Journal of Microscopy, Vol. 00, Issue 00 2015, pp. 1–12

doi: 10.1111/jmi.12252

Received 17 September 2014; accepted 1 March 2015

Automatic quantification of angiogenesis in 2D sections: a precise and timesaving approach C . W E I S ∗ , §, # , J . M . C O V I †, # , J . G . H I L G E R T †, N . L E I B I G ‡, A . A R K U D A S †, R . E . H O R C H †, U . K N E S E R ‡ & V . J . S C H M I D T †, ‡ ∗ Department of Physics, Friedrich-Alexander-University Erlangen-Nuremberg, Erlangen, Germany

†Department of Plastic and Hand Surgery, University Hospital Erlangen, Friedrich-Alexander-University Erlangen-Nuremberg, Erlangen, Germany ‡Department of Hand-, Plastic- and Reconstructive Surgery, BG Unfallklinik Ludwigshafen, University of Heidelberg, Heidelberg, Germany §Department of Diagnostic and Interventional Radiology, University Hospital Heidelberg, Heidelberg, Germany

Summary Introduction: The standardized characterization of angiogenesis is crucial in the field of tissue engineering as sufficient blood supply is the limiting factor of mass transfer. However, reliable algorithms that provide a straight forward and observer-independent assessment of new vessel formation are still lacking. We propose an automatic observer-independent quantitative method (including downloadable source code) to analyze vascularization using two-dimensional microscopic images of histological cross-sections and advanced postprocessing, based on a ‘positive- and negative-experts’ model and a (corrected) nearest neighbour classification, in a vascularized tissue engineering model. Materials and Methods: An established angioinductive rat arteriovenous loop model was used to compare the new automatic analysis with a common 2D method and a µCT algorithm. Angiogenesis was observed at three different time points (5, 10 and 15 days). Results: In line with previous results, formation of functional new vessels that arose from the venous graft was evident within the three-dimensional construct and a significant (p < 0.05) increase in vessel count and area was observed over time. The proposed automatic analysis obtained precise values for vessel count and vessel area that were similar to the manually gained data. The algorithm further provided vectorized parameterization of the newly formed vessels for advanced statistical analysis. Compared to the µCT-based three-dimensional analyses, the presented two-dimensional algorithm was superior in terms of small vessel detection as well as cost and time efficiency. Conclusions: The quantitative evaluation method, using microscopic images of stained histological sections, ‘positiveand negative-experts’-based vessel segmentation, and nearest neighbour classification, provides a user-independent and # These authors contributed equally to this work. Correspondence to: Volker J. Schmidt, BG Unfallklinik Ludwigshafen, University of

Heidelberg, Ludwig-Guttmann-Straße 13,67071 Ludwigshafen. Tel: +49(0) 6216810-0; fax: +49(0) 621-6810-2600, e-mail: [email protected]  C 2015 The Authors C 2015 Royal Microscopical Society Journal of Microscopy 

precise but also time- and cost-effective tool for the analysis of vascularized constructs. Our algorithm, which is freely available to the public, outperforms previous approaches especially in terms of unambiguous vessel classification and statistical analyses. Introduction Angiogenesis is critical in tissue engineering concepts as blood vessels enable oxygen supply and delivery of nutrients. To maintain cell survival with regard to the oxygen diffusion limit of 100–200 µm (Carmeliet & Jain, 2011), a pre-existing microcirculation is an essential condition of larger vascularized tissue transplants. Independent from the processing method, a precise and reliable evaluation of the vascularization is critical for the preclinical assessment of tissue engineered constructs (Rath et al., 2012; Arkudas et al., 2013; Strobel et al., 2014). To investigate the various attempts of angiogenesis, a reproducible tool for quantification is necessary. Nowadays, characterization of vascularization is usually performed using two-dimensional (2D) histological or micro-computed tomography (micro-CT) analysis techniques. Although micro-CT provides observation of the three-dimensional (3D) vasculature and automatic quantitative algorithms are available, the technique is timeand cost-intensive, limited to reference centres and is most commonly performed postmortem in rodent models (Arkudas et al., 2010). Established methods for the analysis of histological cross-sections on the quantification of angiogenesis bear several problems, e.g. speed and user dependence. We circumvent these limitations by introducing an automatic analysis tool, addressing multiple types of images, whether greyscale or colour, with different stainings, creating reproducible results and offering a well-described interface that provides the segmented vessels’ data for further evaluation. For the evaluation of the analysis method, we used a unique model of angiogenesis in vivo which is distantly related to the early work of Erol and Spira (1979, 1980) and was further modified by Morrison and Tanaka (Tanaka et al., 2000). The model is based on a grafted vessel that is interpositioned

