Chapter 4 Classical Stereological Measures This chapter reviews the most commonly used stereological measurements. More detailed discussions and derivations may be found in the collection of traditional texts in the field (Saltykov, 1958; DeHoff and Rhines, 1968; Underwood, 1970; Weibel, 1978; Kurzydlowski and Ralph, 1995; Howard and Reed, 1998). Each section of the chapter focuses on a manual stereological measurement. In each case, the procedure is illustrated with a microstructure and a superimposed grid. Each figure lists: 1. The probe population that is required in the measurement; 2. The specific probe that is used in the illustration; 3. The event that results when this probe interacts with the microstructure; 4. The measurement to be made and its specific result; 5. The stereological relationship that connects the measurement to the geometric property of the structure; 6. The calculation of the appropriately normalized version of the measured result; 7. The calculation of the geometric property of the microstructure that is estimated from this measurement. A discussion of the corresponding procedure used in computer-based image analysis follows the descriptions of the manual measurements. The chapter begins with the measurement of the area of features in a two dimensional structure since this is easily visualized. Measurement of volume fraction, surface area density and line length are then reviewed. Two-Dimensional Structures; Area Fraction from the Point Count Figure 4.1 shows a two dimensional structure consisting of two feature sets. Label the white background a and the colored regions b. The fraction of the area of the structure occupied by the b phase is AAb. This area fraction is measured using point probes. The population of points in this two dimensional world is sampled by superimposing a grid of points on the structure. An outline of the measurement process is given in the caption of Figure 4.1. The square 5 ¥ 5 grid in Figure 4.1 constitutes a sample of 25 points (the intersections in the grid) out of this population of all of the points that could be identified in the area of the specimen. The event of interest that results from the 45 46 Chapter 4 Figure 4.1. Measurement of the area fraction of the dark phase AA. (For color representation see the attached CD-ROM.) Probe population: Points in two dimensional space This sample: 25 points in the grid Event: Points lie in the phase Measurement: Count the points in the phase This count: 8 points in the phase Relationship: ·PPÒ = AA Normalized count: PP = 8/25 = 0.32 Geometric property: AA = 0.32 interaction of the probe sample with the structure is, “the point hits the b phase”. The actual measurement is simply a count of these events, i.e., the number of points in the 5 ¥ 5 grid in Figure 4.1 that lie within the b areas. In Figure 4.1, this count gives 8 points in the b phase; that count is the result for this placement of the probe sample in this field of the microstructure. This point count is related to the area fraction of the b phase in this two dimensional structure by the fundamental stereological relationship, PP b = AAb (4.1) The left side of this equation, ·PPÒ, is the “expected value” of the fraction of points that hit the b phase in a sample set of grid placements; the right hand side is the area fraction of b in the microstructure. Classical Stereological Measures 47 In the example of Figure 4.1, the point fraction is 8/25 = 0.32. In practice the grid will be placed on a number of fields, each producing a particular point count for beta hits. This distribution of number of hits has a mean value P , and a corresponding sample mean value for the point fraction, P P = P /PT, where PT is the total number of points in the grid, 25 in this example. Following usual statistical practice, this sample mean value, P P, is use to estimate the expected value of the point fraction ·PPÒ in the b phase for the population of points in the structure, and, through equation (3.1), the area fraction of the b phase. The area fraction of the a phase in this structure may be evaluated by counting hits in a, or simply by subtracting AAb from 1. Figure 4.2 shows a three phase structure composed of a, b and g areas. In this example, a 5 ¥ 5 grid is placed on the structure. The procedure described in the preceding section may be applied separately to hits in a, hits in b and hits in the g phase. This placement of the grid gives 10 hits on the a phase, 8 on the b particles and 7 hits on g. The corresponding point fractions of the three phases are obtained by dividing by PT = 25. Estimates for the three area fractions for this single Figure 4.2. Measurement of the area fractions in a three phase structure AAa, AAb, AAg. (For color representation see the attached CD-ROM.) Probe population: Points in two dimensional space This sample: 25 points in the grid Event: Points lie in each phase, a, b, g Measurement: Count the points in each phase This count: 10 points in a, 8 points in b, 7 points in g Relationship: ·PPÒ = AA Normalized count: PPa = 10/25 = 0.40; PPb = 8/20.32; PPg = 7/25 = 0.28 Geometric property: AAa = 0.40; AAb = 0.32; AAg = 0.28 48 Chapter 4 placement of the grid are AAa = 0.40; AAb = 0.32; AAg = 0.28. Statistically valid results will require replication of these counts on a number of appropriately chosen fields. Area fractions of constituents can be measured directly in computer assisted image analysis. This is a straightforward application of the point count applied with a very high density of points. In this case, each pixel in the image is a point probe. The phase of interest in the count is segmented from the rest of the image on the basis of its grey shade range or its color, perhaps along with some other geometric characteristics of the phase, such as size or shape. Numerical gray shades or color values for those pixels that satisfy the conditions set up to identify the phase are set to black, and the remaining pixels are set to white. The “detected image” is thus a “binary image”. The computer then simply counts black pixels in this binary image. The point fraction is the ratio of this count to the total number of pixels in the image. These counts can be made essentially in real time. It might seem that the pixel count in an image analysis system would provide a much greater precision in the estimate of AA than a manual count of a comparatively very limited number of points. However, analysis has shown that the manual count may give about the same precision, as reflected in the standard deviation of the counts for a collection of fields. This is partly because much of the variation in AA derives from differences from one field to another. Also, there is a great deal of redundant information in the pixel count, since many pixels lie in any given individual feature. The point count method is most efficient when the grid spacing is such that adjacent points rarely fall within the same feature, cell or region in the image (they are then said to be independent samples of the structure). An advantage of this method is that when the hits produced by the point grid are all independent, the number of hits can be used directly to estimate the measurement precision, as discussed in Chapter 8. The most difficult step in image analysis is detection or segmentation, i.e., making the computer “see” the phase to be analyzed as the human operator sees it. There are always pixels included in the discriminated phase that the operator can “see” are not part of the phase, and pixels not included that “should be”. Detection difficulties increase with the complexity of the microstructure. Such problems frequently can be minimized with more careful sample preparation and additional image processing and more steps in the analysis. In any experimental situation it is necessary to balance the effort required to develop an acceptable sample preparation strategy and discrimination algorithm against the inconvenience of the manual measurement which incorporates the most sophisticated of discrimination systems, the experienced human operator with independent knowledge about the sample and its preparation. In these procedures for estimating area fraction of a phase it is not necessary to know the actual dimensions of the grid, since the measurement involves only ratios of counts of points. The scale of the grid relative to the microstructure, and the number of points it contains, do not influence