Improved detection of fluorescently labeled microspheres and vessel architecture with an imaging cryomicrotome
- First Online:
- Cite this article as:
- van Horssen, P., Siebes, M., Hoefer, I. et al. Med Biol Eng Comput (2010) 48: 735. doi:10.1007/s11517-010-0652-8
- 603 Downloads
Due to spectral overlap, the number of fluorescent labels for imaging cryomicrotome detection was limited to 4. The aim of this study was to increase the separation of fluorescent labels. In the new imaging cryomicrotome, the sample is cut in slices of 40 μm. Six images are taken for each cutting plane. Correction for spectral overlap is based on linear combinations of fluorescent images. Locations of microspheres are determined by using the system point spread function. Five differently colored microspheres were injected in vivo distributed over two major coronaries, the left anterior descending and left circumflex artery. Under absence of collateral flow, microspheres outside of target perfusion territories were not found and the procedure did not generate false positive detection when spectral overlap was relevant. In silico-generated microspheres were used to test the effect of background image, transparency correction, and color separation. The percentage of microspheres undetected was 2.3 ± 0.8% in the presence and 1.5 ± 0.4% in the absence of background structures with a density of 900 microspheres per color per cm3. The image analysis method presented here, allows for an increased number of experimental conditions that can be investigated in studies of regional myocardial perfusion.
KeywordsMyocardial perfusion Fluorescent microspheres Microvascular structure Cardiovascular imaging Separation of fluorescent labels
Labeled microspheres have been used for more than half a century to quantify perfusion distribution, either over the circulatory system as a whole or within an organ . In this method, labeled microspheres are injected into the circulation either as a bolus or by continuous infusion over a period of time . The working principle is that these microspheres, typically with a diameter around 15 μm, travel with the blood stream and become trapped in the microcirculation since they cannot pass the capillaries. Ultimately, they shift from the capillary lumen to the interstitium [4, 10]. The number of microspheres in a certain region, relative to the total amount of injected microspheres and weighted by the blood flow at the site of injection, represents the regional perfusion in volume flow per unit mass.
Several methods have been developed permitting the distinction of different labels. The classical labels were radioactive , colored  or following recent development, fluorescent . Tissue samples for measurements were originally obtained by cutting the tissue in the area of interest into pieces of 0.1–2 cm3 depending on the study aim . The radioactivity in these samples was then measured directly  or these were processed to determine the total fluorescence or number of microspheres .
In a recently introduced and more advanced method, an imaging cryomicrotome is used that allows the position of individual microspheres to be detected . In this method, the tissue block under investigation is frozen and sliced with a cryomicrotome at a preset thickness, e.g., 40 μm at a time. After each cut, fluorescent images are taken with a high-resolution digital camera. Image processing then yields the x–y–z coordinates of the microspheres [2, 11].
The classical and cryomicrotome method share a commonality in that the tissue examined is destroyed. The information retained in the classical method has a spatial resolution of the tissue sample size which has an under limit of 0.1 g, while in the cryomicrotome the resolution is of the voxel size of the camera and slice thickness, which is a factor 1 × 106 higher. Moreover, the region of interest for calculation of microsphere density can be selected at will in the cryomicrotome method but not in the classical methods. In addition, injection of fluorescent cast material into the vasculature prior to freezing allows detection of the arterial tree structures down to small arterioles with a diameter of 20–40 μm [15, 16]. Hence, the imaging cryomicrotome enables the combined measurements of blood vessel structure and flow distribution within an organ. However, until now, only four different fluorescent labels could be distinguished with the imaging cryomicrotome method , which is a limitation compared to the earlier radioactive labeled methods where up to nine different labels could be detected.
The purpose of this study was to increase the number of differently labeled fluorescents. We present a method to improve the fluorescent color separation by correcting for spectral overlap based on a linear combination of the images. Our approach permits differentiation of six fluorescent labels, of which five are used for the microspheres and one for the infused casting material. The proposed method was validated both in silico and with actual data obtained from canine hearts.