2

C. WEIS ET AL.

between the femoral artery and vein creating an arteriovenous (AV) loop. This approach is of special interest as in such grafted vessels embedded in a fibrin matrix, angiogenesis occurs without requiring addition of proangiogenic factors (Kneser et al., 2006; Polykandriotis et al., 2008; Arkudas et al., 2010; Schmidt et al., 2013). The aim of this study was to: (1) establish an automatic quantitative analysis method of vascularization using microscopic images of stained histological cross-sections, (2) assign newly formed vessels to a coordinate system independent from the microscopy image for complex statistical analysis, (3) assess whether the automatic analysis method of vascularization is comparable with standard analysis methods, and (4) analyze the temporal dynamics of axial vessel sprouting using the rat AV loop model as a proof-of-concept. Material and methods Experimental design Experiments were performed in accordance with the German animal protection law and were approved by the Institutional Animal Care and Use Committee of the Regierungspr¨asidium Mittelfranken (54-2532.1-34/09). Eight male Lewis-rats (Charles River Laboratory, Sulzfeld, Germany) with an average weight of 340 g were anaesthetized using isoflurane (Baxter Health Care S.A., H¨ochstadt, Germany). In this proofof-concept study, the sample size was limited to an amount that was necessary to gain significant results. An AV loop was constructed between the left femoral artery and vein by interposition of a venous graft from the contralateral thigh. The AV loop was embedded within a fibringel-filled chamber to induce 3D angiogenesis. New vessel growth was investigated 5, 10 and 15 days after surgery. Isolation chamber and fibrin matrix The cylindrical chamber exhibited a diameter of 10 mm and a height of 6 mm (P. Greil, Department of Materials Science, Glass and Ceramics, University of Erlangen-Nuremberg). To prevent loop dislocation, four tubes were placed centrally (Fig. 1B; Polykandriotis et al., 2007). The chamber was filled R with 800 µL fibrin sealant (Tissucol ; Baxter Health Care) composed of fibrinogen (10 mg/mL), thrombin (2 IU/mL) and aprotinin (1500 KIE/mL; Figs. 1C, D). Surgical procedures Operations were performed by the same investigator (J.M.C) using a surgical microscope (magnification ×16, OPMI IFC, Carl Zeiss, Oberkochen, Germany). Anaesthesia was induced R , Baxter, Health Care) and animals received (5% Isofluran R ¨ tramadol (7.5 mg/kg i.v.; Tramal ; Grunenthal, Aachen, Germany), benzylpenicillin-streptomycin (0,5 mL/kg i.m.; R -compositum, Albrecht, Aulendorf, Germany) and Veracin

R heparine (80 IU/kg Liquemin ; Ratiopharm, Ulm, Germany). After midventral incision of the medial left thigh, femoral artery and vein were dissected and mobilized from the pelvic artery downstream to the knee level. Afterwards, a 20 mm vein graft was harvested from the femoral vein of the contralateral thigh and placed between the left femoral artery and vein by microsurgical techniques creating an AV shunt (Fig. 1A). The AV loop was then embedded into the fibrin-filled isolation chamber (Figs. 1B, C), which was closed and fixed onto the underlying adductor fascia (Prolene 3/0, Ethicon, Norderstedt, Germany) subsequently (Fig. 1D). Finally, the skin was closed by means of Vicryl 4/0 (Ethicon). The animals were housed in the animal facility of the University of Erlangen Medical Centre and kept at a 12 h dark/light cycle with free access to standard chow (Sniff) and water. At the end of the experiment, animals were killed by intracardial injection of a combination of embutramid, mebezonium R and tetracain (15 mL/kg: T61 : Intervet, Unterschleißheim, Germany) under deep general anaesthesia (5% isoflurane).

Histological and Micro-CT analysis For vessel detection, the distal descending aorta was cannulated (24-gauge catheter) after abdominal midline incision. The downstream vascular system was flushed with 150 mL isotonic salt solution containing heparin (100 IE/mL) followed by an injection of 30 mL warmed (37°C) India ink solution (50% v/v India ink [Lefranc & Bourgeois Nan King] in 5% gelatin and 4% mannitol) into the aorta to visualize perfused vessels around the constructed AV loop. For micro-CT analyses, 20 mL yellow Microfil (MV-122) containing 5% of MV Curing Agent (both Flowtech, Winterthur, Switzerland) was applied instead. After removing the sample from the chamber, constructs were fixed (formaldehyde 4%), dehydrated and embedded in paraffin. Histological slices (thickness 5 mm) were obtained at standardized planes perpendicular to the longitudinal AV loop axis by means of a microtome (Leica Microsystems, Wetzlar, Germany). Slices were stained by hematoxylin and eosin using an autostainer (ST5010, Leica Microsystems) according to standard protocols. Rat endothelial cells were detected immunohistochemically using the isolectin B4 (Sigma L 2140, Sigma-Aldrich, Hamburg, Germany) as described elsewhere (Hayes & Goldstein, 1974; Arkudas et al., 2009). For rehydration, specimens underwent a descending xylol-alcohol series. After incubation with tris-buffer (Merck KGaA, Gernsheim, Germany, 2 min) and citrate-buffer (pH 6.0, 1 min), sections were treated with 3% peroxidase for 10 min and blocked using avidin, biotin (each 15 min) and finally 10% normal goat serum (30 min). Subsequent samples were incubated with biotinlabeled lectin (1:120, isolectin B4 , Sigma-Aldrich) in trisbuffer at 4°C overnight. After treatment with tris-buffer and R Tween 20 (Carl Roth GmbH & Co. KG, Karlsruhe, Germany), streptavidin HRP (horseradish peroxidase, K0679, LSAB-kit,  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

AUTOMATIC QUANTIFICATION OF ANGIOGENESIS IN 2D SECTIONS

3

Fig. 1. The arteriovenous loop was completed (A). Subsequently, the construct was embedded in a fibrin-filled chamber and four centrally placed tubes prevented displacement of the loop (B). The chamber was fixed onto the fascia and the second firbin layer was added to cover the loop (C). Prior to skin suture, a lid was placed on top of the chamber (D).