the expected value relationship, equation (4.1). These choices do influence the precision of the estimate obtained, through the spread in the number of counts made from one grid placement to the Classical Stereological Measures 49 next, but not the estimated expected value. For example, as a limiting case, if the grid were very small in comparison with the features in the structure then most of the time the entire grid would either lie within a given phase, or be outside of it. For a 5 ¥ 5 grid, counts for most placements of the grid would give a count of 25 or zero. The mean number of counts would still be the area fraction for a large number of observations, but the standard deviation of these counts would be comparatively large. A correspondingly large number of fields would have to be viewed to obtain a useful confidence interval on the estimate of the mean. The point count can also be used to estimate the total area of a single feature in a two dimensional structure. The solution to this problem is an example of a sampling strategy that is pervasive in the design of stereological experiments called the “systematic random sample”. Figure 4.3 shows a single particle of the b phase. Figure 4.3. Area of a feature in two dimensions. (For color representation see the attached CD-ROM.) Probe population: Points in two dimensional space This sample: 6 ¥ 8 = 48 points in the grid Calibration l0 = 5.7 mm Event: Points lie in the feature Measurement: Count the points in the feature This count: 20 points in the feature Relationship: A = l02·P Ò Geometric property: AA = (5.7 mm) · 20 = 650 mm2 50 Chapter 4 Design a grid of points with outside dimensions large enough to include the whole particle. In order to estimate the area of the feature with a point count it is necessary to calibrate the dimensions of the grid at the actual magnification of the image. A stage micrometer may be used to measure the overall dimensions of the grid and compute the spacing between grid points. Let the spacing thus measured be l0, and the total number of points in the grid be PT. For the 6 ¥ 8 point grid in Figure 4.3, there are 48 grid points. Calibration of the magnification gives the grid spacing as l0 = 5.7 mm. The population of points that the grid samples is contained in the area AT given by PT · l02 (= 48 · (5.7)2 = 1560 mm2 in Figure 4.3). Focus on the square box in the upper left corner of the area in Figure 4.3. A specific placement of the grid may be specified by locating the upper left grid point Q at any point (x, y) within this box. A given choice of (x, y) specifies the location of the remaining 48 probe points in the grid. Imagine that the point (x, y) is moved to survey all of the points in the corner box. Then the remaining grid points systematically sample all of the points in the area AT, i.e., the population of points of interest. The number of points P that lie within the feature is counted for a given placement of the grid (20 points hit the particle in Figure 4.3). The point fraction PP for that placement of the grid may be computed as P/PT. If the point Q at (x, y) that locates the grid is chosen uniformly from the points in the corner box then the point fraction for this sample value is an unbiased estimate of the expected value for the population ·PPÒ and equation (4.1) applies. Statistically speaking, the point fraction in this experiment is an unbiased estimate of the fraction of the area AT occupied by the b particle. The expected value relation is PP = AA = A A = AT PT l02 (4.2) The number of points PT in the grid in this experiment is fixed. The number of points that hit the area of the figure, P, varies from trial to trial. Thus PP = P PT (4.3) where ·PÒ is the expected value of the number of points observed within the feature on each trial. Insert this result into equation (4.2) and solve for A A = (PT l02 ) PP = (PT l02 ) P = l02 P PT (4.4) This simple result is the statistical equivalent of tracing the figure on graph paper and counting the number of squares that lie in the figure. The area of the b feature is seen to be the area associated with a given point (l02) times the number of points that hit the figure. This is approximately true for any single placement of the grid. The argument that leads to this perhaps obvious result demonstrates that the area is exactly equal to the area of a grid square times the expected value of the number of hits on the feature. This transforms the simple geometric approximation into a statistical result, with the attendant potential for replication of the Classical Stereological Measures 51 experiment, evaluation of standard deviation and estimate of the precision of the result. For the b feature in Figure 4.3, the 20 hits for the placement of the grid shown estimates the area of the feature to be 650 mm2. Volume Fraction from the Point Count The most commonly measured property of three dimensional feature sets is their volume, usually reported as the volume fraction, VV, of the phase. This property may be estimated using either plane, line or point probes; the simplest and most commonly used measurement relies on point probes. The population of points to be sampled by these probes is the set of points contained within the volume of the specimen in three dimensional space. Point probes are normally generated by first sectioning the sample with a plane, and generating a grid of points on the plane section. As in the two dimensional structure described in the last section, the event of interest is “point hits the b phase” where “b phase” is taken to mean the set of features at the focus of the analysis. These points are simply counted. The stereological relation that connects this point count with the volume of the structure is PP b = VVb (4.5) Visualize a specimen composed of two phases a and b. This structure is revealed by sectioning it with planes and examining fields on these sections, as shown in Figure 4.4. The population of points in the three dimensional specimen is probed by the 5 ¥ 5 grid of points superimposed on this structure. For this sample, 5 points lie within the b phase. This count is replicated on a series of fields on the set of sectioning planes. The mean P and standard deviation sP of these counts are computed. The mean point fraction, P /PT, is used to estimate the expected value for the population of points, ·PPÒ, and, through equation (4.5), the volume fraction occupied by the b phase. For the field shown in Figure 4.4, this estimate is 5/25 = 0.25. In order to obtain an estimate of the volume fraction of b in the three dimensional structure represented by the plane section in Figure 4.4 it is necessary to repeat this measurement on a number of fields that are chosen to represent the population of points in three dimensions and average the result. Table 4.1 provides an example of results that might be derived from a set of observations of 20 such fields. The mean number of counts on these fields is 7.60 and the standard deviation of the set of 20 observations is found to be 1.5. The standard deviation of the population mean in this experiment is given sP by s P = where n = 20, the number of readings in this sample. P ± 2s P is the n 95% confidence interval associated with this set of readings. The result in Table 4.1 may be interpreted to mean that the probability is 0.95 that the expected value of P for the population of points lies within the interval 7.60 ± 0.68. Since each field was sampled with 25 points, the normalized point count and its confidence interval is obtained by dividing both numbers by 25. There is thus an 0.95 probability that the expected value of the point fraction lies in the interval 0.277 to 0.231. 