Available online at www.sciencedirect.com R Microvascular Research 66 (2003) 113–125 www.elsevier.com/locate/ymvre Automated tracing and change analysis of angiogenic vasculature from in vivo multiphoton confocal image time series Muhammad-Amri Abdul-Karim,a Khalid Al-Kofahi,a Edward B. Brown,b Rakesh K. Jain,b and Badrinath Roysama,* a b Department of Electrical, Computer and Systems Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180, USA Edwin L. Steele Laboratory, Department of Radiation Oncology, Massachusetts General Hospital and Harvard Medical School, Boston, MA 02114, USA Received 15 January 2003 Abstract Automated methods are described for tracing and analysis of changes in angiogenic vasculature imaged by a multiphoton laser-scanning confocal microscope. Utilizing chronic animal window models, time series of in vivo 3-D images were acquired on approximately the same target volume of the same specimen while undergoing angiogenic change (typically every 24 h for 7 days). Objective, precise, 3-D, rapid, and fully automated vessel morphometry was performed using an adaptive tracing algorithm that is based on a generalized irregular cylinder model of the vasculature. This algorithm was found to be not only adaptive enough for tracing angiogenic vasculature, but also very efficient in its use of computer memory, and fast, taking less than 1 min to trace a 768 ⫻ 512 ⫻ 32, 8-bit/pixel 3-D image stack on a Dell Pentium III 1-GHz computer. The automatically traced centerlines were manually validated on six image stacks and the average spatial error was measured to be 2 pixels, with an average concordance of 81% between manual and automated traces on a voxel basis. The tracing output includes geometrical statistics of traced vasculature and serves as the basis of statistical change analysis. The computer methods described here are designed to be scalable to much larger hypothesis testing studies involving quantitative measurements of tumor angiogenesis, gene expression relative to known vascular structures, and impact of drug delivery. © 2003 Elsevier Science (USA). All rights reserved. Keywords: In vivo change analysis; Angiogenesis; Tumor vasculature; Vasculature tracing; Vasculature segmentation; Automated morphometry; Multiphoton microscopy Introduction An automated method is described for statistical change analysis of in vivo tumor vasculature. It consists of three phases: intravital microscopy, automated vasculature morphometry, and statistical change analysis. Changes occurring in tumor vasculature, and in preexisting vasculature bed next to a tumor, provide insight into tumor pathophysiology, which includes gene expression, angiogenesis, vascular transport, and drug delivery (Jain et al., 2002; Folkman, 2001; Carmeliet and Jain, 2000; Auerbach et al., 1991). * Corresponding author. ECSE Department, Rensselaer Polytechnic Institute, Troy, NY 12180-3590, USA. Fax: ⫹1-518-276-8715. E-mail address: roysam@ecse.rpi.edu (B. Roysam). First, in vivo image acquisition by intravital microscopy is performed using the multiphoton laser-scanning microscope (MPLSM), aided by a variety of chronic animal window preparations described by Brown et al. (2001). The live specimen, while undergoing angiogenesis, is imaged on approximately the same volume, over a period of time (typically every 24 h for 7 days), producing a time series of 3-D image stacks (see Figs. 1– 4). These images reveal how a preexisting vascular bed is altered as a tumor grows into the imaged region. Precise vasculature segmentation is then performed on the image stacks using a fully automated 3-D vasculaturetracing algorithm. This model-based algorithm extends our prior work on tracing dye-injected neuron images (Al-Kofahi et al., 2002) and retinal angiograms (Shen et al., 2001; Can et al., 1999), using a set of directional edge detectors to 0026-2862/03/$ – see front matter © 2003 Elsevier Science (USA). All rights reserved. doi:10.1016/S0026-2862(03)00039-6 114 M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 Using the statistics generated by the tracing algorithm on time series images, change analysis is performed by comparing morphometric statistics of the vasculature within a region of interest common to all images in the sequence. Tabulation of the statistics together with visual displays in the form of progression graphs is performed to reveal and describe vasculature changes (Table 1 and Fig. 6). These change measurements may then be used for hypothesis testing, pattern classification, or other decision-making processes. Materials and methods Specimen preparation and imaging 3-D image stacks were acquired in vivo using an MPLSM on Severe Combined Immunodeficiency Disease Fig. 1. (A) Sample 768 ⫻ 512 ⫻ 32, 8-bit/pixel 3-D in vivo image of a fluorescently labeled vasculature of skin altered by growth of a nearby tumor, presented by its projections (x–y, y–z, and x–z). Note the amount of background noise and intensity nonuniformities within the vasculature. The size scale varies from 5 to 20 pixels wide. Assumption of Gaussian intensity profile within vasculature is not valid in this image. Volumes highlighted are common overlapping regions across all images in this set of time series images. (B) Projections of the traces on each image projection. For illustrative purposes, each vasculature segment is labeled. The tracing algorithm adds several extra slices before the top slice and after the bottom slice for computational consistency issues. Note that the program ignores most of the background noise. trace and segment vasculature that satisfy a generalized cylinder model (Fig. 5). For this study, the algorithms were modified to handle higher tortuosity, high size-scale variability, nonuniform brightness, and irregular structure of tumor vasculature, which are not typical features in neuron images. Being model-based, the algorithm introduced in this paper overcomes limitations of intensity-based methods by having the built-in notion of a physical object model, rather than being based solely on intensity. Additionally, the algorithm overcomes the limitations of line-filtering methods by avoiding the Gaussian cross-sectional profile assumption, and by performing calculations only on the image foreground. Finally, the algorithm