DAKO, Hamburg, Germany) was applied for 30 min followed the application of diaminobenzidine (DAKO) for 10 min. After washing with PBS, counterstaining occurred with hemalum (Mayer´s H¨amalaunl¨osung, Merck KGaA). Finally, samples R were washed with PBS and sealed with Aquatex (Merck, KGaA). For micro-CT scans (see Fig. 2), explanted constructs were placed on the table of a high-resolution, cone-beam micro-CT scanner developed at the Institute of Medical Physics (ForBild, University Erlangen, Germany). Scanning parameters: tube voltage 40 kV, 250 mA, 15 µm voxel size, 250 mm tube–detector distance. To compare µCT results with the values from the presented 2D algorithm, data were aligned via image registration (Rigid and Non-Rigid Image Registration, Chimeara GmbH, Effeltrich, Germany). Microscopy Images of each stained cross-section were acquired with an Olympus IX81 inverted microscope (Tokyo, Japan) with 10× magnification (objective 10x, coupler 1.0x). Exposure time was set by the microscope, respectively, its control software (Olympus cellSens dimension). The whole sample was covered by multiple images that were acquired and automatically stitched together by Olympus’ Multiple Image Acqui C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

Fig. 2. 3D reconstruction of the AV loop-associated neovascularization C by means of micro-CT data. Vessels (yellow) were detected after microfil perfusion. The orientation of the histological cross-sections is indicated in light blue.

sition technique (MIA). Greyscale images were taken with a XM10 camera (pixel width/height: 6.45 µm) of unstained cross-sections just prepared with India ink, truecolour images

4

C. WEIS ET AL.

were obtained by a SC30 camera (pixel width/height: 3.2 µm) for the stained cross-sections (lectin/H&E + India ink). Afterwards, the images were saved as jpg and/or tif and transferred to the postprocessing computer. R If perfusions with either India ink or Microfil were incomplete, the blank vessels in the histological images had to be filled manually. Therefore, the lumina, which could be easily detected in case of the lectin staining, were dyed black R R Photoshop CS5 Extended (Version manually with Adobe 12.0.1.x64) and thus every vessel could be acquired with the analysis tool. Incomplete filling was present in approximately R 50% of the samples following Microfil perfusion but only in about 20% of the samples that were visualized by means of India ink. By refining our perfusion protocol, we were able to decrease this number to single figure percentages. Shrinking R of the dye/Microfil was more evident in larger vessels and rather negligible in smaller vessels. Automatic image analysis The automatic vessel segmentation is based on a ‘negative & positive experts’ model (Kalal et al., 2010) and a (corrected) nearest neighbour classification/k-means clustering (MacQueen, 1967; Seber, 1984) and is realized in MATLAB (using the Image Processing Toolbox and the Statistics Toolbox). The necessary source code for building the application is available for download in the supplementary information. The greyscale images and colour images contain a different amount of information (ambiguous and nonambiguous) and as a result they undergo slightly different postprocessing workflows. Both workflows result in binary maps, containing the segmented vessels, which are then statistically analyzed by the same algorithm. The novelty of the statistics algorithm is the vessels’ processing as objects with a vectorized distance from the main vessels instead of absolute coordinates within the entire image. In the following, the workflow for greyscale input images from unstained histological sections is described (Fig. 3). After the image is loaded in MATLAB and saved as a matrix, the user selects positive and negative experts via a graphical user interface to build a model for the nearest neighbour classification. Areas that represent vessels should be specified as positive experts, areas representing the sample’s background (fibrin, etc.) should be specified as negative experts. For each expert, the mean value and the SD of the signal intensity, and the selected area are calculated. The one-dimensional Euclidean distances to the mean signal intensity of the negative and the positive experts are computed for the signal intensities of all pixels of the images. The pixels are then classified depending on the minimal Euclidean distance (Figs. 4C, D). The result of the nearest neighbour classification is a binary map (segmented vessels: TRUE, other: FALSE). To validate and adjust the segmented vessels, the greyscale image is overlaid with the binary mask (Fig. 4E). The classification can furthermore be influ-

Fig. 3. Algorithm for segmenting vessels from greyscale images. Mean( . . . ) represents the mean value, std( . . . ) represents the SD. POS and NEG are abbreviations for the positive and negative experts. SI is the signal intensity.

enced by adding/subtracting multiples of the positive expert’s SD to its mean value. Increasing the amount of SD leads to more pixels classified as vessels and vice versa. To eliminate single-pixel artefacts, the area of all segmented vessels is analyzed. Only vessels larger than the smallest positive expert selection are transferred to the final binary map. Whereas the unstained samples (Fig. 4A) contain only a small amount of ambiguous information regarding the definite vessel segmentation, the stained histological sections show dark objects, e.g. the cell nuclei in H&E-stained sections that impair the segmentation process based on greyscale images as described above. Therefore, truecolour images have to be acquired to be able to use additional information gained from the different colours of dark objects and stained vessels. In the following, the workflow for truecolour images is described (Fig. 5) using an H&E-stained section as an example (complete image: Fig. 6A, magnified subsection: Fig. 6B). The algorithm is basically the same as for greyscale images, but with added features  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