52 Chapter 4 Figure 4.4. Measurement of the volume fraction VV of a phase in three dimensions. (For color representation see the attached CD-ROM.) Probe population: Points in three dimensional space This sample: 25 points in the grid Event: Points lie in the phase Measurement: Count the points in the phase This count: 5 points in the phase Relationship: ·PPÒ = VV Normalized count: PP = 5/25 = 0.25 Geometric property: VV = 0.25 Since the expected value of the point fraction estimates the volume fraction by equation (4.5), this range is also the confidence interval for the estimate of the volume fraction. Figure 4.5 shows a structure consisting of three phases, a, b and g. In this structure the small g particles lie within the b phase; there are no g particles in the a matrix. This is a very common structure in life science applications where an organelle (g ) is a part of the b cells. The a phase is not of interest in the Table 4.1. Point counts from the structure in Figure 4.4 P 7.60 sP 1.51 sP 0.34 P ± 2sP 7.60 ± 0.68 PP ± 2s PP 0.304 ± 0.027 VV ± 2sVV 0.304 ± 0.027 Classical Stereological Measures 53 Figure 4.5. Small particles of the g phase (black) are located within b features (gray) in this three phase structure. There are no g particles in the a phase (white). A 70 point grid is used to estimate the fraction of the volume of b occupied by g by separately estimating the volume fractions of the b and g phases in the structure. (For color representation see the attached CD-ROM.) Probe population: Points in three dimensional space This sample: 70 points in the grid Event: Points lie in a, b or g phase Measurement: Count the points in each phase This count: 30 points in b, 8 points in g Relationship: ·PPÒ = VV Normalized count: PPb = 30/70 = 0.43; PPg = 8/70 = 0.11 Geometric property: VVb = 0.43; VVg = 0.11; VVb,g = 0.11/(0.43 + 0.11) = 0.20 experiment; the fraction of the volume of the b cells that are occupied by g organelles is the object of this example. There are 70 points in the grid. For the field shown, 30 points hit the b phase and 8 hit g; the remaining 32 points are in a. These counts are replicated on a number of fields. The mean number of hits in each phase, P b and P a, are computed, along with their standard deviations. Corresponding point fractions are obtained by dividing by PT = 70 for this case. The resulting point fractions are used to estimate their corresponding expected values and hence the volume fractions, VV b, VVg, and, by difference, VVa by applying equation (4.5). The volume fraction of the structure occupied by the b cells including the g organelles contained within them is the sum VVb + VVg. To find the fraction of that volume occupied by g organelles, take the ratio, VVg/(VV b + VVg ). For the single field shown the estimates are: VVg = 8/70 = 0.11; VV b = 30/70 = 0.43 and VVa 0.46. The fraction of the volume of the b regions occupied by g is 0.11/(0.43 + 0.11) = 0.20. In this analysis it is important to obtain valid estimates of the volume fractions of the three phases separately, and then combine the results to obtain the desired comparison. As an alternate (incorrect) procedure, imagine taking counts of b and g for each field, Pb and Pg. Add the b and g counts, (Pb + Pg). Next, take 54 Chapter 4 the ratio [Pg /(Pb + Pg)]. Average this ratio over a number of fields to estimate the fraction of b cells occupied by g organelles. This procedure does not give the same result as that described in the previous paragraph because the average of a sum of ratios is not the same as the ratio of the averages. To obtain valid estimates it is important to measure the volume fractions of the various phases with respect to the structure as a whole, and subsequently manipulate these results to provide measures of relative volumes among the phases. Figure 4.6 illustrates the use of the point count to estimate the total volume of a single feature in a three dimensional microstructure. A three dimensional array of points is used to sample the feature. Visualize a box large enough to contain the feature. N equally spaced planes are sliced through the feature; let h be the distance between planes. N = 5 and h = 11.5 mm in Figure 4.6. Each plane has an (m ¥ n) grid of points with a calibrated grid spacing l0 imposed on it. Grids with (5 ¥ 5) points with spacing of 5.2 mm are shown in Figure 4.6). In this way a three dimensional grid of points with (N ¥ m ¥ n) points {(5 ¥ 5 ¥ 5) = 125 in Figure 4.6} is constructed which includes the entire feature. The box associated with each grid point has dimensions (v0 = l0 · l0 · h), {(5.2 · 5.2 · 11.5 = 311 mm3 in the figure}. Counts are made of the number of points that hit the feature on each of the five sectioning planes. The sum of these five counts is the total number of points in the three dimensional grid that hit the particle. In Figure 4.6 the total number of hits is (1 + 3 + 2 + 3 + 1 = 10 hits). Visualize a small box with the dimensions v0 = l0 · l0 · h at the upper left rear corner of the large box that contains the feature. The upper, left, back corner of the Figure 4.6. The Cavalieri principle is used to estimate the volume of a single feature by counting points on a series of sectioning planes. The spacing between the planes is h. (For color representation see the attached CD-ROM.) Probe population: Points in three dimensional space This sample: 5 ¥ 5 ¥ 5 = 125 points in the grid Calibration: h = 11.5 mm, l0 = 5.2 mm n0 = l0 · l0 · h = 5.2 · 5.2 · 11.5 = 311 mm3 Event: Points lie in the feature Measurement: Count the points in the feature This count: 1 + 3 + 2 + 3 + 1 = 10 points in the feature Relationship: Vb = PTv0·PPÒb = v0·P Òb Geometric property: Vb = 311 mm 3· 10 = 3110 mm3 Classical Stereological Measures 55 three dimensional grid will be located at some point Q = (x, y, z) within this small box. For any given choice of Q, the positions of the remaining (N ¥ m ¥ n) points in the grid is determined. They provide a systematic sample of the population points within the containing box. As Q is moved to all of the points in the small corner box, the grid of points samples all of the population of points within the containing box. Thus, a random choice for the position of Q from its uniform distribution of possible points in the small box produces a systematic random sample of the population of points in the containing box. The fraction of points in the three dimensional grid thus provides an unbiased estimate of the expected value for the population of points in the containing box and equation (4.5) may be applied: PP b = VV b = Vb Vb = VT PT v0 (4.6) The number of points that hit the feature varies from trial to trial. Since the total number of points PT is the same in all placements of the three dimensional grid, the expected value of this point count estimates the point fraction PP b = P PT b (4.7) Insert this result into equation (4.6) and solve for Vb: b V b = PT v0 PP b = PT v0 P = v0 P PT b (4.8) Thus, the volume of the b feature is the volume associated with an individual point (l0 · l0 · h) times the expected value of the number of hits that points in the grid make with the feature. For a single position of the grid, v0Pb provides an estimate of the volume of the feature. For the example illustrated in Figure 4.6, a total of 10 hits are noted. The volume associated with a grid point was computed earlier to be 311 mm3. Thus the volume of the feature is estimated at (10)(311) = 3110 mm3. The structured random sample is a much more efficient procedure than random sampling in which grid points are independently placed in the volume. In the latter case, points will inevitably cluster in some areas producing oversampling, while being sparse in others producing undersampling. It will take on the average nearly three times as many independent random points to reach the same level of precision as with the use of the structured approach, but the same answer will still result, namely that the point fraction that hit the phase or structure measures the volume. The point counting procedure to estimate the volume of a three dimensional object is an example of the oldest of the stereological procedures based upon the “Cavelieri Principle” (Howard & Reed, 1998a). The object to be quantified is sliced into a collection of slabs of known thickness. Some method is used to measure the area of each slab, such as the point count described in an earlier section. An alternate procedure might use a planimeter, an area measuring mechanical instrument 56 Chapter 4 used by cartographers before the advent of the computer, to measure the individual cross section areas. If the slabs are of uniform thickness, and uniform content, weighing each slab, together with a calibration of the weight per unit area, could be used. The underlying principle takes the volume of each slab to be its cross sectional area times its thickness, so that the volume of the object is the sum of the volumes of the slabs N V = Â Ai h = A ( Nh) (4.9) i =1 where ·AÒ is the average cross section area of the slabs and (Nh) is the total height of the object. Two-Dimensional Structures; Feature Perimeter from the Line Intercept Count Each of the collection of