1.

Introduction

There are many areas of biomedical research where detection and enumeration of rare circulating cells in the bloodstream of small animals are important, including cancer metastasis, hematological malignancies, immunology, and reproductive medicine.^1–7 The classical method for enumerating circulating cells is to draw small peripheral blood (PB) samples from the animals, which are subsequently analyzed with techniques such as flow cytometry, fluorescence microscopy, or more recently with microfluidic devices.^8–12 However, this “ex-vivo” approach is known to be fundamentally problematic, since handling, storage, and enrichment of PB samples have been shown to affect cell viability and behavior.^13–16 These limitations motivated the development of “in vivo flow cytometry” (IVFC) technology by Lin et al. in 2004,^17,18 wherein circulating cells are optically counted without the need for drawing blood samples. Briefly, in IVFC, a laser slit illuminates a small blood vessel in the ear or retina of a mouse. As fluorescently labeled cells pass through the field of view (FOV), fluorescent “spikes” are detected by the instrument. As such, circulating cells can be noninvasively and continuously counted in vivo allowing, for example, detection of rapid changes in circulating cell populations over time in response to drug treatment.^19–21 Other IVFC variants have also been developed, including multiphoton IVFC, photo-acoustic flow cytometry, and photo-thermal flow cytometry.^22–30 IVFC has been used for the in vivo study of many cell types, including breast cancer, prostate cancer, hepatocellular carcinoma cells, melanoma, multiple myeloma, T-lymphocytes, infections, and sickle cells.^{18,21,22,28–36}

Despite its great utility in small animal research, IVFC has one significant limitation: the relatively low cell detection sensitivity, which is a consequence of the small blood vessels that are sampled. Specifically, blood flow rates in an ear arteriole are on the order of $1\mu \mathrm{L}/\min$, whereas a mouse has approximately 2 mL of circulating blood. Therefore, in a 15-min scan, only about 1% of the total circulating blood volume is sampled, so that low-abundance cell types may be entirely missed. Motivated by this limitation, our group (and others^31,32) has developed new IVFC approaches that interrogate larger circulating blood volumes and, therefore, increase detection sensitivity.^37–41

Most relevant to the present work, we recently developed and validated a technique termed Computer Vision IVFC (CV-IVFC).⁴⁰ Briefly, we developed an instrument that utilizes macroscopic fluorescence imaging of a relatively large ($5\times 5{\mathrm{mm}}^2$) region of a mouse ear, where circulating blood volumes are on the order of $10–15\mu \mathrm{L}/\min$, i.e., an order of magnitude greater than microscopy-IVFC. The imager used a red (660 nm) excitation laser and a filtered high-sensitivity electron multiplied charge-coupled device (EMCCD) camera that acquired fluorescence images with a 19 Hz frame rate. As fluorescently labeled circulating cells passed through the ear, they were detected by the camera. We also developed an algorithm to automatically detect and track circulating cells from the image sequences. This was necessary for three reasons: (1) the rarity of the low-abundance cells that we were studying (i.e., the infrequency of arrival in the instrument FOV), (2) the large FOV (relative to cell size), and (3) the relatively large background autofluorescence and camera noise. We validated CV-IVFC, first in flow-phantom models with calibrated microspheres, and second in mice with injected multiple myeloma (MM) cells labeled with the vybrant-DiD near-infrared fluorescent dye. We demonstrated that CV-IVFC was capable of detecting and tracking circulating cells at concentrations of $20\mathrm{cells}/\mathrm{mL}$ of circulating blood volume in a 30-min scan. This is potentially extremely valuable for applications involving rare circulating cells; for example, in measurement of the dissemination of circulating tumor cells in animal models of cancer metastasis where typical cell concentrations are on the order of $1–100\text{cells}/\mathrm{mL}$. Moreover, unlike microscopy IVFC, CV-IVFC allows visualization of in vivo cell behavior of interest such as homing and docking events in the vasculature.

