within the Jurkat sample. The FISH probe imagery is superimposed over a reduced contrast brightfield image of the cells to provide a sense of scale and verify that the probes are located within the cell. In the upper panel, cell imagery, collected in the standard imaging mode, shows that at least one of the FISH probes is positioned out of the plane of focus. Consequently, the area of the out of focus probe increases and sometimes engulfs the second probe in the cell. Like the bead imagery shown in Figure 6, the intensity falls off significantly with defocus, making it difficult to automatically segment the probe or even see it by eye. In at least one case, object 916, two probes appear to be in close proximity. Slight defocus may have caused these relatively bright probe images to blur into each other, creating what appears to be a single large probe. Although it cannot be specified with certainty, this cell is thought to have two probes due to the total intensity of the probe signal and the elongated shape of the probe. In marked contrast, the EDF imagery presented in the lower half of the figure shows discrete FISH spots even when they are positioned in close proximity to each other. Unlike the standard collection mode, the EDF imagery exhibits no blurring of FISH spots. This is also readily apparent when comparing the larger selection of images shown in Figures 11 and 12. Each FISH labeled chromosome appears as a bright, tightly focused spot.
Figure 8. Standard and EDF ImageStream FISH imagery of cells with disomy for chromosome Y. Cells were selected from each run for display due based on the level of difficulty in chromosome enumeration due to presentation of the probes. In this imagery FISH probes are dispersed in the nucleus along the optic axis or in close proximity in the lateral direction. EDF improves image presentation by removing focus induced variation.
Download figure to PowerPoint
Figure 9. First and second steps in the process for classification and enumeration of chromosomes: elimination of cell fragments, aggregates and other artifacts via processing of brightfield imagery; initial segmentation and refinement of possible FISH probes in cells. Example imagery (extended depth of field) shown for various steps.
Download figure to PowerPoint
Figure 10. (Upper) Histograms showing first pass enumeration results for monsomy, disomy and polysomy cells for both standard and EDF collection modes. (Middle) Refinement of monosomy populations using peak intensity and area to eliminate false positive events. (Lower) Refinement of disomy populations to eliminate false positive events. EDF imaging isolates 2-fold more true disomy positive cells and provides clearer differentiation between true positive and false positive events.
Download figure to PowerPoint
Figure 11. Randomly selected cell imagery corresponding to “Monosomy Refinement” gates for standard and EDF image collection modes. EDF imagery (upper right) clearly shows single copy chromosome events. Probe imagery from the standard collection mode (upper left) is highly variable due to limited depth of field and random location of the probes along the optic axis.
Download figure to PowerPoint
Figure 12. Randomly selected cell imagery corresponding to “Disomy Refinement” gates for standard and EDF image collection modes. EDF imagery (upper right) clearly shows double copy chromosome events. Probe imagery from the standard collection mode (upper left) is highly variable due to limited depth of field and random location of the probes along the optic axis.
Download figure to PowerPoint
Development of FISH spot enumeration classifier and classification results.
To determine the efficacy of EDF imaging on the enumeration of chromosomes, a simple five step classifier was developed using the IDEAS analysis software. The first two steps involved segmentation and the selection of appropriate objects within the data file for subsequent analysis (16). Object selection was accomplished by plotting the brightfield aspect ratio vs. brightfield area as shown in the dot plot in Figure 9. A gate was drawn that encompassed primarily single cells (example object No. 424) and excluded cell fragments/debris (example object No. 845) and groupings of cells (example object No. 75). The gate defined a population named “Cells” containing 595 individual objects, to which subsequent analysis was applied. A similar segmentation and selection process was performed on the standard collection file and resulted in 588 individual objects.
The third step in classification, shown in the lower portion of Figure 9, involved refinement of the standard segmentation mask to isolate areas of local maxima in each fluorescence cell image. A fluorescence image of object 89, collected in Channel 3, is shown prior to segmentation (a), after initial segmentation (light blue overlay) to identify all areas containing light above background (b), and after Morphology segmentation (light blue overlay), a form of contour masking to identify areas of local maxima contained in the initial segmentation mask (c).
The fourth step in classification employed an IDEAS feature called “FISH Spots” which used the Morphology mask to perform a connected components analysis to enumerate discrete spots contained within each fluorescent image. The results of this computation and the final gating of disomic cells, the fifth step in the classification, are shown in Figure 10. As shown in the FISH Spot enumeration histograms at the top of Figure 10, the first pass analysis using the standard collection mode yielded enumerations of 506 monosomic, 67 disomic and 15 polysomic (three or more spots) cells. In contrast, the first pass enumeration with EDF imaging yielded 421 monosomic, 136 disomic and 38 polysomic cells. The EDF collection mode therefore produced a 2× increase in the number of disomy and polysomy-classified cells, with a 17% decrease in monosomic cells. Manual review of the monosomy and disomy-classified cells in both collection modes revealed a significant number of false classifications where hybridization had failed, leaving only nonspecific binding in the nucleus.
To improve classification accuracy, each population of monosomy and disomy-classified cells was further analyzed by plotting peak intensity vs. area for the fluorescence channel. Nonspecific binding generally has low peak intensity and large area and is therefore plots of peak intensity vs. area provide for good discrimination of nonspecific binding events. Bivariate plots of this analysis are shown in the middle and lower plots of Figure 10. The discrimination boundaries for the standard collection mode are not clear. This is most evident in the boundary drawn to discriminate true and false positives for disomy refinement (lower left plot of Fig. 10). The boundary is complex and arbitrary and unsuitable for fully automated classification. In contrast, the boundaries drawn for the EDF analysis are clear with the true and false positive populations showing excellent separation. Minor shifts in feature values due to preparation differences or instrument variations will not significantly affect the results of the classifications making these features and boundaries suitable for fully automated classification. The refined classifications result in 45 and 96 disomy events for the standard and EDF collection modes respectively and a respective 191 and 189 monosomy events.
Figures 11 and 12 display a random set of images from each set of “Refined” and “False Positive” populations defined in Figure 10. A review of this imagery sheds light on why the EDF collection mode exhibits a 2-fold increase in discrimination of disomic events. First, the EDF collection mode largely eliminated focus variation, producing small, bright, tightly focused spots for each hybridization event as evident in both “Refined” monosomy and disomy populations (upper left hand galleries in Figs. 11 and 12). Second, these tightly focused spots dramatically improved the performance of the morphological segmentation algorithm and final classification steps by forming a clear demarcation between well-hybridized probes and nonspecific binding events (center and lower right hand side bivariate plots of Fig. 10). The false positive events shown in the lower right hand side galleries of Figures 11 and 12 are a result of nonspecific binding and result in large segmentation masks with low peak intensities. The use of peak intensity and FISH Spot area effectively discriminates false positive events. By contrast, it is very difficult to discriminate between nonspecific binding and highly defocused probes as shown in the lower left hand galleries of Figures 11 and 12. Third, by eliminating focus variation, probes located away from the ideal focal plane still appear as small spots. This is in contrast to the probe imagery found in cell numbers 213, 228, 245, 255, 257, 275 etc. shown in the upper left hand gallery of Figure 11 or cell numbers 15, 251,279, 465 and 624 etc. in the upper left hand gallery of Figure 12. Tight focus substantially reduces the probability of events where a defocused probe image engulfs or contacts a second probe image in the cell. With standard imaging it is a rarity to find imagery similar to cells Nos. 55 and 247 from Figure 12, where two probes are in close proximity and tightly focused. More likely than not, one of these probes will blur, engulfing the other leading to a misclassification of a disomic event as a monosomic event.