features in the two dimensional structure shown in Figure 4.7 has a boundary, and each boundary has a length commonly referred to as its perimeter. The normalized parameter, LA, is the ratio of the total length of boundaries of all of the features in the specimen divided by the area that the specimen it occupies. This perimeter length per unit area can be estimated by probing the structure with a set of line probes represented in Figure 4.7 by the four horizontal lines in the superimposed grid. The event of interest in the measurement is “line intersects boundary”. A simple count of these events forms the basis for estimating LA through the fundamental stereological formula (Underwood, 1970a): PL = 2 LA p (4.10) PL in this equation is called the “line intercept count”; it is the ratio of the number of intersections counted to the total length of line probe sampled (only the horizontal lines in the grid were used in this example). Each of the lines in Figure 4.7 is found by calibration to be 17.7 micrometers (mm) long. The total length of line probe sampled in this placement of the grid is 4 ¥ 17.7 = 70.8 mm. Since there are 17 intersections marked and noted in Figure 4.7, for this example PL is 17/70.8 = 0.24 counts/mm probed. The left hand side of equation (4.10) is the expected value of this measurement for the population of lines that can be constructed in two dimensional space. Inverting equation (4.10) gives the estimate of the perimeter length per unit area for the features shown in Figure 4.7: LA = p p Ê mm ˆ PL = 0.24 = 0.38Á ˜ Ë mm2 ¯ 2 2 (4.11) The area within the grid shown in Figure 4.7 is (17.7)2 = 313 mm2. A rough estimate of the boundary length of the features contained in the grid area is thus 0.38 (mm/mm2) · 313 mm2 = 119 mm. There are 8.5 particles in the area of the grid. (Particles that lie across the boundary are counted as 1/2.) A rough estimate of the Classical Stereological Measures 57 Figure 4.7. Use of the line intercept count to estimate the total length of the boundary lines of a two dimensional feature set. (For color representation see the attached CDROM.) Probe population: Lines in two dimensional space This sample: Four horizontal lines in the grid Calibration: L0 = 17.7 mm, total probe length = 4 · 17.7 = 70.8 mm Event: Line intersects the feature boundary line Measurement: Count the intersections This count: 17 intersections Relationship: ·PL Ò = (p/2) · LA Normalized count: PL = 17 counts/70.8 mm = 0.24 counts/mm Geometric property: LA = (p/2)PL = (p/2) · 0.24 = 0.38 (mm/mm) average perimeter of particles may be estimated by dividing the total boundary length in the area by the number of particles: 119 mm/8.5 = 14 mm. Inspection of the features in Figure 4.7 indicates that this result is plausible. Each member of the population of lines in two dimensional space has two attributes: position, and orientation. The set of lines in the grid used for a probe in Figure 4.7 represent a few different positions in the population of lines, but only a single orientation. In order for a sample of line probes to provide an unbiased estimate of a value for the population of probes it is necessary that the probe lines uniformly sample all orientations of the circle. A representative sample of the population of orientations of lines could be obtained by rotating the stage by 58 Chapter 4 uniform increments between measurements. A much more direct strategy, which guarantees uniform sampling of line orientations, uses test line probes in the shape of circles to make the PL measurement. Figure 4.8 shows a microstructure with phase boundaries that tend to be aligned in the horizontal direction. In Figure 4.8a and 4.8b a square grid of points is superimposed on this structure. Calibration shows that the grid is square 23.7 mm on a side. The set of five horizontal lines are used to sample the population of lines in two dimensional space in Figure 4.8a. The total length of lines probed in the a Figure 4.8a. Use of horizontal line probes to measure the total projected length of boundaries of two dimensional features on the vertical axis LA,proj(vert). (For color representation see the attached CD-ROM.) Probe population: Horizontal lines in two dimensional space This sample: Five horizontal lines in the grid Calibration: L0 = 23.7 mm, total probe length = 5 · 23.7 = 118.5 mm Event: Line intersects the feature boundary line Measurement: Count the intersections This count: 8 intersections Relationship: ·PLÒ(horz) = LA,proj(vert) Normalized count: PL = 8 counts/118.5 mm = 0.068 counts/mm Geometric property: LA,proj(vert) = PL(horz) = 0.068 (mm/mm2) Classical Stereological Measures 59 b Figure 4.8b. Use of vertical line probes to measure the total projected length of boundaries of two dimensional features on the horizontal axis LA,proj(horx). (For color representation see the attached CD-ROM.) Probe population: Vertical lines in two dimensional space This sample: Five vertical lines in the grid Calibration: L0 = 23.7 mm, total probe length = 5 · 23.7 = 118.5 mm Event: Line intersects the feature boundary line Measurement: Count the intersections This count: 62 intersections Relationship: ·PLÒ(horz) = LA,proj(vert) Normalized count: PL = 62 counts/118.5 mm = 0.52 counts/mm Geometric property: LA,proj(vert) = PL(horz) = 0.52 (mm/mm2) horizontal direction is (5 · 23.7 = 118.5 mm). The events, line intersects boundary, are marked with red markers in Figure 4.8a. There are 8 intersections, giving a line intercept count in the horizontal direction of {8/118.5 = 0.068 (counts/mm)}. If the set of vertical lines in the grid are used as a line probe (Figure 4.8b), the resulting count is 62; for this direction, PL = 62/118.5 = 0.52 counts/mm. It is clear that the line intercept count is different in different directions in this aligned structure, and that this reflects the anisotropy of the structure. 60 Chapter 4 c Figure 4.8c. Use of circular line probes to measure the true total length of boundaries in two dimensional features, LA. (For color representation see the attached CD-ROM.) Probe population: Lines in two dimensional space This sample: Set of lines equivalent to 8 circles Calibration: d = 23.7/4 = 5.9 mm, total probe length = 8 · p · d = 148.9 mm Event: Line intersects the feature boundary line Measurement: Count the intersections This count: 55 intersections Relationship: ·PLÒ = (p/2)LA Normalized count: PL = 55 counts/148.9 mm = 0.37 counts/mm Geometric property: LA = (p/2)PL = 0.58 (mm/mm2) Indeed, counts in different directions may be used to characterize the anisotropy quantitatively. Replications on a collection of fields produce mean and standard deviations of the counts made in each direction. A polar plot of PL as a function of q, the angle the line direction makes with the horizontal direction, as shown in Figure 4.9, called the “rose of the number of intersections” [Saltykov, 1958; Underwood, 1970b] provides a graphical representation of this anisotropy in a two dimensional structure. The true length of boundaries in this anisotropic structure may be estimated by applying test lines that are made up of circular arcs, as shown in Figure 4.8c. Classical Stereological Measures 61 Figure 4.9. Rose-of-the-number-of-intersections for a two-dimensional anisotropic structure. The population of orientations in this set of circular line probes provides a uniform sample of the population of line orientations in two dimensional space. This line grid is equivalent to eight circles, each with a calibrated diameter of 23.7/4 = 5.92 mm. The total length of this collection of line probes is thus 8 [p · (5.92 mm)] = 148.9 mm. The number of intersections with these line probes marked in Figure 4.8c is 55, giving a line intercept count PL = 55/148.9 = 0.37 (counts/mm). Note that this result is within the range between the result for horizontal lines (0.068) and the vertical lines (0.52). Equation (4.10) gives the corresponding estimate for the true perimeter length: LA = (p/2) 0.37 = 0.58 mm/mm2. The number of features per unit area in Figure 4.8 is counted to be 21/(23.7)2 = 0.037 (1/mm2). A rough estimate of the average perimeter of features in this structure is LA/NA = 0.58/0.037 = 16 mm. Inspection of Figure 4.8 shows that this is not an unreasonable estimate of the average particle perimeter. Most image analysis programs provide a measurement of the perimeter of individual features directly on the digitized image. Boundary pixels are