While our previous work demonstrated proof of principle of the approach, the performance of CV-IVFC as a function of fluorescence contrast (cell to background) was poorly understood. In general, three methods for fluorescent labeling of circulating cells for IVFC have been reported in the literature:⁴² (1) “ex-vivo” labeling with a membrane-staining fluorescent dye such as vybrant-DiD (as in our previous work), (2) labeling of a cell line with a constitutively expressed fluorescent protein such as GFP, YFP, or mCherry, and (3) anti-body targeting of fluorophores. However, in general, methods (2) and (3) are expected to yield lower contrast than (1) due to increased autofluorescence and reduced light penetration associated with visible excitation and emission wavelengths (inherent to the physics of light propagation in biological tissue), or less efficient fluorescent probe targeting and uptake. Our initial “proof-of-principle” work used method (1), which represented a relatively high-contrast imaging model. Therefore, the performance of CV-IVFC with alternate fluorophores with lower contrast imaging conditions, such as (2) and (3) above, was of significant interest.

To address this issue, in the present work, we tested CV-IVFC with a series of tissue-mimicking optical phantoms with varying levels of autofluorescence, and with calibrated cell-simulating microspheres with varying fluorescent brightness. Depending on instrument parameters, CV-IVFC allowed robust detection of microspheres with significantly lower contrast; for example, detecting at least 50% of microspheres with over two orders of magnitude lower “heterogeneous contrast” (which we define as the number of high-intensity background pixels) than our previously reported in vivo conditions. Occurrence of false-positive detection events was found to correlate with temporal and spatial clustering of high-background pixels. As we discuss, these results are significant since they support the potential utility of CV-IVFC for a much wider range of biological models than our previous proof-of-principle studies.

2.

Methods and Materials

2.1.

Computer Vision In Vivo Flow Cytometry Instrument

A schematic diagram and photograph of the CV-IVFC instrument (as previously described in detail in Ref. 40), are shown in Figs. 1(a) and 1(b), respectively. The samples (mice or phantoms) were placed on the translation stage [Fig. 1(c)] and trans-illuminated with a 660 nm diode laser (DPSS-660; Crystalaser Inc., Reno, Nevada) filtered with a 660 nm “clean-up” filter (d660/20× Chroma Technology, Rockingham, Vermont) to remove any residual out-of-band light. A plano-convex lens pair ($f=50\mathrm{mm}$ and 200 mm; Edmund Optics, Barrington, New Jersey) expanded the beam to 5 mm at full width at half maximum at the sample where the laser intensity was $10\mathrm{mW}/{\mathrm{cm}}^2$. Light was collected with a 2× objective (2× Mitutoyo Plan Apo Infinity-Corrected Long WD Objective Edmund Optics, Barrington, New Jersey) and 1× lens tube. Fluorescence video sequences were obtained by placing a 710 nm filter (et710/50 m; Chroma) in the collection path. Video sequences were captured with a high sensitivity, 14-bit EMCCD camera (${\mathrm{iXon}}^{\mathrm{EM}}+855$ Andor Technology, Belfast, Northern Ireland) operating at a frame rate of 19 Hz. In combination with the collection lenses, this resulted in a $\sim 5\times 5{\mathrm{mm}}^2$ image. The EMCCD camera was also equipped with an adjustable gain setting, which was used in fluorescence image sequences. Monochrome white light images were also obtained by backillumination with a white-light source (Digi-Slave L-Ring 3200, Edmund Optics, Barrington, New Jersey) and removal of the 710 nm emission filter.

Fig. 1

(a) Schematic and (b) photograph of the computer vision in vivo flow cytometry (CV-IVFC) imaging system. M, mirror; Lin Pol; linear polarizer; Obj, 2x objective. (c) Photograph of a mouse ear positioned on the CV-IVFC imaging stage during data acquisition.

2.2.