2 Materials and methods
2.1 Canine dataset
Datasets for this study were obtained from three animals of a series of physiological experiments carried out in canines according to a protocol that was approved by the Institutional Animal Care and Use Committee of the University of Utrecht Medical Center. After pre-medication with Dormitor 0.03 mg/kg i.m. (intra-muscular), Ketamine 0.03 mg/kg i.m. and Atropine 0.5 mg i.m. anesthesia was induced by Sufenta forte 10 μg/kg/h i.v. the animals were intubated, and a surgical level of anesthesia was maintained by mechanical ventilation with a mixture of Propofol 24 mg/kg/h i.v., Sufenta forte 3 μg/kg/h i.v. and oxygen. The hearts were exposed by left thoractomy via the fourth intercostal space. Under fluoroscopic imaging, 8F guiding catheters were positioned at the ostium of the left circumflex artery, LCX, and the left anterior descending artery, LAD. Microspheres were gradually infused over a period of 1 min (standard infusion pump, Harvard Apparatus, USA). The microsphere dose for each infusion was approximately 1.0 × 105 (15μm, Invitrogen, The Netherlands) as determined from microsphere concentration in the storage vial and volume withdrawn. The colors used (with corresponding excitation and emission maxima in nm between brackets) were: green (454/492), yellow (518/521), red (570/598), carmine (582/614) and scarlet (651/680). Microspheres were injected at different physiological conditions such as baseline and hyperemia. However, for this study, it was not attempted to correlate distributions to these physiological conditions.
After the experiment, the animal was killed by fibrillating the heart after which it was excised, and the right and left main coronary arteries were cannulated and perfused with a buffered solution containing adenosine for vasodilation of the microcirculation. Subsequently, the main vessels were infused with cast replica material (Batson no. 17, Polysciences, USA) consisting of a monomer base solution, a catalyst and a promoter as described previously [15, 16]. Potomac Yellow (440/490) (Radiant Colour, Belgium) provided the fluorescent base for the replica plastic. The cast material was allowed to harden over a period of 24 h at ambient temperature. The fully prepared heart was embedded in carboxymethylcellulose sodium solvent 5% (Brunschwig Chemie, The Netherlands) and Indian ink 5% (Royal Talens, The Netherlands), frozen at −20°C and placed in the cryomicrotome for subsequent cutting .
2.2 The imaging cryomicrotome
Filter combinations used, with the corresponding imaged fluorescents
Cast material only
Green microspheres and cast material
Red microspheres and yellow microspheres
Yellow microspheres and green microspheres
Carmine microspheres and red microspheres
Scarlet microspheres and red microspheres
2.3 Image processing
Three-dimensional (3D) computer reconstructions of the specimen were obtained based on 40 × 40 × 40 μm3 voxels. The image stack resulting from reflection imaging delivered an outline representation of the heart, while the stacks of fluorescent images yielded a 3D reconstruction of the vasculature and distributions of microspheres with different fluorescent labels. A series of algorithms was developed and applied to correct for imaging artifacts as described below.
2.4 Correction for transparency effects
Without compensation, light coming from deeper layers of the sample introduces false detections of microspheres. The effect can span a depth of up to 200 μm and hence, the majority of microspheres visible at the surface are not actually present in the surface layer, slice n. Up until now, the cutting depth at which the light disappeared was taken as the location of the microsphere . However, subsequent microspheres will be counted as one, which results into loss of observations. Therefore, we propose an image processing step to correct for transparency effects prior to detection.
2.5 Correction for spectral overlap
Linear combinations of images obtained with different filter settings to yield a unique color
Combination of images
2.6 Correction for auto-fluorescence
2.7 Microsphere detection
A binary representation of the data is created by applying a threshold to the images, based on the median value of the intensity for the stack. Each cluster of bright pixels that are connected is initially counted as one microsphere observation. The observations are then checked for four criteria before the final microsphere location is accepted and stored.
Only if the four criteria are passed, the 2D location of the identified microsphere in the current image is stored in a text file for further processing.