identified in the detected binary image. The boundary line is approximated as a broken line connecting corresponding points in adjacent boundary pixels. The perimeter of the feature is then computed as the sum of the lengths of the line segments that make up this broken line. The attendant approximation may produce significant errors. The magnitude of these errors depends upon how many different directions are used to construct the broken line boundary. Consider the circle shown in Figure 4.10. Its digitized image is represented by the blue plus red pixels. Identification of which pixels lie on the boundary may vary with the choice of the number of neighbors that are required to be “non-b”. The broken line used to compute the perimeter is illustrated for two different cases. In Figure 4.10a only four directions are used in constructing the perimeter. It is easy 62 Chapter 4 Figure 4.10. Illustration of the bias in measuring perimeter of a circle in a digitized image: a) four directions gives the perimeter of a square; b) eight directions gives that of an octagon. (For color representation see the attached CD-ROM.) to see that the line segments connecting neighboring pixels can be combined to make the square that encloses the circle. If the diameter of the circle is d, then its actual perimeter is (p d ). The reassembled broken lines make a square with edge length d; its perimeter is (4 d ). This overestimates the perimeter of the circle by a factor of 4/p = 1.27. That is, use of this algorithm gives a built in bias for every particle of about +27%. In Figure 4.10b the segments assemble into an octagon; the ratio of perimeter of the broken line to the circle is computed to be 1.05. Thus, even for this more sophisticated algorithm, evaluation of the perimeter of simple particles has a bias of about +5%. Constructing an n-sided polygon with sides running in 16 possible directions (a 32-gon) further improves the accuracy of the perimeter. For example, measuring circles varying from 5 to 150 pixels in diameter using this technique gives results that are within 1% as shown in Table 4.2. However, the bias is always positive (the measured value is longer than the true value). Another new method for perimeter measurement is based on smoothing the feature outline and interpolating a superresolution polygon using real number values (fractional pixels) for the vertices (Neal et al., 1998). This method has even better accuracy and less bias, but is not yet widely implemented. The presence of any bias in measurements is anathema to stereologists and so they generally prefer to use the count of intercepts made by a line grid with the outlines. A simple experiment will permit the assessment of a given software in its measurement of perimeter. Generate, by hand or computer graphics, a collection of black circles or different sizes. Acquire the image in the computer. Detect the collection of circles. Instruct the software to measure the area A and perimeter P of each feature and compute some measure of the “circularity” of the features, such Classical Stereological Measures 63 Table 4.2. Perimeter measurements using a 32-sided polygon Diameter (pixels) 5 10 15 20 30 40 50 75 100 125 150 Actual Perimeter 15.7078 31.4159 47.1239 62.8319 95.2478 125.6637 157.0796 235.6194 314.1593 392.6991 471.2389 Measured Perimeter 15.6568 31.8885 47.2022 63.4338 95.4562 126.6030 158.1100 236.6480 314.6880 393.1660 471.8700 Error (%) -0.326 +1.504 +0.166 +0.958 +0.219 +0.747 +0.656 +0.437 +0.168 +0.119 +0.134 as the formfactor 4pA/P2. This parameter has a value of 1 for a perfect circle. Measured values may be consistently smaller than 1.00, and may vary with size of the circle in pixels. Use of the manual line intercept count to estimate perimeter does not suffer from this bias. Three-Dimensional Structures: Surface Area and the Line Intercept Count Figure 4.11 shows a specimen that contains internal interfaces. Usually interfaces of interest in microstructures are part of the boundaries of particles, e.g., the ab interface in a two phase structure, but this is not a requirement for estimating Figure 4.11. A three-dimensional specimen with internal surfaces. (For color representation see the attached CD-ROM.) 64 Chapter 4 their surface area, reported as the surface area of interface per unit volume of structure. This property, SV, is sometimes referred to as the surface density of the two dimensional feature set imbedded in a three dimensional structure. Line probes are used to sense the area of surfaces. A sample of the population of lines in three dimensional space is usually constructed by sectioning the structure with a plane to reveal its geometry, then superimposing a grid of lines on selected fields in that plane. The events to be counted are the intersections of the probe lines with traces of the surfaces revealed on the plane section. The line intercept count, PL, described in the preceding section in the context of analyzing a two dimensional structure, i.e., the ratio of the number of intersection points to the total length of probe lines in the sample, is made for the grid in each field examined. The governing stereological relationship is PL = 1 SV 2 (4.12) In the preceding section the structure under analysis was viewed as a two dimensional microstructure, and the properties evaluated were two dimensional (perimeter, area). If indeed the microstructures shown are plane probe samples of some three dimensional microstructure, then the features observed have the significance of sections through three dimensional features. To obtain an unbiased estimate of the expected value of PL for the population of lines in three dimensional space, the probes in the sample must be chosen uniformly from the population of positions and orientations of lines in three dimensional space, and the geometric properties estimated are those of the three dimensional microstructure sampled. In this context, the line intercept count obtained from Figure 4.7, PL = 0.24 (counts/mm) takes on new meaning. If the field chosen and the line probes are a properly representative sample of the population of lines in three dimensions, then this value is an unbiased estimate of the expected value ·PLÒ for that population. This requires a carefully thought out experimental design that prepares and selects fields and line directions to accomplish this task. For the single sectioning plane represented in Figure 4.7, this will only be valid if the surfaces bounding the features in the figure are themselves distributed uniformly and isotropically. Then inversion of equation (4.12) then gives SV = 2(0.24) = 0.48 mm2/mm3 for the surface density of the ab interface in this structure. It is difficult to visualize the meaning of this value of surface density for the particles in Figure 4.7. As an aid in seeing whether this result is plausible, estimate the mean lineal intercept for the b particles in the structure by applying equation (3.2) in Chapter 3. The points in the grid may be used to estimate the volume fraction of the b phase by applying equation (4.5). For the placement shown, 3 points hit the b phase, giving a rough estimate of VV b = 3/16 = 0.19. This result may be combined with the surface density estimate to calculate a rough value of the mean lineal intercept of the b particles: l b = 4VV b 4 ◊ 0.19 = = 1.6 mm. 0.48 SV ab (4.13) Classical Stereological Measures 65 The grid points in Figure 4.7 are about 6 mm apart. The mean lineal intercept averages longest dimensions with many small intercepts near the surface of the particles. Thus, this very rough estimate appears to be reasonable. The structure in Figure 4.8 may also be viewed as a section through a three dimensional anisotropic microstructure. This structure is squeezed vertically and elongated horizontally. However, a single orientation of sectioning plane is insufficient to establish the nature of this anisotropy in three dimensions. Indeed, out of the population of line probes in three dimensional space only those that lie in the section plane of Figure 4.8 are sampled. The observation that PL is different in different directions takes on a three dimensional character, as does the anisotropy that this difference measures. The concept of the rose of the number of intersections as a quantitative description of the anisotropy in the microstructure extends to three dimensional microstructure. The three dimensional rose is a plot of the