Computer Vision In Vivo Flow Cytometry Detection and Tracking Algorithm

Given the large illumination area, circulating cells appeared as small targets (approximately 1 to 5 pixels in size) in image sequences. The relatively large tissue autofluorescence (see Sec. 3.1 below) and camera noise (due to gain) required discrimination of moving cells from the significant background signal. Moreover, the rarity of target circulating cell populations ($<1000\text{cells}/\mathrm{mL}$) meant that their arrival in the imaging FOV was infrequent (about 1/min), making manual counting by a human operator tedious and error prone. To address this, we developed a computer vision algorithm to automatically detect and track circulating cells from image sequences. The details of this algorithm were previously described in detail in Ref. 40, but are restated here briefly for completeness: The algorithm (Fig. 2) was coded in MATLAB^® (Mathworks, Natick, Massachusetts) and used two main analysis steps corresponding to “detect” and “connect” operations, respectively. In Step 1, cell candidates were identified by simple pixel-by-pixel mean background subtraction and comparison to a threshold. Any pixel or groups of pixels that exceeded this threshold were assigned “1” values (zero otherwise) and were classified as cell candidates. We ranked all pixel values and chose a threshold in the range from 99.90th to 99.99th percentile of maximum, which was previously determined to work well with in vivo data. In Step 2, the dynamic behavior of cell candidates was analyzed to “connect” them into cell trajectories. This step had multiple operations. First, when a cell candidate was identified, a search was performed inside a fixed radius from the candidate in the subsequent image in the sequence. Second, search regions (“cones”) corresponding to the extrapolated position of the cell candidate given its last known position, speed, and direction were searched in the subsequent 15 images in the sequence. After completion of Step 2, any residual cell candidates that were not merged into trajectories were discarded, and all remaining trajectories were saved and counted by the algorithm as circulating cells. On a 64-bit personal computer with 16 GB of RAM, analysis of a 1000 frame video took approximately 2.5 min. The performance of the algorithm was analyzed for sensitivity and false alarm rate (See Sec. 2.4.2 below).

Fig. 2

Block diagram of the CV-IVFC tracking algorithm, showing major steps in the analysis. Cell candidates are identified (Step 1), and dynamically merged into trajectories (Step 2). The resulting cell trajectories are counted and overlaid on the white light image (see text for details).

2.3.

Phantom Experiments

We developed a set of tissue (ear) mimicking optical phantoms and used a series of cell-simulating fluorescent microspheres. Our goal was to produce imaging conditions with equal or lower contrast than our previous in vivo work. Example images from these studies where vybrant-DiD labeled MM cells were tracked in mice ears in vivo are shown in Fig. 3.⁴⁰ The variation in contrast (defined as $C=[\mathit{I}(\mathit{s})-\mathit{I}(\mathit{b})]/[\mathit{I}(\mathit{b})]$, where $\mathit{I}(\mathit{s})$ is the microsphere intensity and $\mathit{I}(b)$ is the mean background intensity in an image) is evident, and ranged between 0.18 and 1.59 (average of 0.73) between experiments, which we attribute primarily to differences in cell labeling. As we quantify in detail, the phantom experiments performed here yielded conditions with significantly lower contrast, allowing us to study the performance of the CV-IVFC with increased background autofluorescence or weaker cell labeling.

Fig. 3

Example images of circulating cells obtained from our previous in vivo studies⁴⁰ in which fluorescently labeled multiple myeloma (MM) cells were detected and tracked with CV-IVFC, along with the corresponding cell contrast for (a) high, (b) medium, and (c) low levels of contrast. The red circles and arrows indicate the position of the cell.

2.3.1.