AUTOMATIC QUANTIFICATION OF ANGIOGENESIS IN 2D SECTIONS

5

Fig. 4. Sample images from the greyscale vessel segmentation workflow: (A) Overview of an AV loop’s histological section. The vessels were stained solely with India ink, no H&E staining was used. The image was acquired with brightfield microscopy (greyscale, resolution + MIA + automatic registration). The postprocessing stream is shown in detail for the highlighted section (grey box). (B) Detailed image of the highlighted area in (A). Vessels are black, background spreads from greyscales white to grey. (C) Binary mask calculated from the selection of the positive experts (vessels). Affiliation to the class ‘positive expert’ is shown in white. (D) Binary mask calculated from the selection of the negative experts (background). Affiliation to the class ‘negative expert’ is shown in white. (E) Overlay of the original image (B) with the segmented vessels (green).

for the evaluation of the colour information. After loading the image and the selection of positive and negative experts (POS: vessels; NEG: dark objects such as nuclei, tissue background), the image is transformed from Red-Green-Blue (RGB) to the Lab colour space CIELAB(Hunter, 1958). In terms of finding chromatic similarities between two colours, the CIELAB colour space is superior to RGB, because the 2D Euclidean distance can be used to measure the similarities instead of a more complicated 3D model. As in the greyscale workflow (Fig. 3), for each expert the mean value and SD of the signal intensity is calculated – for each CIELAB channel (L: lightness, a & b: colour information) a pair of variables is obtained. In a first step, the L channel is used to classify the image’s pixels in the three classes background, nuclei, and vessels according to their minimal (1D) Euclidean distance. Based on this classification method, the sample’s bright background (Fig. 6C) and the darker objects (vessels, nuclei, etc.) can be separated very easily. The resulting binary map for the vessels still contains  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

dark objects because they are in the same L value range as the vessels. To correct this, the additional colour information from channels a and b is taken into account. For all positive classified pixels, the (2D) Euclidean distance to the positive experts (vessels) and the negative experts (e.g. nuclei) on the a/b plane is calculated. Based on the nearest neighbour classification, binary maps for the dark objects (Fig. 6D) and the vessels (Fig. 6E) are created. For further validation and redefining the segmentation, the truecolour image is overlaid with the vessels’ binary map (Fig. 6F). Analogous to the greyscale workflow, the segmentation can be redefined by changing the amount of SD used in the classification. To eliminate singlepixel artefacts, the area of all segmented vessels is analyzed. Only vessels that are bigger than the smallest positive expert selection are transferred to the final binary map. Once the vessels are segmented from either the unstained sample, imaged in greyscale (Fig. 4A), or the stained sample (Fig. 6A), imaged in truecolour, the vessels are statistically

6

C. WEIS ET AL.

to calculate a whole new set of parameters (e.g. cumulative amount of vessels vs. distance to main vessels) and allows on the other hand a better comparison between the samples (even of different sizes). Before the final statistics are calculated, the user is presented with a graphical user interface to assign input and output to the two main vessels and top and bottom of the complete sample, to be able to compare between multiple probes. For each sample, the user can choose between following statistical analyses:

r

r r r

Fig. 5. Algorithm for segmenting vessels from RGB images. Mean( . . . ) represents the mean value, std( . . . ) represents the SD. POS and NEG are abbreviations for the positive and negative experts. L, A and B are the channels of the CIELAB colour space.

Standard analysis: The sample is divided into four regions: input (centre), input (periphery), output (centre) and output (periphery). The centre parts cover the area between the two main vessels. The periphery parts cover the rest. For all areas, the amount of vessels and the cumulative area of the vessels are calculated. Cumulative analysis: For the input, the output and the complete sample, the cumulative amount of vessels and the cumulative vessels’ area are calculated for the distance to the corresponding main vessel. Binned analysis: According to the vessels’ distance to the main vessels, they are grouped in distance bins predefined by the user. The amount of vessels and the vessels’ area are calculated for each bin. Ideal Map: To be able to compare the vessel´s locations and sizes for different samples, the main vessels are idealized. Each vessel is plotted at the same location as before, it is represented as a circle with the same area size as the original vessel.

Reproducibility and inter-reader variability of the algorithm’s results were tested by analyzing two samples; each of them five times independently. Afterwards, the average and the SD of the quantity and the cumulative area of vessels were calculated. The SDs were then normalized (SD/average), resulting in a percentage of how widely scattered the results are compared to the average values. Comparison to other image analysis methods

analyzed with the same workflow described in the following (Fig. 7). At the beginning, the user defines the region of interest for which the two main vessels are then automatically assigned and the vectorized Euclidean distance for each vessel of the binary map is calculated. If the automatic assignment did not work or the main vessels were not properly stained, the algorithm provides the possibility for correction and subsequent evaluation. For each vessel, the Euclidean distances to both main vessels are calculated, the minimum is saved along with the corresponding vector and the vessel’s area. This vectorized parameterization, which is the equivalent of a coordinate system transformation, enables on the one hand

To evaluate our algorithm’s results, we compared it to a previously established image analysis method (Arkudas et al., 2009). The number of blood vessels, SD of blood vessel size and average size of blood vessels were determined automatically for all stained cross-sections by separating the truecolour image in its three colour channels and thresholding the desired colour with a predefined value. Statistical analysis Unless noted otherwise, all results are presented as mean ±SE. Statistical analysis was performed using Excel (Microsoft Corporation) and MATLAB (MathWorks). Statistical significance  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