line intercept count as a function of longitude and latitude on the sphere of orientation. An understanding of the meaning of such a construction may be aided by visualizing the meaning of the directed line intercept count. Consider the subpopulation of lines in three dimensional space that all share the same orientation, specified by the longitude, q, and the co-latitude (angle between the direction and the north pole), f, as shown in Figure 4.12. The line intercept count may be performed on a set of directed line probes taken from this subpopulation. This directed count, PL(q, f) measures the total projected area, A(q, f) of surface elements in the structure on the plane that is perpendicular to the direction of the test lines. Thus, in Figure 4.8, the low value for counts for horizontal line reflects the fact that the features have a small footprint when projected on the plane perpendicular to this direction. The large number of counts for vertical line probes derives from the large Figure 4.12. Orientation in three-dimensional space is specified by a point on a sphere represented by spherical coordinates (q, f). The population of line orientations is sampled without bias if the line probes used are cycloids on vertical sections (see text). The dimensions of the cycloid can be expressed in terms of the height, h, of the circumscribed rectangle (sometimes referred to as the minor axis). (For color representation see the attached CD-ROM.) 66 Chapter 4 projected area of these particles on the horizontal plane. The governing stereological relationship is (Underwood, 1970b): PL (q , f ) = SV cos a (4.14) where ·cos aÒ is the average value of the angle between the direction of the test probes and the directions normal (perpendicular) to all of the surface elements in the structure. The circular test lines used for the two dimensional analysis of the structure in Figure 4.8 provide an ingenious and economical mechanism for automatically averaging over the population of line orientations in two dimensional space. However in three dimensional space line probe orientations are uniformly distributed over the sphere of orientation. Averaging over the circle of orientations in a given plane does not provide an unbiased sample of the population of line orientations in three dimensions, as is demonstrated by the shaded regions in Figure 4.12a. Circular test probes in a plane sample the set of orientations in a great circle on the sphere, such as the circle containing the pole in Figure 4.12a. This subset of line orientations in a great circle is uniformly sampled in all directions by a circular line probe in the plane. This implies that the length of line probes with orientations within five degrees of the pole (green cap) is the same as then length sampled within five degrees of the equator (red stripe). The green cap is clearly a very much smaller fraction of the area of the sphere of orientation than is the red stripe. The green cap represents a much smaller fraction of the set of line orientations in three dimensions than does the red stripe. But the circular test probe samples these regions with the same length of line probe. Thus, such a probe design over-samples orientations near the pole and under-samples them near the equator. This design does not provide a uniform sample of the population of line orientations in three dimensions, and will produce a biased estimate of expected values for the line intercept count. An unbiased sample of this population is provided by a sample design known as the method of vertical sections (Baddeley et al., 1986), illustrated in Figure 4.13. A reference direction is chosen in the macroscopic specimen, shown in Figure 4.13a. In the remainder of the analysis, this is called the “vertical direction”. Sectioning planes are prepared which contain (are parallel to) this reference direction, as shown in Figure 4.13b. These sections are chosen so that they represent the variations in the structure with position, and with orientation in “longitude”, i.e., around the vertical direction. In a typical experiment, a longitude direction is chosen at random and a section is cut that contains this direction and the vertical direction. Two other vertical sections, 120° away from the first, are also prepared. This provides a random systematic sample of the longitude orientations. Fields are chosen for measurement in these sections that are uniformly distributed with respect to position on the vertical plane. The vertical direction is contained in each field, as shown in Figure 4.13c, and is known. In order to provide an unbiased sample of the population of line orientations on the sphere, a grid is constructed of line probes that have the shape of a Classical Stereological Measures 67 Figure 4.13. The stereological experimental design based on vertical sections gives an unbiased sample of the population of line orientations in three-dimensional space. (For color representation see the attached CD-ROM.) cycloid curve1, shown in Figure 4.13d. This curve has the property that the fraction of its length pointing in a given direction decreases as the tangent rotates from the vertical direction in a manner that is proportional to the sine of the angle from the vertical. The need for this sine-weighting is indicated in Figure 4.12b. Two equal intervals in the range of f values are shown, one at the pole (green) and the other at the equator (red). The cycloid test line exactly compensates for the sine weighting required of the f direction with longer length of lie probe near the equator (red segment) and shorter near the pole (green segment). This sine weighting of the line length produces an unbiased sample of the population of orientations in three dimensional space. A grid consisting of cycloid test lines with known dimensions, so that the total length of lines probed is known, forms the basis for an unbiased PL count in three dimensional space. Figure 4.14 reproduces the anisotropic structure in Figure 4.8 assuming it is a representative vertical section through a three dimensional structure. The height of the box containing the grid is calibrated at 28.8 mm. The grid is made up of 12 cycloid segments joined together in rows of four. The height of each cycloid segment is (28.8/6) = 4.8 mm. The length of a cycloid segment is twice its height, or 9.6 mm. (Figure 4.12 has the dimensions of a cycloid segment.) Thus, the total length of lines in the grid is (12 segments ¥ 9.6) = 115.1 mm. A total of 24 intersections with ab boundaries are marked and counted. The line intercept count, PL = 24/115.1 = 0.21 1 The cycloid curve is generated by a point on the rim of a wheel as the wheel rolls along a horizontal road. Cycloid segments used in the method of vertical sections as shown in Figure 4.13 correspond to the curve traced out by the first half turn of the wheel. 68 Chapter 4 Figure 4.14. Line intercept count made on a presumed vertical section using a cycloid test line grid gives an unbiased estimate of the surface area per unit volume, SV. Arrow indicates the vertical direction in the specimen. (For color representation see the attached CD-ROM.) Probe population: Lines in three dimensional space This sample: Vertical section with cycloid line probes; 12 cycloid units joined together in three lines Calibration: L0 = 28.8 mm. Each unit has height h = 28.8/6 = 4.8 mm length of each cycloid = l = 2 · h = 9.6 mm total probe length = 12 · l = 115.1 mm Event: Cycloid probe intersects the feature boundary line Measurement: Count the intersections This count: 24 intersections Relationship: ·PLÒ = (1/2)SV Normalized count: PL = 24 counts/115.1 mm = 0.21 counts/mm Geometric property: SV = 2PL = 0.42 (mm2/mm3) (counts/mm). This is an unbiased estimate of the expected value ·PLÒ for lines in three dimensional space. It may thus be used to provide an estimate of the surface area density through the equation: SV ab = 2 PL ab Ê mm2 ˆ Ê 1 ˆ = 2 ◊ 0.21Á ˜ = 0.42 Á ˜ Ë mm ¯ Ë mm3 ¯ (4.15) In the analysis of Figure 4.8 as a two dimensional structure, line intercept counts were made in the vertical and horizontal directions (Figure 4.8). Results of this analysis gave PL,h = 0.059 (1/mm) and PL,v = 0.45 (1/mm). Interpret these counts as directed probes in the three dimensional population of lines. According to Classical Stereological Measures 69 equation (4.14), these counts measure the total projected area of particles on a plane perpendicular to the