Optical phantom preparation

Tissue-mimicking optical phantoms were prepared by mixing polyester resin (Castin’ Craft, Environmental Technology Inc., Fields Landing, California), ${\mathrm{TiO}}_2$ powder (Sigma Aldrich, St. Louis, Missouri), and India Ink (Higgins, Leeds, Massachusetts) to approximately match the reduced scattering coefficient (${\mu }_s^{\prime }=15{\mathrm{cm}}^{-1}$) and absorption coefficient (${\mu }_a=0.1{\mathrm{cm}}^{-1}$) of tissue at 700 nm.⁴³ To replicate the approximate dimensions of the mouse ear [Figs. 4(a) and 4(b)], the resin was placed in a 20-mm diameter and a 2-mm thick mold before hardening. The phantom material alone exhibited modest autofluorescence, so we added AlexaFluor (AF)-680 dye (Life Technologies, Carlsbad, California) to simulate biological tissue autofluorescence. Likewise, sebaceous glands are a prominent feature of in vivo fluorescence images in the mouse ear and appear as bright spots [Fig. 4(c)] that are occasionally falsely identified as circulating cells by the CV-IVFC algorithm. To mimic this, we added $6\text{-}\mu \mathrm{m}$ fluorescent microspheres (Peakflow Claret cytometry beads, P-24670, Life Technologies) to the phantoms. In total, three sets of phantoms ($N=3$ for each) were fabricated with either (1) P1; $0.4\mu \mathrm{M}$ AF680 and 850 microspheres per mL of resin material added, (2) P2; $0.4\mu \mathrm{M}$ AF680 and $1700\text{microspheres}/\mathrm{mL}$ added, or (3) P3; $0.6\mu \mathrm{M}$ AF680 and $1700\text{microspheres}/\mathrm{mL}$ added. Phantom type P2 corresponded to the approximate in vivo imaging conditions observed in our previous work, and type P3 represented a significantly elevated background fluorescence. Before hardening, a strand of Tygon tubing with a $250\text{-}\mu \mathrm{m}$ internal diameter (TGY-010C, Small Parts, Inc., Seattle, Washington) was placed in the resin to simulate a blood vessel.

Fig. 4

Example images of a mouse ear (a-c) and optical phantoms (d-g). (a) Photograph of mouse ear with red arrows indicating large blood vessels. (b) White light image of the mouse ear taken with the CV-IVFC camera, (c) example fluorescence image of the same ear region with sebaceous glands and electron multiplied charge charge-coupled device (EMCCD) shot noise appearing as groups of bright pixels. (d) Photograph of a phantom with a dashed black line indicating the position of the embedded Tygon tubing. (e) White light image of the phantom taken with the EMCCD camera showing position of the Tygon tubing. (f) Fluorescence image of a phantom without fluorescent dye or microspheres added. (g) Fluorescence image of an example P2 phantom type, showing qualitative similarity to (c).

Example white light images of a phantom are shown in Figs. 4(d) and 4(e). Two example fluorescence images are also shown from a phantom without fluorescent material added, and with fluorescent material added (phantom type P2) in Figs. 4(f) and 4(g), respectively, showing the qualitative similarity to the in vivo fluorescence image shown in Fig. 4(c).

2.3.2.

Fluorescent microspheres

We used calibrated fluorescent microspheres suspended in phosphate buffered saline to simulate circulating cells and flowed them through the phantoms using a microsyringe pump (70-2209, Harvard Apparatus, Holliston, Massachusetts) configured to produce a flow speed of $1.7\mathrm{mm}/\mathrm{s}$ in the phantom tube (typical of flow speeds observed in vivo). To simulate the effect of different cell labeling techniques, we used a set of near-infrared microspheres with varying fluorescence intensities and with similar excitation and emission spectra to, e.g., Cyanine 5.5, Alexafluor-680. These are denoted as follows for the balance of this paper: MS1 (Peakflow Claret cytometry beads, Life Technologies), MS2 (Flash Red Intensity Standard FR5, Bangs Laboratories, Inc., Fishers, Indiana) and MS3 (Flash Red Intensity Standard FR4, Bangs). As shown in Fig. 5, these three microsphere types had fluorescence intensities equal to 200%, 62.8%, and 6.8% of the vybrant-DiD labeled MM cells that we used in our previous work. Therefore, this combination of spheres allowed us to test the performance of CV-IVFC with almost two orders of magnitude variation in fluorescence labeling.