does not suffer from the subjectivity and tedium associated with manual tracing. Fig. 2. (A) The image of approximately the same volume as in Fig. 1 on Day 2. The volume that represents the common overlapping region among all images in this time series is highlighted. The background noise and intensity nonuniformities within the vasculature are more visible in this image compared to the image of the first day (Fig. 1). (B) Projections of the traced image. As expected, the traced centerlines are minimally affected by the imaging artifacts. M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 115 Table 1 Experimental change analysis results for time series tumor vasculature imagesa,b,c Day 1 2 3 4 Total vessel segments Total length of vessels (m) Average horizontal width (m) Average vertical width (m) Data Change index Data Change index Data Change index Data Change index 140 160 170 95 1.143 1.063 0.559 6572.1 7001.0 7334.8 4541.1 1.065 1.048 0.619 4.81 5.13 5.75 5.81 1.068 1.120 1.011 2.89 2.82 2.95 3.02 0.976 1.046 1.022 — — — — a A set of four time series images taken every 24 h, shown in Figs. 1– 4. Measurements are restricted to the common overlapping volume among all images. c Change Index is the ratio of the current data over the data on the previous day. b (SCID) mice prepared using various chronic window preparations. The reader is referred to a recent paper by Brown et al. (2001) for a much more detailed description of the specimen preparation methods. In Figs. 1– 4, a murine mammary adenocarcinoma (MCaIV) was implanted in the center of the dorsal skinfold window. Under anesthesia, blood vessels were highlighted using fluorescently labeled dextrans injected intravenously and a volume of preexisting vasculature was imaged every 24 h for 7 days as the tumor grew over the site. In Fig. 7, vasculature in the cranial window was imaged after highlighting using fluorescent dextrans injected intravenously. The resulting 3-D images are typically 22–141 optical slices of size 768 ⫻ 512 each, with typical intraslice resolution of 0.72 m and interslice resolution of 5 m per voxel. Each 4-D dataset has typically seven 3-D images, taken 24 h apart over a week. The cranial window images (Fig. 7) required smoothing before tracing could be attempted. To process such images, a morphological preprocessor toolkit was integrated into the implementation of the tracing algorithm. It consists of standard 3-D grayscale mathematical operations such as erosion, dilation, closing, and opening, as well as rank filters such as the median filter. The built-in structuring elements for these operations are ellipsoids and cubes with configurable size and axial scales. For example, a z scale of 1 is a special case corresponding to 2-D structuring elements on the x–y plane. These operations can be performed prior to tracing if necessary to remove unwanted image artifacts. Unless stated otherwise, all results presented in this paper were produced without any preprocessing. Automated vessel tracing methods Segmentation is an essential first step for vessel morphometry, which in turn is used for statistical quantitation of vascular changes (Jain et al., 1997; Barbareschi et al., 1995; Leunig et al., 1992). Methods for vascular segmentation can be grouped into three main categories: (i) intensity-based segmentation; (ii) line-filtering methods; and (iii) manual and semiautomatic methods. Methods in the first category utilize intensity-based thresholds and are generally susceptible to background noise (e.g., Brey et al., 2002; Holmes et al., 2002; Brown et al., 2001; Wild et al., 2000; Parsons- Wingerter, et al., 1998; Seifert et al., 1997; Tjalma et al., 1998; Toi et al., 1996; Fox et al., 1995; Jakobsson et al., 1994; Avinash et al., 1993; Kowalski et al., 1992; Rohr et al., 1992). Some methods utilize frequency and size thresholds (Schoell et al., 1997; Iwahana et al., 1996; Nissanov et al., 1995). Some improvements were reported by preprocessing the images with morphological filters (e.g., KumarSingh et al., 1997; Merchant et al., 1994). Methods in the second category use a line filter as the vasculature model (Streekstra and Pelt, 2002; Frangi et al., 1999; Sato et al., 1998). These methods assume a Gaussian cross-sectional vessel profile, and perform expensive calculations on each image voxel; therefore they scale poorly with image size. Methods in the third category involve manual tracing, counting, or visual inspection of the vasculature (Dellas et al., 1997; Dellian et al., 1996; Kirchner et al., 1996; Li et al., 1994; DeFouw et al., 1989; Endrich et al., 1979). These methods, despite being the gold standard, are unavoidably subjective and overly labor-intensive (Al-Kofahi et al., 2003). The vessel-tracing algorithm used in this work is an extension of prior work in the context of tracing 3-D confocal images of dye-injected neurons, which was shown to yield a 97% accuracy compared to multi-user gold standards (Al-Kofahi et al., 2003; Al- Kofahi et al., 2002). The tumor vasculature images of interest exhibit several structural complexities that necessitated several improvements to the previously developed algorithms. Specifically, they exhibit a much higher tortuosity, size-scale variability, and irregularity of geometrical structure. Furthermore, images of tumor vasculature often exhibit nonuniform intensity levels within vessels, mostly due to the dynamics of blood flow relative to the spatiotemporal imaging window. There is also significantly higher variability from one vessel to the next. Change analysis errors often originate from segmentation or tracing errors, which are caused mainly by the type and density of imaging artifacts present in an image. Since the image acquisition phase is performed in vivo and over several days, several types of imaging artifacts are inevitable, caused partly by flow of red blood cells, dye leakage, variability of dye injection, and vasculature structural changes due to respiration, among others. The tracing algo- 116 M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 defined by a base point b, which is the center location of the first LPD correlation kernel in the template, e.g., the point bR. The template orientation in 3-D is defined by the set ⫽ {V, H}, where H is a