2.8 In silico validation
In order to evaluate the robustness of the spectral overlap correction, the full detection method for all five colors of microspheres along with the cast material was additionally tested with a simulated dataset. The parameters to create the test datasets were chosen such that they corresponded to actual in vivo measurements as measured by Rolf et al. We used Eq. 4 to produce the datasets, with varying attenuation coefficients, in-plane and depth-dependent standard deviation for the Gaussian kernel and microsphere density. For σ1, we chose a fixed value of 0.1 pixel, σ2 ranged from 0.8 for B2, 1.0 for C3, 1.2 for B4/D4 and 1.4 pixel for E5 and the attenuation coefficient ranged in pixels from 0.5 for B2, 0.4 for C3, 0.3 for B4/D4 and 0.2 for E5 accordingly. The microspheres were randomly distributed in the in silico datasets, although in vivo this is typically not the case . To overcome this limitation, we created two distributions with different densities; a high density set, with approximately 900 microspheres/cm3 to resemble local high flow regions and a density stack of approximately 500 microspheres/cm3 comparable to the average number of microspheres/cm3 as observed in animal experiments. The individual microspheres within the in silico distributions were assigned different intensity levels between 100 and 255. This is in accordance with in vivo observations where the intensity can vary due to the position of the microsphere in the current slice, variation of intensity of microspheres and regional tissue properties. Finally, a realistic background image was added by superimposing in vivo images onto the in silico datasets as a real life representation of the variation in background intensity due to the camera properties and the surrounding tissue. These images were analyzed with the corrections as described above.
3.1 Transparency effects and auto-fluorescence
3.2 Microsphere counts and location
3.2.1 In silico dataset
The simulated images allowed us to validate the detection of microspheres as well as the performance of the algorithm in the presence of microsphere clusters.
Detection agreement for the in silico datasets with and without noise added to the images: (a) approximately 900 microspheres/cm3 per color; (b) approximately 500 microspheres/cm3 per color
% of total
% of total
(a) High density distribution
98.5 ± 0.4
97.7 ± 0.8
(b) Normal density distribution
99.1 ± 0.8
98.2 ± 1.3
3.2.2 In vivo dataset
From each of these colors, saddle points and high intensity detections were evaluated. The percentage of detected saddle points (indicating doubles) and high intensities (clusters of microspheres) amounted to a mean value and standard deviation of 0.75 ± 0.53 and 1.81 ± 0.85%, respectively. In addition to the microsphere injections with spectral overlap, a third color of microspheres (green) was injected into the LAD that had no spectral overlap with the prior injections. Scarlet microspheres combined with yellow microspheres were also injected into the LAD of the third heart. This was done to evaluate the amount of clustering for differently labeled microspheres without spectral overlap. For all the three hearts, the number of such clusters was determined and resulted in a mean of 0.26% with a standard deviation of 0.15%.
We presented a method to correct for spectral overlap of fluorescent emission light, which allows the use and successful separation of a larger number of fluorescent labels in imaging cryomicrotome applications. With our image processing algorithms, we successfully distinguished one color for the casting material and an additional five colors for microspheres, thereby increasing the number of fluorescent colors successfully detected from 4 to 6. An increase in labels is important, since it enlarges the maximum number of experimental conditions that can be investigated in a single animal which in turn lowers the number of animal experiments needed for a particular protocol and reduces the organ-to-organ variability across animals.
The in silico simulations demonstrated that over 98% of the microspheres were correctly counted for a realistic range of microsphere densities. The number of missed counts was only 1.5 ± 0.4% for high microsphere density (900 spheres/cm3) and decreased to 0.9 ± 0.8% for a lower concentration (500 spheres/cm3). With background images added, the number of missed microspheres increased to 2.3 ± 0.8 and 1.8 ± 1.3%, respectively. The error was not specific for any color.
The data obtained from frozen canine hearts showed a clear demarcation zone between the LAD and LCX perfusion territories indicating the absence of false positive detection of microspheres. Although, the number of saddle points and high intensity observations was very low, it increased with the wavelength of the emission light. This is due to the widening of the point spread function as a consequence of the lower attenuation coefficient of the tissue for longer wavelengths in combination with the chromatic aberration of the lens system.
4.1 Factors influencing microsphere detection
The unwanted contribution of light due to spectral overlap is already minimal for selective excitation and emission filter combinations, hence β(λex, λem) ≪ α(λex, λem) (Eq. 10). For the casting material (Potomac Yellow), the single filter B1 (440/480) is sufficient, since the emission extends further into the near ultraviolet range, with no overlap up to 490 nm. For the other colors, the value for the scale factor ε (Eq. 12) typically ranges from 0.5 to 1. Especially for the red and carmine microspheres, β(λex, λem) ≈ α(λex, λem) as the overlap in emission spectra is clearly discernible (Fig. 2). Hence, spectral overlap correction is a necessity for these colors.