Fig. 5

Comparison of average fluorescence intensities of the three microsphere types (MS1, MS2, and MS3) used in these studies and vybrant-DiD labeled MM used in our previous work, normalized to MS1 intensity.

3.

Results

3.1.

Computer Vision In Vivo Flow Cytometry Performance with Varying Background Intensity

Example data generated by the CV-IVFC instrument are shown in Fig. 6. Here, an example fluorescence image sequence for phantom type P2 and MS1 is shown [Figs. 6(a)–6(c)] where each image frame is separated by 0.25 s. After microsphere candidates were identified in Step 1 [Figs. 6(d)–6(f)], they were merged into trajectories in Step 2 and then overlaid on white-light images [Figs. 6(g)–6(i)]. As shown, even though contrast was poor and significant background autofluorescence (AF680 and embedded fluorescent microspheres) was present in the phantom, the algorithm was capable of robustly distinguishing moving spheres from stationary background.

Fig. 6

Example image sequence of a single microsphere traveling through a phantom, separated by 0.25 s. Fluorescence (a-c) and (d-f) binary images obtained after the thresholding operation (Step 1) of the alogrithm are shown. Red circles and arrows indicate the position of the microsphere. In Step 2, dynamic analysis of image sequences was performed, and cell candidates were merged into trajectories, rejecting the stationary or spurious noise evident in (d-f). The recovered microsphere trajectory was overlaid (g-i) on the original white light image.

We systematically tested the CV-IVFC algorithm with phantoms with varying levels of background autofluorescence (phantom types P1-P3) with a single microsphere intensity (MS2). We analyzed image sequences and applied multiple detection thresholds in Step 1 of the algorithm as described in Sec. 2.2 above and in more detail in Ref. 40. Example representative results from three individual phantoms are summarized in Fig. 7. As indicated (and consistent with our previous work), the use of a lower threshold increased overall detection sensitivity [Fig. 7(a)] at the cost of increased FAR [Fig. 7(b)], whereas the use of a higher threshold reduced both parameters. In general, adjustment of this threshold allowed us to trade-off sensitivity and FAR with the system. Figures 8(a)–8(c) are three examples showing extracted microsphere trajectories overlaid on white light images, and (d) is an in vivo video sequence for comparison.

Fig. 7

(a) Sensitivity and (b) false alarm rate (FAR) as a function of threshold level (Step 1) for three different phantom types (P1, P2, and P3) and MS2 microspheres.

Fig. 8

Extracted microsphere trajectories overlaid on white light images for (a) phantom type P1 [Video 1 ( http://dx.doi.org/10.1117/1.JBO.20.3.035005.1)], (b) phantom type P2 [Video 2 ( http://dx.doi.org/10.1117/1.JBO.20.3.035005.2)], (c) phantom type P3 [Video 3 ( http://dx.doi.org/10.1117/1.JBO.20.3.035005.3)], and (d) in a mouse ear in vivo [Video 4 ( http://dx.doi.org/10.1117/1.JBO.20.3.035005.4)]. The left side of the video displays the original fluorescence image sequence, and the right side displays processed video (Video 1, QuickTime, 90 KB; Video 2, QuickTime, 707 KB; Video 3, QuickTime, 71 KB; Video 4, QuickTime, 72 KB).

3.2.

Computer Vision In Vivo Flow Cytometry Instrument Performance with Varying Microsphere Intensity

Next, we tested a single phantom (P2) with three different microsphere types with decreasing fluorescence intensities (MS1-MS3). Again, we computed sensitivity and FAR for a series of threshold values, and the data are summarized in Fig. 9. Generally, this followed the same trend as above, in that increasing the detection threshold resulted in lower sensitivity [Fig. 9(a)] and lower FAR [Fig. 9(b)]. We also note that, in general, shorter trajectories were extracted by the CV-IVFC algorithm in lower contrast conditions. Specifically, the average microsphere track lengths were 2.14, 1.42, and 0.82 mm for MS1, MS2, and MS3 types, respectively (see Fig. 10). Although circulating microspheres were visible for only brief periods in low contrast conditions, CV-IVFC was capable of robustly detecting them.