AUTOMATIC QUANTIFICATION OF ANGIOGENESIS IN 2D SECTIONS

7

Fig. 6. Sample images from the truecolour vessel segmentation workflow: (A) Overview of an AV loop’s histological section. The vessels were stained with India ink, the completed sample was stained with H&E. The image was acquired with brightfield microscopy (RGB, resolution + MIA + automatic registration). The postprocessing stream is shown in detail for the highlighted section (grey box). (B) Detailed image of the highlighted area in (A). Vessels are black, nuclei are dark blue and sample background is red-violet. (C) Binary mask calculated from the selection of the negative experts (background). Affiliation to the class ‘negative expert’ is shown in white. (D) Binary mask calculated from the selection of the negative experts (nuclei). Affiliation to the class ‘negative expert’ is shown in white. (E) Binary mask calculated from the selection of the positive experts (vessels). Affiliation to the class ‘positive expert’ is shown in white. (F) Overlay of the original image (B) with the segmented vessels (green).

of the difference between groups was determined using the Mann-Whitney-Wilcoxon test. Statistical significance was established at p < 0.05 (*). Number and area of blood vessel per cross-section were calculated using MATLAB (MathWorks). Results

verified due to a successful perfusion of Ink solution in each animal. At the three different explantation points, the matrix exhibited no signs of fibrinolysis or degradation. The perfused main vessel axis was macroscopically visible through the matrix in each animal. All explanted constructs underwent further morphometrical analysis.

Basic data Complications caused by anaesthesia or surgery were absent in all animals (n = 8). There were no signs of wound dehiscence or infection in either group. A fibrous capsule surrounding the chamber was evident in all animals at the stage of explantation. Two rats (15d group) underwent serous fluid puncture 1 week following surgery. At the time of explantation, patency of the AV loop and absence of thrombosis was  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

Automatic image analysis After loading the merged histological section of the complete sample, the user was guided step by step through the model building, segmentation, and statistics processes. On average, the analysis of an RGB image (15 000 pixel × 10 000 pixel) took 25 min on a commercial 4GB 2GHz DuoCore Unix computer. The translation from RGB to CIELAB and the creation of

8

C. WEIS ET AL.

were only a proof-of-concept, the data were not interpreted (but are shown in the SI). Evaluating the inter-reader reproducibility of the parameters calculated with our method showed that the cumulative area varied only 3.68% from the average value (SD/average, two blinded observers, two samples each evaluated five times). The quantity of vessels varied 10.35% from the average (SD/average, two blinded observers, two samples each evaluated five times). The inter-reader reproducibility of the method, using channel separation and thresholding, depends on how the threshold was set. If the threshold was the same for all samples, the inter-reader deviation was 0% (cumulative area and amount of vessels). If each reader set the threshold by herself/himself for each sample individually, the cumulative area varied 7.46% from the average value (SD/average, two blinded observers, two samples each evaluated five times). The total amount of vessels varied 18.64% from the average (SD/average, two blinded observers, two samples each evaluated five times). Comparing our method’s results (Fig. 8B) with the thresholding approach (Fig. 8C) revealed that the conventional colour-channel thresholding counted 10% more vessels than our approach (Fig. 8F, shown for three representative samples). The cumulative area, however, was nearly the same (Fig. 8G, shown for three representative samples). Microscopic and morphometrical analysis

Fig. 7. Algorithm to analyze the segmented vessels to create statistics.

the segmentation model were the most time-consuming processes of the analysis (90% of the processing time), whereas user interaction was limited to 10% of the algorithm’s run time. All user interactions were announced by the program with detailed descriptions of the required input. Therefore, training of new users took only two to three (supervised) image analyses (1 h). Because of the detailed description preceding the steps requiring user input, there was no difference between the results of self-trained users and users with supervised training. Although the user input steps are very convenient, basic computer skills are advantageous to run the automatic analysis, described in this paper. Based on the vectorized vessel coordinates, all the analyses described above were performed (standard analysis, cumulative analysis [Fig. 8D] and binned analysis), as well as the ideal maps (Fig. 8E) were calculated. Since the analyzed constructs

In all samples, main vessel axis was detected easily as it featured the largest diameter (Fig. 9). Five days after AV Loop preparation, vascular growth was spare and located only in close proximity to the main axis which was formed by the venous AV loop graft (Fig. 9). At day 10 and 15, strong neovascularization was evident without the need of additional growth factors (Figs. 9 and 10). Although mean vessel area increased between day 10 and 15, there was no significant differences in vessel number, suggesting that vessel differentiation becomes more important during this late period. Inflammatory cells such as neutrophils, macrophages, lymphocytes and fibrocytes were present at every time point, whereas migration into the centre of the matrix was only observed at day 10 and 15. In two samples, areas with grey discolouration were observed, which indicated leakage of India ink during perfusion. However, specimens were not excluded from the trial and were further evaluated. Micro-CT The AV loop-associated neovascularization was visualized C perfusion during by means of micro-CT following Microfil explanting. After 15 days (Fig. 2), strong luminal neovascularization was observed that arose in a perpendicular manner from the venous graft. Sprouting of small vessels was evident over the entire length of the graft. A high amount of vessel  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