rotation about the line HH⬘ (i.e., horizontal) and V is a rotation about the line VV⬘ (i.e., vertical) (see Fig. 5). Lines VV⬘ and HH⬘ are orthogonal to each other. In this work, each element in is discretized to 32 discrete values, giving an angular precision of 11.25°, therefore yielding a total of 1024 unique 3-D directions. Our software implementation allows the user to select other discretization values to best fit the images of interest. The template length, k, is defined by the number of LPD correlation kernels stacked together to form the template. Fig. 5 illustrates templates of length 8 in 3-D and Fig. 8 illustrates templates of various lengths in 2-D. The template that is most closely oriented along the generalized cylinder and centered on the cylinder boundary produces the maximum template response, i.e., correlation coefficient (Al-Kofahi et al., 2002). Fig. 3. (A) Day 3 image of approximately the same volume as in Fig. 2. The common overlapping region is highlighted. At this time, more vasculature segments are visible. The amount of background noise is much more apparent than images of previous days. (B) Projections of the traced image. As expected, the tracing algorithm is robust to structural irregularities and intensity nonuniformities present in the image. rithms presented here are designed to achieve a high automation level while being robust to these imaging artifacts. A 3-D mathematical model for describing vessels Al-Kofahi et al. (2002) used the generalized three-dimensional cylinder as a geometric model for dye-injected neurons. This geometrical model is illustrated in Fig. 5, and is sampled at four edges, denoted T ⫽ {tT, tB, tL, tR}, with the subscripts denoting their locations either at the top, bottom, left, or right edge of the vessel cross-section. The four edges are detected using a set of directional templates adapted from the work of Sun et al. (1995) and Can et al. (1999). Each template is composed of a stacked set of directional 1-D low-pass differentiator (LPD) (Sun et al., 1995) correlation kernels of the form [⫺1, ⫺2, 0, 2, 1]T. They are illustrated in Fig. 5. The location of a template in the 3-D image space is Fig. 4. (A) Day 4 image of approximately the same volume as in Fig. 3. The common overlapping region is highlighted. The number of visible vasculature is much less compared to images of previous days. (B) Projections of the traced image. Geometrical statistics generated from these traces are shown in Fig. 6. M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 117 Fig. 5. Isometric view of the vasculature model over a short distance with structural irregularity. Each template shown here is of length 8. The two angular directions, ⫽ {H, V} are rotations around the line HH⬘ (parallel to the z axis) and VV⬘ (orthogonal to HH⬘, i.e., on the x–y plane). Other than the strength of the edge at the four boundaries, this model does not have any inherent assumption on the cross-sectional shape or the intensity profile of the vasculature. This figure also illustrates the robustness of using median statistics for the template responses to handle vascular structural irregularities. As noted in the Introduction, the above model was originally developed for neuron tracing. It must be modified to handle the higher tortuosity, high size-scale variability, nonuniform brightness, and irregular structure of tumor vasculature. As illustrated in Fig. 5, tumor vessels can have a rough surface and a nonuniform cross-section, making the generalized cylinder model inexact. In other words, the need exists for a systematic method for handling limited deviations from the pure generalized cylinder model. In keeping with the need to keep the tracing algorithms scalable to larger data sets, it is also desirable to seek computationally inexpensive methods. A simple method for meeting these requirements is described below. In the prior body of work (Al-Kofahi et al., 2002; Can et al., 1999; Sun et al., 1995), the response of each template was length-normalized by averaging correlation kernel responses along the template length. It is well-known in the statistical literature that the average response is greatly affected by intensity nonuniformities within the structures (Huber, 1981). A more robust alternative to the average response is the median response, which requires computing the median value of correlation kernel responses along the template length. By definition, the median value is robust to as many as 50% outliers (Huber, 1981). The median response R for a template of empirically chosen lengths defined in the set K, at a 3-D location b ⫽ [x y z]T, along the 118 M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 119 Fig. 6. The statistics generated by traces of the images in Figs. 1– 4, also shown in Table 1. (A) Statistics for total vasculature segments present in an image. (B) Statistics for total length of vasculature. (C) Statistics for average horizontal widths. (D) Statistics for average vertical widths (See Fig. 5 for width definitions). Notice that the changes highlighted by the statistics qualitatively agree with the image contents. For example, the average horizontal width increased in the first 3 days, but stayed relatively the same on Day 4 although 38% of the vessels “disappeared.” Currently, the accuracy of vertical width measurements is greatly affected by axial blurring in the images. orientation defined by the unit vector u, is expressed mathematically as: R共b,u,K兲 ⫽ arg max {median (r(b ⫹ ju,u ⬜ ))}, (1) k僆K j⫽1. . .k where r(b,u⬜) is the response of a single 1-D LPD correlation kernel at b along the direction u⬜ that is perpendicular to u. Recall that a template of length k is comprised of k 1-D LPDs stacked together, hence, r is essentially a template of length 1. Notice that this median version of R is also length-normalized since it can only have values returned by r, independent of all k僆K. The maximum median value over all lengths k 僆 K is chosen to be the median response of the template at location b along the direction u. Another aspect of model generalization and refinement originated from observing that tumor vessels resemble a deformed cylinder over a short length in most cases. In other words, over a short distance, linear approximations of