probe direction, and are related to the average of the cosine of the angle that surface elements make with the probe direction. Measurement with the cycloid test line grid provides an estimate of the unbiased value for SV. These results may be combined to provide a first estimate of the average of the cosines of surface element normals with the horizontal direction and the vertical direction in Figure 4.8: cos a cos a h v PL,h 0.059 = = 0.14 0.42 SV 0.45 P = L ,v = = 1.09 0.42 SV = (4.16) The maximum value of the cosine function is of course 1.00. The estimate for the vertical projection violates this limting value. This impossible result arises because counts were obtained from a single placement of the grids involved, and these counts are only estimates of their expected values. Further, the variance of the ratio of two statistics, like PL,v/SV is the sum of the variances of the individual statistics; the precision of the estimate from a single field is low. Nonetheless, it is clear from these results (as from a casual inspection of the structures, that most of the surface elements in this structure have normal directions near the vertical where a is near zero and cos a is near 1.0. In order to help visualize the geometric significance of the surface area value extimated above it is useful to compute the mean lineal intercept of the features in the structure. This requires an estimate of the volume fraction of the b phase. Apply the point count to the grid in Figure 4.8: there are three hits on b particles. Thus, a very rough estimate of VV is 3/25 = 0.12. Insert this result into the expression for the mean lineal intercept given in equation (3.2), Chapter 3. l = 4 ◊ 0.12 = 1.2 mm 0.42 (4.17) For plate-like structures such as those shown in Figure 4.8, ·lÒ reports the thickness of the plate. This value is plausible for the average plate thickness in this structure. Three-Dimensional Microstructures; Line Length and the Area Point Count Figure 4.15 shows a specimen with some lineal features as part of its microstructure. The triple lines in a cell structure or grain structure described in Chapter 3 provide a common example of such lineal structural features. Dislocations in crystals observed in the transmission electron microscope provide another example. Edges of surfaces are also lineal features. Tubular structures, such as capillaries, or plant roots, may be approximated as collections of space curves. The total length of a one dimensional feature set may be estimated from observations on a set of plane probes. Line features intersect plane probes in a collection of points, as shown in Figure 4.16. These points are counted over a field of known 70 Chapter 4 Figure 4.15. Illustration of the use of a plane probe to intersect linear structures in a volume. Lineal features in a three-dimensional specimen reveal themselves on the plane probe as a collection of points. Counting the number of intersections provides a tool to measure the total length of the structures as shown in Figure 4.16. (For color representation see the attached CD-ROM.) area. The area point count PA is the ratio of the number of points of emergence of these line features in the area, normalized by dividing by the area. If the collection of fields included in the measurement provides a uniform sample of the population of positions and orientation of planes in three dimensional space, then this count provides an unbiased estimate of LV, the length of line features in unit volume of structure: PA = 1 LV 2 (4.18) Construction of a representative sample of the population of planes in three dimensional space is a challenge, since sectioning the sample to form a plane probe divides the sample. For the special situation in which the lineal features are contained in a transparent medium and can be viewed in projection a procedure similar to the method of vertical sections is available (Gokhale, 1992). Figure 4.17 shows a single phase cell structure (e.g., grains in a polycrystal, or botanical cells). The triple points on this section result from the intersection of the plane probe with the triple lines in the three dimensional structure. Calibration shows that the square area to be sampled is 31.1 mm on a side; the area thus probed in this field is (31.1 = 967 mm). The triple points in this area are marked and counted. The grid facilitates counting the triple points by dividing the area into sixteen small squares, which are counted systematically. There are 37 triple points in the area. Thus, PA = 37/967 = 0.038 (counts/mm). Assuming this field is an unbiased sample of the population of planes in the specimen, equation (4.18) may be used to estimate the length of triple lines in this cell structure: Classical Stereological Measures 71 Figure 4.16. Point count made on a section through a structure containing linear features (shown here with finite cross-sectional area for clarity). (For color representation see the attached CD-ROM.) Probe population: Planes in three dimensional space This sample: Square area contained within the measurement frame Calibration: L0 = 40.0 mm; probe area A0 = (40) = 1600 mm Event: Plane intersects the linear features (observed as a small feature) Measurement: Count the intersections This count: 13 intersections Relationship: ·PAÒ = (1/2)LV Normalized count: PA = 13 counts/1600 mm2 = 0.0081 counts/mm2 Geometric property: LV = 2PA = 2 · 0.0081 = 0.016 (mm/mm3) = 16 (km/cm) LV = 2 PA = 2 ◊ 0.038 = 0.076 mm mm3 (4.19) In more familiar units, this length corresponds to 76 (km/cm). This at first surprisingly large result is not unusual for line lengths in microstructures. Figure 4.18 shows a two phase structure (a + b) with particles of the b phase distributed along the aa interfaces and aaa triple lines. This is inferred from the qualitative observation that the b particles are all situated at the former aaa triple lines. One can visualize three different types of triple lines in this structure through their corresponding triple points: aaa, aab, and aaa triple points that are not 72 Chapter 4 Figure 4.17. Section through a cell structure; a count of the triple points measures length of triple lines in the three-dimensional structure, LV.. (For color representation see the attached CD-ROM.) Probe population: Planes in three dimensional space This sample: Square area contained within the grid Calibration: L0 = 31.1 mm; probe area A0 = (31.1) = 967 mm2 Event: Plane intersects the triple line (observed as triple point) Measurement: Count the intersections This count: 37 intersections Relationship: ·PAÒ (1/2)LV Normalized count: PA = 37 counts/967 mm2 = 0.038 counts/mm2 Geometric property: LV = 2PA = 2 · 0.038 = 0.076 (mm/mm3) = 76 (km/cm3) present because they are occupied by b particles (aaa¢). These three classes of triple points are marked with respectively red, green and blue markers in the field be analyzed in Figure 4.18. These results are tabulated in Table 4.3. The total length of aaa triple line, occupied plus unoccupied, is 0.36 (mm/mm). The fraction of the aaa triple line occupied by b particles is 0.16/0.36 = 0.44. Since the volume fraction of b is small (probably only a few percent, since there are no hits on b for the points in the grid in Figure 4.18) this observation that almost half of the cell edge network in three dimensional space is occupied by the b phase measures the evidently strong tendency of b to be associated with the cell edge network. Further, although the b phase is sparse, the triple line length in the category aab is about the same as the total length as that of the cell edge network. Classical Stereological Measures 73 Figure 4.18. A two phase microstructure in which the a phase is a cell structure. The length per unit volume, LV, of three kinds of triple lines may be characterized: aaa, aab, and aaa (occupied). (For color representation see the attached CD-ROM.) Probe population: Planes in three dimensional space This sample: Square area contained within the grid Calibration: A0 = 182.6 mm2 Event: Plane intersects each class of triple line Measurement: Count the intersections This count: aaa = 19, aab = 39, aaa(occ) = 15 Relationship: ·PAÒ = (1/2)LV Normalized count: PAaaa = 19/182.6 mm2 = 0.10 counts/mm2 PAaab = 39/182.6 mm2 = 0.21 counts/mm2 PAaaa(occ) = 15/182.6 mm = 0.08 counts/mm2 Geometric property: LVaaa = 2PAaaa = 2 · 0.10 = 0.20 (mm/mm3) LVaab = 2PAaab = 2 · 0.21 = 0.42 (mm/mm3) LVaaa(occ) = 2PAaaa(occ) = 2 · 