AUTOMATIC QUANTIFICATION OF ANGIOGENESIS IN 2D SECTIONS

9

Fig. 8. Comparison between conventional thresholding and our model-based algorithm. (A) H&E-stained sample of an AV loop. Vessels express a dark colour due to the Microfil perfusion. (B) Result of the conventional thresholding method. The sample is shown in greyscale. The dark objects segmented by the conventional approach as vessels are overlayed in blue. By using signal intensity thresholding solely, nuclei stained by H&E were also segmented. (C) Result of our model-based algorithm. The sample is shown in greyscale. The segmented vessels are overlayed in blue. Our model-based segmentation is able to distinguish between H&E-stained nuclei and vessels. (D) Cumulative vessel count versus distance to main vessel for the complete sample and separated for input (vessels located closer to the chamber entrance, inflow of the main vessel axis) and output (vessels located closer to the chamber exit, outward flow of the main vessel axis). The segmentation shown in C is represented by the output curve. (E) Ideal map of the sample shown in A. The main vessel is located in the centre (blue cross). The idealized vessels are placed according to their vectorized parameterization (distance and diameter). (F) Amount of segmented vessels shown for three representative samples. The conventional approach counts 10% more vessels than our model-based approach. (G) Cumulative area of vessels shown for three the same representative samples. The cumulative area calculated by both algorithms is nearly the same.

formation was also apparent at the entrance of the chamber most likely due to an ingrowth from the surrounding tissue. After aligning the histological sections with the corresponding slices from the CT data set and analyzing them with the proposed algorithm, the quantity of vessels for the sample imaged with a microscope was 18-fold higher than the vessel count for the micro-CT slice (Fig. 11A). However, the cumulative vessel area was 1.5-fold larger for micro-CT (Fig. 11B). Discussion Angiogenesis plays a crucial role in tissue engineering and is essential to overcome the limitations of mass transfer in this context. Therefore, formation of new blood vessels is a core element of any tissue construct grown in vitro or in vivo regardless of the tissue-specific goal (Lovett et al., 2009) . This study introduces an automatic quantitative analysis method of angiogenesis that uses microscopic images of stained histological cross-sections to assess not only the amount but also the spatial relations of the newly formed vessels. Compared to the conventional threshold-based method, a similar accuracy was obtained with the new algorithm as we analyzed the AV loop-associated angiogenesis at different time points. However, our algorithm was able to distinguish between different types of dark objects. Thus, it was superior to  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

exclude dark stained nonvessel structures. Due to the demonstrated resolution limit, micro-CT was not able to compete with the 2D analyses in terms of precise evaluation of vascular growth in small tissue-engineered constructs. To test the reliability and effectiveness of our algorithm to characterize vessel formation, we used the rat AV loop model as it was described to provide calculable neovascularization without the need of additional angiogenic factors (Polykandriotis et al., 2008; Boos et al., 2010; Beier et al., 2011). The model was first established by Erol and Spira (1980) and latter modified by Mian et al. (2000), who embedded the shunt into a chamber. This construct allows examining and characterizing neoangiogenesis in a setting isolated from the donor’s body. An automatic quantitative micro-CT evaluation algorithm for the associated angiogenesis was already introduced for the AV loop model (Arkudas et al., 2010). To evaluate our algorithm’s applicability to the everyday lab routine, we analyzed the temporal dynamics of the AV loop-associated vessel formation. We observed a significant increase of vessel amount until day 10 following AV loop preparation. In contrast to the total vessel area, the increase of vessel number between day 10 and day 15 was not significant, suggesting that in this late period vessel differentiation is predominant. The results obtained from the characterization of the AV loop angiogenesis served to assess the quality of the

10

C. WEIS ET AL.

Fig. 9. De novo angiogenesis 5 and 15 days following AV loop preparation. After injection of black-appearing India ink into the aorta cross-sections of the explanted and paraffin-embedded AV loop, constructs were obtained perpendicular to the longitudinal AV loop axis which are the two large central vessels. Samples are stained with hematoxylin and eosin. Five days after AV loop creation (A), new vessel formation is nearly undetectable. However, after 15 days (B) strong luminal neovascularization arising from the venous graft vessel is evident. Scale bars: 1000 µm

introduced algorithm. Thus, the newly established evaluation tool was able to segment vessels unambiguously und user-independently. Due to high resolution of the histological images acquired by means of bright field microscopy (pixel width/height of 3.2 µm for RGB and 6.45 µm for greyscale), even small vessels in the range of the microscope’s detection limit could be localized. Due to the ‘positive & negative experts’ model used for classification (Figs. 3 and 5) the algorithm is applicable to greyscale images (Fig. 4) and RGB images with any staining type (Fig. 6). Furthermore, the tool is able to disR tinguish between the dark Microfil -stained vessels and other dark objects, e.g. dark nuclei resulting from the H&E staining (Fig. 6F). Repetitive vessel segmentation was performed in the same sample to test the reproducibility of the algorithm. Whereas the cumulative area of the segmented vessels varies only 4%, the quantity of segmented vessels varies 10% on average. This is due to the user’s selection of the smallest positive expert (the smallest selected vessel). A lower area threshold results in more vessels that are transferred to the final binary map. Although the small vessels are counted with the same weight as vessels with a big area, their contribution to the cumulative vessel area is negligible.

Fig. 10. Vessel count and area over time (5 d, n = 2; 10 d, n = 3; 5 d, n = 3). Significance level is p < 0.05 (*). The error bars represent the SE. (A) Vessel count: the quantity of vessels increases over time. There is a significant difference (*) between day 5 and day 10. However, after 10 days increase of vessel amount is not significant. (B) Vessel area: the cumulative area of the vessels increases significantly (*) over the whole observation period.