vasculature boundaries are not parallel to each other. This violates the assumption of parallel edges in our previous work (Al-Kofahi et al., 2002), which stated that each element in the set of boundary points B must have the same orientation . Here, each template is allowed to shift, ex- Fig. 7. (A) An optical section from a 768 ⫻ 512 ⫻ 141 image stack. Observe the intensity nonuniformity within the vasculature. This was due to moving red blood cells (RBC) during horizontal scan of the specimen. (B) The same optical section, filtered 20 times using a median filter with a spherical structuring element (x–y diameter ⫽ 7 voxels, z diameter ⫽ 1 voxel). The horizontal scan artifacts due to moving RBCs are minimized, while vascular boundaries are preserved. (C) Optical sections 1 through 17 of the image stack in the x–y projection. (D) Segmented vasculature of the optical sections in (C), obtained by tracing, also in the x–y projection, which highlights the accuracy of boundary detection by the tracing algorithm. (E) The entire image stack, 705 m deep, in red– blue anaglyph. (F) Segmented vasculature of the entire image stack, in red– blue anaglyph. This shows that the tracing algorithm is fully 3-D, and applicable to deep image stacks in addition to relatively shallow image stacks shown in Figs. 1– 4. Fig. 10. Illustrates both manual and automated centerline traces superimposed on a maximum intensity projection image for qualitative performance evaluation (quantitative manual-automated concordance measure is 85%), highlighting several areas of interest. (A) False-negative; the vessel has poorly defined edges and poor contrast relative to the local background. (B) Here is where the automated centerline traces are smoother than the manual traces, showing its “steady-hand” effect. (C) Both manual and automated traces agree here in this image, but in subsequent images in the time series dataset, this section becomes more convincing as dye-leakage, hence creating false-positive errors for the automated tracing algorithm in those subsequent images. (D) False-positive; here the strip of dye leakage closely resembles a vessel. (E) False-positive; this relatively dim vessel is occluded by brighter optical slices, hence missed by manual tracing, but not by the automated tracing algorithm. Note the robustness of the automated tracing algorithm to irrelevant blob-like structures, and nonuniform dye distribution within and among vessels. 120 M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 Fig. 8. Example of two pairs of templates that gives the highest response at each iteration i in 2-D. One pair of templates in 2-D consists of left and right templates. Note that at a single iteration i, each template is allowed to elongate, shift, and rotate independently subject to constraints of the generalized cylinder model. pand, and rotate independent of other templates (shown in 2-D in Fig. 8). With this in mind, the strict generalized cylinder model is relaxed to account for cross-sectional expansion and shrinkage as one traces along a vessel. 3-D vessel tracing algorithm The tracing algorithm traces the vasculature in an exploratory manner starting from automatically detected starting points (seed points). To find the seed points, a 2-D maximum intensity projection of the 3-D volumetric image is created and a search for local maxima is performed along the lines of a coarse grid on this 2-D image. Each local maxima discovered becomes a seed candidate. For each seed candidate, the z value is found by performing an axial search corresponding to the lateral coordinates of each candidate. Notice that the search for seed point candidate has only been 1-D thus far. Next, each seed point candidate is validated in 3-D using the generalized cylinder model and unfit candidates are rejected. The validation process begins with finding the four boundaries as illustrated in Fig. 5, using the templates and template response function as in Eq. (1), exhaustively at all directions and at all widths. Seed point candidates without the four almost-parallel boundaries, i.e., with relatively low maximal template response, are therefore rejected. At this stage, each validated seed point is associated with vertical and horizontal width infor- mation as well as a direction unit vector corresponding to the morphometrics of the generalized cylinder model where it fits best. Tracing begins at the seed points in an iterative manner until one of the stopping criteria is met, i.e., where the generalized cylinder model is violated. A seed point is used as a starting point twice, along the opposing directions of the cylinder axis. A single iteration of tracing is defined as moving from a centerline point pi to the next centerline point pi⫹1 (p0 is a seed point) with the distance between these two points defined as the step size si, the angle between them as i, and the unit vector along the direction i as ui. At pi, the locations of the corresponding boundary points are denoted by the set Bi ⫽ {bTi, biB, bLi, biR} and are ordinarily computed as the points at which the template responses are maximum. To better handle tortuous structures of interest, this direct method was modified as follows. First, at each point in the template shifting process, the responses of templates along adjacent directions are also computed by rotating the templates. Second, template lengths are adjusted to fit local image features. This is because longer templates are more robust to noise than shorter ones since they perform more averaging along the vessel edge. On the other hand, shorter templates are more accurate around curved vasculature segments. This means that at each shifting and rotating step, the algorithm computes the responses of templates of different lengths (Al-Kofahi et al., 2002). A length-normalized tem- M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 121 Fig. 9. This figure illustrates the intuitions expressed mathematically in Eq. (4) that describes the maximum expansion (or shrinkage) tolerance angle of the vasculature when tracing from pi to pi⫹1 as a function of shift constraint ␣, rotation constraint , step size si, and template length k. plate response defined earlier in Eq. (1) is used to set the tracing direction. Mathematically, this can