Fig. 9

(a) Sensitivity and (b) FAR as a function of threshold level (Step 1) for three different microsphere types (MS1, MS2, and MS3) and a P2 phantom type.

3.3.

Computer Vision In Vivo Flow Cytometry Instrument Performance Summary: All Conditions

The overall performance of the CV-IVFC algorithm for the phantom and microsphere tests performed here is summarized in Fig. 11. We first considered the overall sensitivity of the algorithm for all 27 combinations ($3\text{microsphere intensities}\times 9\text{phantoms}$) as a function of the image contrast. These data are shown in Fig. 11(a). (Here, we selected a threshold of the 99.95th percentile for the algorithm in Step 2, although this curve could be re-generated for any other threshold). All phantoms and microsphere combinations that we tested are shown, where the triangle markers correspond to MS1, squares to MS2, and circles to MS3. As indicated, although three sets of nominally identical phantoms were fabricated (P1–P3), significant inter-phantom variability was observed. For this reason, we plotted all individual data points (rather than means and standard deviations). As shown, the range of average contrasts (C) for different phantom-microsphere combinations was between 0.23 and 1.54, although the contrast of individual microspheres was as low as 0.06. For comparison, the range obtained from our earlier in vivo experiments is indicated in red [Fig. 3 and Ref. 40], where an average contrast of 0.73 was observed. By inspection, C was a poor predictor of algorithm performance for low contrast phantom-microsphere combinations (i.e., when $C<0.5$), which as we have noted was our primary interest in this study. Review of our data suggested that the algorithm yielded highest tracking sensitivity when the background autofluorescence was relatively homogenous and could be effectively removed by subtraction. Therefore, as an alternative metric, we next considered the heterogeneity of the background intensity, which we defined as the number of “bright pixels” per frame in the image sequence [Fig. 11(b)]. Here, “bright pixels” were defined as the number of pixels per image after background subtraction that exceeded the mean intensity of a microsphere. In other words, these are the number of pixels where the noise on the pixel was comparable to the intensity of a microsphere. For the phantom experiments, a range of 0.2 to 518.3 bright pixels per image were observed, whereas on average 2.86 bright pixels per image was observed in our previous in vivo data. By inspection, the in vivo data generally agree with the phantom data, suggesting that our phantom-microsphere model mimicked in vivo imaging conditions well. Somewhat surprisingly [as shown in Fig. 11(c)], the number of bright pixels correlated only weakly with the FAR, since the CV-IVFC algorithm (Step 2) was efficient at rejecting bright pixels that were physically well separated. Further analysis revealed that a more complex quantity––the number of “bright pixel clusters” per minute––correlated well with FAR as shown in Fig. 11(d). Here, “bright pixel clusters” were defined as the number of occurrences for which two bright pixels were detected in subsequent image frames within a fixed radius (6 pixels). Analysis of our data indicated that, in general, these led to erroneous formation of microsphere trajectories. As above, our previous in vivo data were coplotted with the phantom data showing good general agreement. The implications of these results are discussed below.

Fig. 10

Example microsphere trajectories obtained in phantom models for (a) MS1, (b) MS2, and (c) MS3 microsphere types for a single phantom type P2. (d) The average trajectory length for each of the microsphere types.

4.

Discussion and Conclusion

We recently developed and validated CV-IVFC as a new technology for detection, enumeration, and tracking of rare circulating cells in small animals in vivo. The major advantage of CV-IVFC compared to previous IVFC designs is the relatively large circulating blood volume that is sampled, leading to improved detection of low-abundance cells. We previously showed that this instrument was capable of detecting circulating cells in the range of $20\text{cells}/\mathrm{mL}$. The purpose of the studies performed here was to characterize CV-IVFC under conditions of low imaging contrast compared to our previous in vivo studies.