Comparing our method’s reproducibility with conventional thresholding showed that our algorithm expressed a higher inter-reader variability than approaches with one predefined threshold for all samples. This was a result of the elimination of user interaction within these fully automatic approaches. But compared to methods where users set the threshold for each sample individually, our approach expressed a lower interreader variability. This was a result of our segmentation model based on ‘positive experts’ and ‘negative experts’. Users with the same skill set tended to classify vessels and nonvessels in the same way. These classifications were then used to create the segmentation model, which itself was based on multiple parameters and not on only one colour channel. In addition to an increased user independence, our algorithm has one big advantage compared to conventional thresholding: it is able to distinguish between different types of dark objects. Analyzing stained samples (Fig. 8A), where  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

AUTOMATIC QUANTIFICATION OF ANGIOGENESIS IN 2D SECTIONS

Fig. 11. Vessel count and area analyzed for the complete sample to compare the micro-CT’s and the histological sections’ results. Histological samples were registered to the corresponding cross-sections of the microCT data set prior to analysis. The graphs show the results for the 15 day explantation interval. The significance level is p < 0.05 (*). The error bars represent the SE. (A) Vessel counts: the quantity of segmented vessels is significantly (*) higher for the histological sections compared to the CT’s cross-section. (B) Vessel area: the cumulative vessel area calculated for the CT data set is larger than the area of the segmented vessels in the histological sections. However, the difference is not significant.

darker pixels can result from vessels as well as from H&Estained nuclei, revealed that conventional thresholding also counted the nonvessel structures (Fig. 8B), whereas our algorithm only counted the vessels (Fig. 8C) and ignored other dark objects based on the experts described above. This unambiguous classification is also the reason why our algorithm counts 10% less vessels than the conventional thresholding approaches (Fig. 8F). Whereas all objects, which were darker than a certain threshold, were classified by conventional approaches as a vessel, our method only classified and counted dark objects, which exactly met the criteria specified by the ‘positive experts’ (size & coordinates in CIELAB). This effect, however, is only valid for very small objects. Comparing the cumulative areas calculated of our and the conventional approach, showed that there were nearly no deviations (Fig. 8G). In previous studies, micro-CT along with 3D reconstruction was proposed as an additional method for the analysis of AV loop constructs (Arkudas et al., 2010). This method is ideal for analyses of the complete 3D volume, e.g. the vessel tree  C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy 

11

hierarchy. For 2D analyses, for example, vessel density, vessel area and quantity of vessels, the proposed algorithm, however, surpassed the micro-CT in terms of applicability (micro-CT and the analysis of the 3D reconstruction is time-consuming and expensive), user independence, and, due to imaging with microscopy, spatial resolution (in our case 0.32 µm for the microscope and 8–15 µm for the micro-CT). The analysis based on histological sections was able to identify a higher quantity of vessels than the 3D reconstruction based on micro-CT (Fig. 11A). The calculated vessel area per cross-section, however, appeared to be bigger in the micro-CT images than in the histological sections (Fig. 11B). Both effects were a result of the different spatial resolution limits of the used techniques. Aggregations of multiple objects, of which the dimensions are below the resolution limit, cannot be distinguished properly. Microscopy was able to resolve those regions around the main vessels where the density of small vessels was very high. MicroCT, however, was not able to resolve those regions. Furthermore, the radio-dense material (in our case Microfil within the vessels) has a stronger influence on the CT’s results, than the surrounding radiolucent tissue. Therefore, areas with a high vessel density are ‘detected’ as one big vessel (covering even the nonvessel regions) instead of multiple smaller ones. As a result, micro-CT obtained a bigger cumulative area than microscopy (Fig. 11B). Single objects, of which the dimensions are below the resolution limit, cannot be detected properly. Our algorithm utilized microscopy’s high resolution, to be able to distinguish small vessels against the sample’s background. The resolution limit of micro-CT was too high to perform this differentiation. This was one of the reasons why the amount of vessels detected by microscopy was higher than the amount of vessels detected by micro-CT (Fig. 11A). The second reason was the inhomogeR neous perfusion of the samples with Microfil that cannot be corrected manually before the CT scans. Haemodynamic conditions of the neovasculature arising from the AV loop graft are unique due to the orthogonal orientation of the newly formed vessels to the loop axis impairing dye distribution into the generated microvessels in a few cases.

Conclusion In this study, an automatic quantitative histological evaluation algorithm was established to improve the evaluation of angiogenesis in tissue engineered constructs but also in other field of research where vessel growth is of special importance (e.g. cancer-related angiogenesis). The proposed method is applicable to greyscale images and RGB images acquired of stained histological samples. The algorithm segments and analyzes vessels unambiguously and user independently. In terms of resolution, speed, costs, reproducibility and the results’ reliability, the 2D automatic quantitative histological evaluation algorithm surpasses previous presented methods.