be written as follows: 共b i,u i,k i兲 ⫽ arg max 兵R共b,u,K兲其, (2) M {(b,u,k) 兩 b⫽pi⫹mu⬜,m⫽1, . . . , ,u僆U,k僆K} 2 where U is the set of unit vectors along directions in the neighborhood of ui and K is the set of all template lengths. The vector u is a unit vector along a particular , while u⬜ is the unit vector perpendicular to u. The parameter M is the user-defined diameter of the widest expected vessel. Values (bi, ui, ki) are the results of this exhaustive search at iteration i, each representing the location, orientation, and length, respectively, of the template that returns the maximum response R. This search is performed four times corresponding to the four templates that make up the generalized cylinder model. The centerline point pi, cylinder direction ui, and cylinder length ki are calculated as a function of the sets {bTi, biB, bLi, biR}, {uTi, uiB, uLi, uiR}, and {kTi, kiB, kLi, kiR}, respectively. Continuing with centerline tracing, we proceed from pi to pi⫹1 using the equation p i⫹1 ⫽ 共p i共B i兲 ⫹ s iu i兲 ⫹ c i⫹1 共B i⫹1兲, (3) where c i⫹1(B i⫹1) is the correction (refinement) vector as a function of the set of boundary points B i⫹1 (Al-Kofahi et al., 2002). In other words, the location of the centerline point at i ⫹ 1 is not exactly known until the algorithm determines the corresponding boundary points at i ⫹ 1. The step size s i acts as the scaling factor for the unit vector u i . It is adaptive and calculated from the length of the shortest template among the four corresponding templates with maximum responses. To allow small deviations from the generalized cylinder model, the template expansion is allowed to be fully independent, while template shift and rotation are subject to a set of constraints. For a template of length k, its shift is limited to an empirically set range ⫾ ␣ and its rotation is similarly limited to the range ⫾ . The following equation defines the maximum expansion (or shrinkage) tolerance angle when tracing from pi to pi⫹1 as a function of shift constraint ␣, rotation constraint , step size si, and template length k (see Fig. 9). 冉 冉 冊 冉 冊冊 ⫽ max tan⫺1 ␣  , sin⫺1 si k . (4) To prevent multiple traces of the same vessel, traced vessels are marked using ellipsoids that are individually sized according to the estimated horizontal and vertical widths at each traced point. Seed points located within traced vessels are deemed invalid and hence ignored. After tracing all structures in an image, individual traces are merged to form branching points. For a detailed description of these issues, the reader is referred to AlKofahi et al. (2002). Change analysis results Our primary intent was to quantitate temporal vessel changes in a set of time series images. There are two broad methods for change analysis. One method is to compute 122 M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 morphometric data from images at each temporal sampling point, and perform statistical comparison of these data. A more ambitious approach is to register the images over time and extract detailed changes on a vessel segment by vessel segment basis. In this work, the less ambitious approach was adopted as a starting point. Vessel lengths, widths, and count can be readily generated from the traces generated by the automatic tracing algorithm described in the previous section. Fig. 6 and Table 1 show the statistics corresponding to the four time series images shown in Figs. 1– 4. Naturally, only vasculature segments located in the volume common to all four images contributed to these statistics. An overall Change Index is calculated as the simple ratio of the current measurement over the previous measurement. For this study, 30 vasculature images, of varying planar dimensions and volume depth, were traced. Typical examples are presented here. SCID mice with window preparations were implanted with MCaIV at the center of the window and fluorescent labels were injected intravenously to highlight the vasculature. Imaging was performed on anesthetized mice using the MPLSM which consisted of a Spectra-Physics MilleniaX pumped Tsunami Ti:Sapphire laser, Bio-Rad MRC600 scanhead, and the Zeiss Axioskop20 microscope (Brown et al., 2001). Several images with a relatively higher degree of intensity nonuniformity were preprocessed using a median filter prior to tracing. These images reveal how a preexisting vascular bed is altered as a tumor grows into the imaged region. In addition to the trace projections, the program generates the actual traces drawn onto the 3-D input image and the segmented vasculature in a 3-D binary image (generated by drawing ellipsoids along the centerline, and sizing the ellipsoid according to the local vertical and horizontal width estimates). Results in Fig. 7 are examples of such binary images. The program also generates the length and width statistics in an output text file. Of the 30 images used in this study, 28 images belong to four sets of time series images (768 ⫻ 512 ⫻ 32 stacks, 8 bits/voxel, 7 days, 1 stack/day). The images were traced and statistics from each temporal set were gathered. A region of interest was defined for each image by the intersection of the image and all other images in the set. Statistics collected within these regions include total vasculature length, average vasculature segment length, average horizontal width, and average vertical width. Statistics and traces generated outside these regions were ignored. The generated statistics were entered into a spreadsheet and plotted to highlight the changes. Fig. 6 and Table 1 show the change analysis results for the first 4 images corresponding to the first 4 days from the second time series set that are shown in Figs. 1– 4. Actual change measurements such as percentage of reduction in total number of vasculature segments or percentage of increase in total length of vasculature can be obtained directly from the spreadsheet. Method for validating trace results Quantitative validation of the tracing algorithm requires the availability of a ground truth, which must be established manually. Manual tracing of 3-D structures is very difficult and time-consuming. It also suffers from a greater degree of tracer variability. To establish the ground truth, it is vital that one accounts for this intertracer