In our image analysis, we assumed that the point spread function is not a function of spatial coordinates. However, in reality, this is not the case and hence transparency correction is complicated due to the anisotropic optical properties of the tissue. To avoid the problem related to anisotropy, we used an underestimate for the attenuation coefficient as proposed by Rolf et al. . If the sequential slice correction is then applied, transparency may be overcorrected but this results not in undetected microspheres.
Dark current of the camera is a common source of error when imaging with exposure times in the order of seconds and adds to image degradation. It is the result of thermionic emission or other effects when there is no light incident on the charged coupled device and results in relatively bright pixels compared to the median value of all pixels in the image. The effects of this dark current are suppressed by active cooling of the camera, but may remain present to some degree as a consequence of long exposure times. Since the errors due to dark current are consistent in the longitudinal direction of the stack, the applied correction for transparency effects also removes potential dark current errors. The effect of dark current in isolated pixels on the detection of microspheres is further diminished, since a microsphere is classified by two or more pixels.
Clustering of microspheres may additionally lead to incorrect microsphere identification. One may classify clustering into three types: (1) microspheres of the same color are clustered prior to injection, (2) single microspheres of the same color cluster together during the injection in the same blood vessel, and (3) microspheres of different colors cluster together during injection in the same blood vessel. Clustering of the same color as in possibilities 1 and 2 is detected by the saddle point checks and the identification of observations larger than two microspheres. Clustering of two microspheres of different colors, however, gives rise to possible loss of observations when there is spectral overlap, since part of the intensity belonging to a color of interest is removed. The clustering of originally detached microspheres by accidentally following the same path to the point of lodging as defined in possibilities 2 and 3 were addressed in silico and in vivo. In the simulated datasets, 3D distribution of microspheres were stochastically distributed. In only 1.5% of the voxels containing two microspheres, more than two were found. When counting the colors without spectral overlap injected into identical perfusion territories, on average only 0.26% of the total number of microspheres shared identical xyz coordinates.
4.3 Comparison with other methods of microsphere density estimation
The imaging cryomicrotome has several advantages over the more classical methods of microsphere density determination, where first the tissue has to be partitioned and then each tissue sample has to be processed to either count the microspheres by a flow cytometric quantification method or by determining the average tracer concentration. For the latter method, a correction is needed to account for the natural decay of fluorescence intensity or radioactivity, depending on the tracer used. This obviously introduces additional sources of error. This problem does not exist with the imaging cryomicrotome, since the microspheres are detected by the shape of corresponding fluorescence patterns rather than by intensity alone. The cytometric quantification method is also free of the intensity decay problem, but the resolution is not improved. With the datasets obtained from a cryomicrotome, any desired 3D analysis of microsphere distribution can be applied, since the spatial relation of the dataset with respect to the outline images is not destroyed. Moreover, in combination with the casting procedure, a detailed analysis of microsphere distribution (flow) with respect to the intramural vascular structure can be performed in organs such as the heart.
We demonstrated the feasibility of using different filter combinations to distinguish five separate microsphere colors in addition to the fluorescently labeled vascular tree. This allows high-resolution spatial measurements of local perfusion for five different experimental conditions in the same organ that can additionally be closely related to the detected vascular structure.
The authors would like to thank J. Weda (Dept. of Biomedical Engineering and Physics) for his assistance in obtaining the spectra of the fluorescent labels presented in this work. This work has been supported in part by The Netherlands Heart Foundation (NHS 2000.082, NHS2006B226, NHS 2006B186), by The Netherlands Organization for Scientific Research (ZONMW 91105008), and by the European Community’s Seventh Framework Program Grant FP7/2007-2013 under grant agreement no. 224495 (euHeart project).
This article is distributed under the terms of the Creative Commons Attribution Noncommercial License which permits any noncommercial use, distribution, and reproduction in any medium, provided the original author(s) and source are credited.