0.08 = 0.16 (mm/mm3) Table 4.3. Triple point counts in Figure 4.18. Calibrated probe area is 182.6 mm2 Category aaa aab aaa(occ) Mark red green blue Counts 19 39 15 PA (1/mm2) 0.10 0.21 0.08 LV (mm/mm3) 0.20 0.42 0.16 74 Chapter 4 Invoking equation (4.18) in these last two examples presumes that the fields chosen are unbiased, though extremely limited, samples of the population of plane probes in three dimensional space. The realization of this condition borders upon intractable, particularly with respect to acquiring planes whose normals sample the population of orientations on the sphere. The method of vertical sections devised for line probes has the advantage that each vertical plane section contains all of the line orientations from vertical to the equator. A plane probe through an opaque specimen has a single orientation; each orientation contained in the sample requires a different sectioning plane, and each section cuts the sample into two parts. Preparation of a series of plane probes with normals that are distributed in some sine weighted fashion with respect to an overarching vertical direction in the specimen may be possible if a large volume of specimen material is available (the Orientator, shown in Chapter 7), but it requires a significant effort. A simpler sample design that may accomplish this is shown in Chapter 6. If the sample is transparent, so that the lineal features may be viewed as a projected image, then a strategy similar to the method of vertical sections for lines may be applied. Figure 4.19 shows the projected image of a thick slab of a microstructure with internal lineal features. A probe line on the projection plane corresponds to a probe plane, defined by that line and the projection direction, that passes through the transparent structure. Intersection points between the probe line and the lines in the projected image have a one-to-one correspondence to intersections of the lineal features in the structure with the associated probe plane. The area probed is L0 · t, where L0 is the length of the line probe on the image and t is the specimen thickness. The PA count is the number of intersections with the probe line, divided by this area. If the lineal structure is not isotropic, PA will be different for different directions of the probe line in the projection plane. Vertical and horizontal lines give different counts in the structure shown in Figure 4.19. The statistical test compares the number of counts using the square root of the number as the standard deviation. In the example, 30 vertical counts is clearly greater than 12 horizontal counts (30 ± 30 or 30 ± 5.5 compared to 12 ± 12 or 12 ± 3.5). In order to obtain an unbiased sample of the population of orientations of planes in space, choose a vertical direction for the specimen. Prepare slices of known thickness that contain this vertical direction and uniformly sample the equatorial circle of longitudes; these are called “vertical slices”. To obtain a uniform sampling of the latitude angle (angle from the pole) it is again necessary to sample more planes with orientations near the equator than near the pole. The orientation distribution of plane normals must be sine weighted. This can be accomplished by using cycloid shaped line probes on the projected image, as shown in Figure 4.20. In this application, the cycloids in the grid must be rotated 90° from the direction used in the line probe sample design because the normals to the cycloid curve (representing the plane normals in the structure) must be sine weighted. Each cycloid curve, when combined with the projection direction, generates a cycloid shaped surface with normals that provide a sine weighting of the area of the surface as a function of colatitude orientation. If PL is the line intercept made on these cycloid line probes in the projected image, then the length of lineal features in unit volume of the structure is given by Classical Stereological Measures 75 Figure 4.19. A microstructure consisting of linear features viewed as a projection through a section of thickness t = 2 mm. (For color representation see the attached CDROM.) Probe population: Planes in three dimensional space This sample: Planes represented by their edges which appear as lines on the grid. (vertical and horizontal lines represent two sets of planes) Calibration: L0 = 20 mm, A0 = L0 · T · 5 planes = 20 · 2 · 5 = 200 mm2 Event: Linear features intersect the horizontal or vertical planes Measurement: Count the intersections with the horizontal and vertical grid lines This count: Vertical (red arrows) = 30 Horizontal (green arrows) = 12 Relationship: ·PAÒ = (1/2)LV Normalized Count: PA = (30 + 12)/200 mm2 = 0.21 (counts/mm2) Geometric Property: LV = 2PA = 0.42 mm/mm3 LV = 2 PA = 2 PL t (4.20) This result also forms the basis for estimating the total length of a single lineal object suspended in three dimensional space, like a the whole root structure of a plant, a wire frame object or the set of capillaries of an organ. In this application it must be possible to view the whole lineal structure from any direction. Choose a vertical direction, then a vertical viewing plane. The plane of observation must have an area A0 large enough to view the entire object. Visualize a box with 76 Chapter 4 Figure 4.20. A line probe on a projected image represents a surface probe in the volume being projected (a). A cycloid-shaped line probe (b) produces a cycloidal surface probe, which uniformly samples the latitudinal angle of the probe. Note that the cycloid is rotated 90 degrees with respect to the orientation used in Figure 4.13. (For color representation see the attached CD-ROM.) the area of the projection plane and depth t large enought to contain the whole object. The volume of the box is A0 · t. Construct a grid of cycloid test lines with major axis parallel to the vertical direction on the projection plane. The total length LT of these probes is calibrated and known. Count intersections of the cycloid line probes with the projected lineal features. If PL is the number of intersections per unit length of line probes, then manipulation of equation (4.20) gives L 2 2 P = PL = A0t t t LT A0 P L= LT LV = (4.21) where ·PÒ is the expected value of the number of intersections that form between the cycloid grid and the projected image of the structure. Thus, a simple count of the number of intersections between the cycloid grid and lines in the projected image, replicated and averaged over a number of orientations for the vertical projection plane, provides an unbiased estimate of the actual length of the feature in three dimensions. Summary Normalized geometric properties, volume fraction VV, surface area density SV and length density LV are accessible through simple manual counting measurements applied to interactions between grids that act as geometric probes and corresponding features in the three dimensional microstructure. Corresponding measurements performed with image analysis software may automate these measurements if the features to be analyzed can be detected satisfactorily. Geometric properties of single features can also be estimated if the observations can encompass the entire feature. Classical Stereological Measures 77 The connecting relationships make no assumptions about the geometry of the microstructure. It may be simple or complex, and exhibit anisotropies and gradients. Each of the relationships involves the expected value of counts of some event that result from the interaction of probe and microstructure. The typical stereological experiment is designed to yield an unbiased estimate of the corresponding expected value. This requires a sample design that guarantees that the population of probes, encompassing all possible positions and, for lines and planes, all orientations on the sphere, will be uniformly sampled. Uniform sampling of all positions requires selection of fields that accomplish that goal. Uniform sampling of orientations can be accomplished with appropriately configured test probes. In two dimensional structures, circular test lines automatically sample all orientations uniformly. The method of vertical sections uses cycloids to guarantee an unbiased sample of orientations of line probes in three dimensional space. A similar procedure, with the cycloids rotated ninety degrees, and applied to projected images, gives isotropic sampling of the population of orientations of planes in three dimensions.