variability by having multiple manual traces of the same image. This constitutes an unreasonable burden and we argue that, for the purposes of this study, it is sufficient to validate the results based on their 2-D maximum intensity projections. In other words, the 3-D automated trace results are projected on a 2-D plane, and validated against manual traces of 2-D maximum-intensity projections of the 3-D image. The tracing results were validated using two performance metrics. The two metrics cater to different user concerns regarding the accuracy of the automated traces. To further avoid the factor of subjectivity in validating the trace results, the comparisons between the manual and automated traces were performed automatically. This validation study is based on six image stacks consisting of the four image stacks shown in Figs. 1– 4 that belong to a time series set, and two other single-shot vasculature image stacks. The first performance metric is the average distance between the manually traced and automatically traced vasculature centerlines. To evaluate this metric, we calculated the Euclidean distance between every traced pixel in the manual traces and every traced pixel in the automated traces that were within a certain Euclidean distance, which was chosen to be 10. This metric is suitable for users concerned with the pixel-wise accuracy of the traces. The average distance over the six image stacks ranged from 1.81 to 2.38 pixels, with an overall average of 2.11 pixels. The second performance metric is the concordance, or agreement between the manual and automated traces. This metric was evaluated by calculating the number of automatically traced pixels that were within a certain distance from corresponding manually traced pixels. The end result tells how much editing may need to be performed manually by the user after the image has been traced automatically. Editing may be as easy as deleting false-positive segments, putting seed points to trace false-negative segments, or manually retracing the false-negative segments. Note that the factor of subjectivity of manual tracing is reduced down to the correspondence between the manual and automated traces. This metric is suitable for users that are concerned about coverage of the automated traces. The concordance measures for the six image stacks ranged from 72 to 89%, with an overall average of 81%. Other than the traces of the vasculature centerlines, the width statistics need to be validated as well. Given the nature of 3-D vasculature images, it is very time-consuming to manually gather the width statistics and, since the results will be subjective, it further reduces the value of the effort. Instead, we used several synthetic images containing tubes M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 123 Table 2 Width statistics results for phantom/synthetic imagesa,b,c Image Description Sine tube on x–y plane Sine tube on x–z plane Sine tube on x–y plane Sine tube on x–z plane Straight tube with parallel boundaries Straight tube with sinusoidal boundaries Horizontal width Vertical width Actual Detected 17.00 19.00 19.00 11.00 17.00 8.64 18.18 19.05 21.77 10.54 16.00 8.62 Error (%) 7.0 0.2 14.6 4.2 5.9 0.2 Average ⫽ 5.4 Actual Detected 5.00 19.00 19.00 11.00 17.00 8.64 6.28 23.73 18.73 12.48 18.95 8.62 Error (%) 25.7 24.9 1.4 13.5 11.5 0.2 Average ⫽ 12.9 a Phantom/synthetic images are generated by rolling an ellipsoid of specified horizontal and vertical width along a specified path (either a straight line or a sinusoid). b The actual width is the known width, and the detected width is the width measurement reported by the tracing algorithm. c Error is the percentage of absolute difference between the actual and detected measurements. of known geometrical dimensions. The tubes were generated by drawing ellipsoids of known horizontal and vertical widths along a known path, which can be straight or tortuous. We stimulated tortuous vasculature by using sinusoidal paths, and vasculature wall irregularities were simulated by using sinusoidal vasculature walls. The deviation of the reported width statistics from the known true values averaged 5.4% for horizontal width and 12.9% for vertical width (see Table 2 ). These measurements are greatly affected by the degree of angular discretization in the tracing algorithm, where finer angular discretization yields higher accuracy at the expense of computational speed. Also, the location of the true boundary was discretized to a voxel location, causing slight deviation from the “true” boundary location, which theoretically lies between the foreground voxel and the neighboring background voxel. Nevertheless, these deviations from the true measurements are reproducible and consistent from one image to another, which is critical for change analysis studies. Conclusions and discussion The change analysis study presented in this paper is based on geometrical statistics generated by a fully automatic 3-D tracing algorithm. The tracing algorithm is fast, accurate, and precise, making it applicable for large-scale applications where speed and reproducibility are important. It is robust to intensity nonuniformities, structural irregularities, and background noise. This work extends our previous work (Al-Kofahi et al., 2002) with more attention given to handling imaging artifacts and structural irregularities which are more apparent in angiogenic vasculature images than dye-injected neuron images (Al-Kofahi et al., 2002) and retinal angiograms (Can et al., 1999). Execution time is up to five times higher than our previous implementation, mainly due to the use of median statistics. Clearly, more intellectual and computational efforts are required in the vasculature tracing (segmentation) phase compared to the change analysis phase. This is because the whole change analysis phase is greatly simplified by just analyzing meaningful statistics and measurements that were extracted from raw image data by the tracing algorithm. It also implies that the accuracy of the change analysis relies heavily on the accuracy of the tracing algorithm. Regardless, the use of an automated tracing algorithm in quantitative change analysis yields reproducible results, which minimizes the factor of subjectivity in generating the measurements