12

C. WEIS ET AL.

In a proof-of-concept study, the algorithm allowed to characterize the temporal dynamics of the AV loop-associated neovascularization in a high-quality manner. Acknowledgements This work was supported by the Interdisciplinary Center for Clinical Research (IZKF) at the University Hospital of the University Erlangen-Nuremberg (to VJ Schmidt), the Emerging Fields Initiative of the Friedrich-Alexander-University Erlangen-Nuremberg, and by grants from the International Max-Planck Research School Erlangen. Additionally, the authors want to thank Ilse Arnold, Marina Milde and Stefan Fleischer for their assistance in the laboratory. Conflicts of interest The Authors disclose any commercial associations that might create a conflict of interest in connection with the submitted manuscript. Competing financial interests of all authors have been appropriately disclosed according to the policy of the Journal. All conflicts of interest, whether they are actual or potential, are disclosed. References Arkudas, A., Pryymachuk, G., Hoereth, T., Beier, J.P., Polykandriotis, E., Bleiziffer, O., Horch, R.E. & Kneser, U. (2009) Dose-finding study of fibrin gel-immobilized vascular endothelial growth factor 165 and basic fibroblast growth factor in the arteriovenous loop rat model. Tissue Eng. Part A 15, 2501–2511. Arkudas, A., Beier, J.P., Pryymachuk, G., et al. (2010) Automatic quantitative micro-computed tomography evaluation of angiogenesis in an axially vascularized tissue-engineered bone construct. Tissue Eng. Part C 16, 1503–1514. Arkudas, A., Balzer, A., Buehrer, G., et al. (2013) Evaluation of angiogenesis of bioactive glass in the arteriovenous loop model. Tissue Eng. Part C 19, 479–486. Beier, J.P., Hess, A., Loew, J., et al. (2011) De novo generation of an axially vascularized processed bovine cancellous-bone substitute in the sheep arteriovenous-loop model. Euro. Surg. Res. 46, 148–155. Boos, A.M., Arkudas, A., Kneser, U., Horch, R.E. & Beier, J.P. (2010) Bone tissue engineering for bone defect therapy. Handchir. Mikrochir. Plast. Chir. 42, 360–368. Carmeliet, P. & Jain, R.K. (2011) Principles and mechanisms of vessel normalization for cancer and other angiogenic diseases. Nat. Rev. Drug Discov. 10, 417–427. Erol, O.O. & Spira, M. (1979) Inlay technique for protection of skin and composite grafts in experimental animals. Plast. Reconstr. Surg. 64, 271–272. Erol, O.O. & Spira, M. (1980) New capillary bed formation with a surgically constructed arteriovenous fistula. Plast. Reconstr. Surg. 66, 109– 115.

Hayes, C.E. & Goldstein, I.J. (1974) An alpha-D-galactosyl-binding lectin from Bandeiraea simplicifolia seeds. Isolation by affinity chromatography and characterization. J. Biol. Chem. 249, 1904–1914. Hunter, R.S. (1958) Photoelectric color difference meter. J. Opt. Soc. Am. 48, 985–993. Kalal, Z., Matas, J. & Mikolajczyk, K. (2010) PN learning: bootstrapping binary classifiers by structural constraints. In Proceedings of the 2010 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, New York, NY, USA. Kneser, U., Polykandriotis, E., Ohnolz, J., et al. (2006) Engineering of vascularized transplantable bone tissues: induction of axial vascularization in an osteoconductive matrix using an arteriovenous loop. Tissue Eng. 12, 1721–1731. Lovett, M., Lee, K., Edwards, A. & Kaplan, D.L. (2009) Vascularization strategies for tissue engineering. Tissue Eng. Part B 15, 353–370. MacQueen, J. (1967) Some methods for classification and analysis of multivariate observations. In: Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, pp. 281–297. University of California Press, California. Mian, R., Morrison, W.A., Hurley, J.V., Penington, A.J., Romeo, R., Tanaka, Y. & Knight, K.R. (2000) Formation of new tissue from an arteriovenous loop in the absence of added extracellular matrix. Tissue Eng. 6, 595–603. Polykandriotis, E., Arkudas, A., Horch, R.E., Sturzl, M. & Kneser, U. (2007) Autonomously vascularized cellular constructs in tissue engineering: opening a new perspective for biomedical science. J. Cell. Mol. Med. 11, 6–20. Polykandriotis, E., Tjiawi, J., Euler, S., et al. (2008) The venous graft as an effector of early angiogenesis in a fibrin matrix. Microvasc. Res. 75, 25–33. Rath, S.N., Strobel, L.A., Arkudas, A., Beier, J.P., Maier, A.K., Greil, P., Horch, R.E. & Kneser, U. (2012) Osteoinduction and survival of osteoblasts and bone-marrow stromal cells in 3D biphasic calcium phosphate scaffolds under static and dynamic culture conditions. J. Cell. Mol. Med. 16, 2350–2361. Schmidt, V.J., Hilgert, J.G., Covi, J.M., Weis, C., Wietbrock, J.O., de Wit, C., Horch, R.E. & Kneser, U. (2013) High flow conditions increase connexin43 expression in a rat arteriovenous and angioinductive loop model. PloS One 8, e78782. Seber, G.F. 1984. Multivariate Observations. Wiley, Hoboken, NJ. Strobel, L.A., Rath, S.N., Maier, A.K., Beier, J.P., Arkudas, A., Greil, P., Horch, R.E. & Kneser, U. (2014) Induction of bone formation in biphasic calcium phosphate scaffolds by bone morphogenetic protein-2 and primary osteoblasts. J. Tissue Eng. Regen. Med. 8, 176–185. Tanaka, Y., Tsutsumi, A., Crowe, D.M., Tajima, S. & Morrison, W.A. (2000) Generation of an autologous tissue (matrix) flap by combining an arteriovenous shunt loop with artificial skin in rats: preliminary report. Br. J. Plast. Surg. 53, 51–57.

Supporting Information Additional Supporting information may be found in the online version of this article at the publisher’s website.

 C 2015 The Authors C 2015 Royal Microscopical Society, 00, 1–12 Journal of Microscopy