The data in Fig. 11 summarize the effects of reduction in contrast from either mechanism. In combination, Figs. 11(a) and 11(b) show the effect of loss of contrast (homogenous and heterogeneous) on detection sensitivity, and illustrate that CV-IVFC can effectively detect and track fluorescent targets with significantly lower contrast than in our previous in vivo studies. To quantify this, we considered the heterogeneous contrast data shown in Fig. 11(b). Although the parent function for these data is unknown, for comparison purposes we fit a linear function [dotted line on Fig. 11(b); note that this appears as a curve due to the logarithmic $x$-axis]. This fit implies that CV-IVFC would retain 80% detection sensitivity for the case corresponding to 109 bright pixels per image, which is equivalent to 38 times lower contrast versus our in vivo studies, where 2.86 bright pixels per image were observed on average. Likewise, 50% of microspheres would still be detected for the case of 457 bright pixels per image, corresponding to 159 times lower contrast. In practice, this reduction in contrast could come from either less-efficient fluorophore cell labeling, lower fluorescence quantum yield, or increase in background autofluorescence. This greatly increases the potential utility of CV-IVFC since it can be used with a wide range of fluorophores and labeling techniques. Absolute quantification of the mean equivalent soluble fluorochrome units for the microspheres was not available from the manufacturer, but comparison with our previous work and the literature values suggests that the intensity of the MS2 type would be comparable to a “typical” well-labeled cell. For example, comparison of extinction coefficient and emission quantum yield data suggests that cells labeled with the Turbo-FP650 red fluorescent protein (RFP; Evrogen, Moscow, Russia)^44–46 would be about three times less-brightly labeled than with vybrant-DiD. Likewise, the use of a green fluorophore would reduce the contrast by approximately a factor of 10 due to increased autofluorescence,^47–50 which (according to the analysis here) would not significantly degrade CV-IVFC sensitivity.

As we have noted, the relationship between FAR and bright pixels was more complex, since FAR was found to correlate with the number of “cell clusters” relating to both the temporal and spatial distribution of bright targets. In practice, this suggests that relatively high background autofluorescence can be tolerated by the algorithm without significant increase in FAR provided that bright pixels are physically separated in the image. As noted above, sensitivity and FAR can be traded-off by adjusting the “threshold” parameter in Step 1. In CV-IVFC, false alarms can be easily discounted by a human operator reviewing the original image sequence. Because the algorithm reports the time of detection in the image sequence, this is significantly more time efficient than manually counting cells with long (30 min or more) data acquisitions. Therefore, in practice, higher FAR is generally preferable to lower sensitivity in CV-IVFC.

We also note that the results shown here pertain specifically to our existing CV-IVFC instrument and algorithms and do not represent a technical limit of the technique. Improvement of the CV-IVFC tracking algorithm is the subject of ongoing work in our lab; we are currently pursuing a number of alternative strategies, for example, by jointly solving the “detect and connect” problem, as opposed to our current two-step process. Likewise, we have noted that stationary but distributed bright regions (such as sebaceous glands) appearing in multiple images frequently trigger false positives in the algorithm. As such, we are studying methods to automatically detect such regions and reject them as cell candidates. Therefore, we expect that a better CV-IVFC performance may be achieved in the future.

Fig. 11

(a) Sensitivity as a function of contrast for all phantoms and micosphere combinations tested. The mean and standard deviation sensitivity and bright pixel numbers for our previously acquired in vivo data are also plotted for comparison (red line). (b) Sensitivity as a function of “bright pixels” per frame (see text for definition) for experimental combinations tested. Unlike sensitivity, (c) FAR correlated poorly with “bright pixels” but correlated well with (d) “bright pixel clusters” per minute (see text for definition). Previous in vivo data are also shown for comparison in red. Dotted lines in (b) and (d) are linear fits to the data.