analyzed for changes. Consequently, collections of change analysis results can be fairly compared between research groups conducting all kinds of different vasculature-related assays. The tracing algorithm does not require any special hardware. The results presented here were obtained using a Dell Pentium III 1-GHz computer. For a typical 8-MB image as shown in Fig. 1, it takes 53 s, and its speed varies depending on the amount of structures present in the image since it only processes the image foreground. At the time of this writing, the authors are considering dissemination plans for the PC-compatible implementation of the algorithm, but they have not been finalized. Nevertheless, the tracing algorithm is still not perfect and suffers from some drawbacks, most noticeably false-negative errors (see Fig. 10). These errors are caused by (1) dim vasculature that has poor contrast with image background, (2) poorly defined edges which can be more appropriately modeled as ramp edges instead of the step edges which are built into our vasculature model, and (3) absence of a seed point on that particular vasculature segment. On the other hand, stretches of dye leakage that closely resemble discontinuity of dye in vasculature and spurious seeds contribute to false-positive errors. Overall, most tracing errors are caused by artifacts caused by the imaging process itself. For example, in the datasets considered for this study, the fluorescent dye is injected over a period of several days, leaving punctate deposits of extravasated dye in the subsequent images. Nevertheless, the erroneous results are reproducible, which means that they can be easily reproduced to 124 M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 further refine the tracing algorithm to be more robust to causes of these errors. Furthermore, unlike manual tracing, the tracing algorithm was shown to have the “steady hand” effect. In fact, Al-Kofahi et al. (2002) have shown that the algorithm is more accurate than a manual tracer in as far as locating the “true centerline” is concerned. The tracing algorithm can be viewed as having a consensual standard where expert human observers agree on a set of criteria to classify image voxels as part of the vasculature or the background. However, it is often not straightforward to formalize these criteria into mathematical forms that can be implemented as computer algorithms. Currently, only the notions of local parallel edges and local contrast are being incorporated in the tracing algorithm. In summary, the statistical change analysis methods described in this paper are just an example of possible utilizations of the tracing output. Other utilizations of geometrical information of curvilinear structures extracted by the tracing algorithm include geometrical change analysis, feature extraction, and image registration. Medical applications of change analysis include testing the efficacy of anti-angiogenic therapies and image-based diagnosis. A potential application in biology is to derive vessel growth parameters which may be correlated with physiological and gene expression profiles in tumors. Acknowledgments Various portions of this research were supported by the Center for Subsurface Sensing and Imaging Systems, under the Engineering Research Centers Program of the National Science Foundation (Award No. EEC-9986821), the NSF Partnerships in Education and Research Program, MicroBrightField Inc. (Williston, VT), the Ministry of Entrepreneur Development of Malaysia (via MARA), and by grants from the NCI (P01CA80124 and R24CA85140) and the Rensselaer Polytechnic Institute. The authors thank colleagues George Nagy, Charles V. Stewart, Richard Radke, Qiang Ji, Omar Al-Kofahi, Vijay Mahadevan, and Kenneth Fritzsche, for discussions and valuable suggestions on the broad topic of tracing algorithms. References Al-Kofahi, K.A., 2000. Algorithms for Rapid Automated Tracing of Neurons from 2-D and 3-D Confocal Images: Applications to Nanobiotechnology. Ph.D. thesis. Rensselaer Polytechnic Institute, Troy, NY. Al-Kofahi, K.A., Can, A., Lasek, S., Szarowski, D.H., Dowell, N., Shain, W., Turner, J.N., Roysam, B., 2003. Median based robust algorithms for tracing neurons from noisy confocal microscope images. IEEE Trans. Inform. Technol. Biomed., in press. Al-Kofahi, K.A., Lasek, S., Szarowski, D.H., Pace, C.J., Nagy, G., Turner, J.N., Roysam, B., 2002. Rapid automated three-dimensional tracing of neurons from confocal image stacks. IEEE Trans. Inform. Technol. Biomed. 6 (2), 171–187. Auerbach, R., Auerbach, W., Polakowski, I., 1991. Assays for angiogenesis: a review. Pharmac. Ther. 51, 1–11. Avinash, G.B., Quirk, W.S., Nuttall, A.L., 1993. Three-dimensional analysis of contrast-filled microvessel diameters. Microvasc. Res. 45 (2), 180 –192. Barbareschi, M., Gasparini, G., Morelli, L., Forti, S., Dalla Palma, P., 1995. Novel methods for the determination of the angiogenic activity of human tumors. Breast Cancer Res. Treat. 36 (2), 181–192. Brey, E.M., King, T.W., Johnston, C., McIntire, L.V., Reece, G.P., Patrick Jr., C.W., 2002. A technique for quantitative three-dimensional analysis of microvascular structure. Microvasc. Res. 63 (3), 279 –294. Brown, E.B., Campbell, R.B, Tsuzuki, Y., Xu, L., Carmeliet, P., Fukumura, D., Jain, R.K., 2001. In vivo measurements of gene expression, angiogenesis and physiological function in tumors using multiphoton laser scanning microscopy. Nature Med. 7 (3), 864 – 868. Can, A., Shen, H., Turner, J., Tanenbaum, H., Roysam, B., 1999. Rapid automated tracing and feature extraction from live high-resolution retinal fundus images using direct exploratory algorithms. IEEE Trans. Inform. Technol. Biomed. 3 (2), 125–138. Carmeliet, P., Jain, R.K., 2000. Angiogenesis in cancer and other diseases. Nature 407, 249 –257. DeFouw, D.O., Rizzo, V.J., Steinfeld, R., Feinberg, R.N., 1989. Mapping of the microcirculation in the chick chorioallantoic membrane during normal angiogenesis. Microvasc. Res. 38 (2), 136 –147. Dellas, A., Moch, H., Schultheiss, E., Feichter, G., Almendral, A.C., Gudat, F., Torhorst, J., 1997. Angiogenesis in cervical neoplasia: microvessel quantitation in precancerous lesions and invasive carcinomas with clinicopathological correlations. Gynecol. Oncol. 67 (1), 27–33. Dellian, M., Witwer, B.P., Salehi, H.A., Yuan, F., Jain, R.K., 1996. Quantitation and physiological characterization of angiogenic vessels in mice: effect of basic fibroblast growth factor, vascular endothelial growth factor/vascular permeability factor, and host microenvironment. Am. J. Pathol. 149 (1), 59 –71. Endrich, B., Reinhold, H.S., Gross, J.F., Intaglietta, M., 1979. Tissue perfusion inhomogeneity during early tumor growth in rats. J. Natl. Cancer. Inst. 62 (2), 387–395. Folkman, J., 2001. Angiogenesis, in: Braunwald, E., Fauci, A.S., Kasper, D.L., Hauser, S.L., Longo, D.L., Jameson, J.L. (Eds.), Harrison’s Principles of Internal Medicine, fifteenth ed., McGraw-Hill, New York, pp. 517–530. Frangi, A.F., Niessen, W.J., Hoogeveen, R.M., Walsum, T., Viergever, M.A., 1999. Model-based quantitation of 3-D magnetic resonance angiographic images. IEEE Trans. Med. Imag. 18 (10), 946 –956. Fox, S.B., Leek, R.D., Weekes, M.P., Whitehouse, R.M., Gatter, K.C., Harris, A.L., 1995. Quantitation and prognostic value of breast cancer angiogenesis: comparison of microvessel density, Chalkley count, and computer image analysis. J. Pathol. 177 (3), 275–283. Holmes III, D.R., Moore, M.J., Mantilla, C.B., Sieck, G.C., Robb, R.A., 2002. Rapid semi-automated segmentation and analysis of neuronal morphology and function from confocal image data. Proc. IEEE Intl. Symp. Biomed. Imag. 233–236. Huber, P.J., 1981. Robust Statistics. Wiley, New York. Iwahana, M., Nakayama, Y., Tanaka, N.G., Goryo, M., Okada, K., 1996. Quantification of tumour-induced angiogenesis by image analysis. Int. J. Exp. Pathol. 77 (3), 109 –114. Jain, R.K., Munn, L.L., Fukumura, D., 2002. Dissecting tumour pathophysiology using intravital microscopy. Natl. Rev. Cancer 2 (4), 266 – 276. Jain, R.K., Schlenger, K., Höckel, M., Yuan, F., 1997. Quantitative angiogenesis assays: progress and problems. Nature Med. 3 (11), 1203– 1208. Jakobsson, A.E., Norrby, K., Ericson, L.E., 1994. A morphometric method to evaluate angiogenesis kinetics in the rat mesentery. Int. J. Exp. Pathol. 75 (3), 219 –224. Kirchner, L.M., Schmidt, S.P., Gruber, B.S., 1996. Quantitation of angiogenesis in the chick chorioallantoic membrane model using fractal analysis. Microvasc. Res. 51 (1), 2–14. M.-A. Abdul-Karim et al. / Microvascular Research 66 (2003) 113–125 Kowalski, J., Kwan, H.H., Prionas, S.D., Allison, A.C., Fajardo, L.F., 1992. Characterization and applications of the disc angiogenesis system. Exp. Mol. Pathol. 56 (1), 1–19. Kumar-Singh, S., Vermeulen, P.B., Weyler, J., Segers, K., Weyn, B., Van Daele, A., Dirix, L.Y., Van Oosterom, A.T., Van Marck, E., 1997. Evaluation of tumour angiogenesis as a prognostic marker in malignant mesothelioma. J. Pathol. 182 (2), 211–216. Leunig, M., Yuan, F., Menger, M.D., Boucher, Y., Goetz, A.E., Messmer, K., Jain, R.K., 1992. Angiogenesis, microvascular architecture, microhemodynamics, and interstitial fluid pressure during early growth of human adenocarcinoma LS174T in SCID mice. Cancer Res. 52 (23), 6553– 6560. Li, V.W., Folkerth, R.D., Watanabe, H., Yu, C., Rupnick, M., Barnes, P., Scott, R.M., Black, P.M., Sallan, S.E., Folkman, J., 1994. Microvessel count and cerebrospinal fluid basic fibroblast growth factor in children with brain tumours. Lancet 344 (8915), 82– 86. Merchant, F.A., Aggarwal, S.J., Diller, K.R., Bovik, A.C., 1994. In-vivo analysis of angiogenesis and revascularization of transplanted pancreatic islets using confocal microscopy. J. Microsc. 176 (3), 262–275. Nissanov, J., Tuman, R.W., Gruver, L.M., Fortunato, J., 1995. Automatic vessel segmentation and quantification of the rat aortic ring assay of angiogenesis. Lab. Invest. 73 (2), 734 –739. Parsons-Wingerter, P., Lwai, B., Yang, M.C., Elliot, K.E., Milaninia, A., Redlitz, A., Clark, J.I., Sage, E.H., 1997. A novel assay of angiogenesis in the quail chorioallantoic membrane: stimulation by bFGF and inhibition by angiostatin according to fractal dimension and grip intersection. Microvasc. Res. 55 (3), 201–214. Rohr, S., Toti, F., Brisson, C., Albert, A., Freund, M., Meyer, C., Cazenave, J.P., 1992. Quantitative image analysis of angiogenesis in rats implanted with a fibrin gel chamber. Nouv. Rev. Fr. Hematol. 34, 287–294. Sato, Y., Nakajima, S., Shiraga, N., Atsumi, H., Yoshida, S., Koller, T., Gerig, G., Kikinis, R., 1998. Three-dimensional multi-scale line filter 125 for segmentation and visualization of curvilinear structures in medical images. Med. Image Anal. 2 (2), 143–168. Schoell, W.M.J., Pieber, D., Reich, O., Lahousen, M., Jenicek, M., Guecer, F., Winter, R., 1997. Tumor angiogenesis as a prognostic factor in ovarian carcinoma: quantification of endothelial immunoreactivity by image analysis. Cancer 80 (12), 2257–2262. Seifert, W.F., Verhofstad, A.A.J., Wobbes, T., Lange, W., Rijken, P.F.J.W., van der Kogel, A.J., Hendriks, T., 1997. Quantitation of angiogenesis in healing anastomoses of the rat colon. Exp. Mol. Pathol. 64 (1), 31– 40. Shen, H., Roysam, B., Stewart, C.V., Turner, J.N., Tanenbaum, H.L., 2001. Optimal scheduling of tracing computations for real-time vascular landmark extraction from retinal fundus images. IEEE Trans. Inform. Technol. Biomed. 5 (1), 77–91. Streekstra, G.J., Pelt, J.V., 2002. Analysis of tubular structures in threedimensional confocal images. Network Comput. Neural Syst. 13, 381– 395. Sun, Y., Lucariello, R., Chiaramida, S., 1995. Directional low-pass filtering for improved accuracy and reproducibility of stenosis quantification in coronary arteriograms. IEEE Trans. Med. Imag. 14 (2), 242–248. Tjalma, W., Van Marck, E., Weyler, J., Dirix, L., Van Daele, A., Goovaerts, G., Albertyn, G., van Dam, P., 1998. Quantification and prognostic relevance of angiogenic parameters in invasive cervical cancer. Br. J. Cancer 78 (2), 170 –174. Toi, M., Kondo, S., Suzuki, H., Yamamoto, Y., Inada, K., Imazawa, T., Taniguchi, T., Tominaga, T., 1996. Quantitative analysis of vascular endothelial growth factor in primary breast cancer. Cancer 77 (6), 1101–1106. Wild, R., Ramakrishnan, S., Sedgewick, J., Griffioen, A.W., 2000. Quantitative assessment of angiogenesis and tumor vessel architecture by computer-assisted digital image analysis: effects of VEGF-toxin conjugate on tumor microvessel density. Microvasc. Res. 59 (3), 368 –376.