Chapter 12 Unfolding Size Distributions In the preceding chapter, methods were discussed for predicting the distribution of intercepts that a random probe would make with objects in a threedimensional space. This kind of modeling can be done using either line or plane probes. With line probes the length of the intersection can be measured, and with plane probes the area as well as the shape of the intersection can in principle be determined. This approach requires making several assumptions: the size and shape of the objects being sampled must be known, and the probes must be isotropic, uniform and random with respect to the objects. In actual practice, the experiment is carried out in the opposite direction. A real specimen is probed with lines or planes, and the data on intercept lengths, areas and perhaps shape information are recorded. Then the goal is to calculate the (unknown) sizes of the objects that must have been present in order to produce those results. Note that it is not possible to determine the size of any single object this way, from the intersection that the probe makes with it. It is only in the statistical aggregate that a distribution of objects can be inferred. And even then it is far from easy to do so. There are several aspects to the problem, which will be discussed below. First, mathematically the inverse problem is ill-conditioned. The small statistical variation in measured data (which arises because a limited number of measurements are made, even if they are an unbiased sample of the specimen) grows to a much larger variation in the calculated answer because of the ways that these variations propagate through the calculation. Second, a critical assumption must be made for the method to be applied at all: we must assume that we know what the shape of the objects is, and usually that they are either all of the same shape or have a very simple distribution of shapes. The most popular shape assumption is that of a sphere, because its symmetry makes the mathematics easy and also relaxes the requirements for isotropic sampling. But in fact not very many structures are actually spheres, and even a small variation from the assumed shape can introduce quite a large bias in the calculated results. To make matters worse, many real specimens contain objects that have a wide variety of shapes, and the shape variation is often a function of size. This presents major problems for unfolding techniques. In spite of these difficulties, the unfolding approach was the mainstay of classical stereology for decades and is still used in spite of its limitations in many situations. Certainly there are specimens for which a reasonable shape assumption can be made—spheres for bubbles, cubes for some crystalline materials (e.g., tungsten carbide), cylinders for fiber composites, and so on. And the use of computer-based 297 298 Chapter 12 measurement systems makes it possible to collect enough data that the statistical variations are small, limiting the extent of the errors introduced by the inverse solution method. Linear Intercepts in Spheres The preceding chapter developed the length distribution of lines intersecting spheres. The frequency distribution for a sphere of diameter D is simply a straight line, and when measured intercept lengths are plotted in a conventional histogram form we would expect to see a result as shown in Figure 12.1. The presentation of data in histogram form is the most common way to deal with measurement data of all types. The bin width d is typically chosen to permit each bin to accumulate enough counts for good statistical precision, and enough bins to show the shape of the distribution. In most cases from 10 to 15 bins are linearly spaced in size up to the largest value obtained in measurement. One consequence of the number of bins used is the ability to measure spheres (or other objects) whose sizes vary over that range of dimensions. Ten or 15 bins allows determining spheres in that number of sizes. If the size range of the spheres is much larger, say 50 : 1, and information on the smallest is actually needed, then a histogram with at least that many bins is required. This greatly increases the Figure 12.1. Frequency distribution of linear intercepts in a sphere of diameter D, shown as a continuous plot and as a discrete histogram. (For color representation see the attached CD-ROM.) Unfolding Size Distributions 299 Figure 12.2. Frequency distribution of linear intercepts from a mixture of sphere sizes, shown as a continuous plot and a discrete histogram. (For color representation see the attached CD-ROM.) number of measurements that must be made to obtain reasonable counting precision. When a mixture of sphere sizes is present, the measured distribution represents a summation of the contribution of the different sizes each in proportion to the abundance of the corresponding size. As shown in Figure 12.2, the result is that the histogram bins corresponding to smaller intercept lengths contain counts from intersections with many different size spheres, while the largest bin contains counts only from the largest spheres. This offers a straightforward graphical way to unfold the data (Lord & Willis, 1951). As shown in Figure 12.3, knowing that the largest bin contains counts only from the largest spheres and that those spheres should also generate a proportionate number of shorter intercepts allows subtracting the expected number of shorter intercepts from each of the other (smaller) bins. This process can then be repeated for the next bin, using the number of counts remaining, and so on. Notice that the number of counts in the smaller size bins (and hence the estimated number of smaller size spheres) is obtained by successive subtractions. Subtraction is a particularly poor thing to do with counting data, since the difference between two numbers has a standard deviation that is the sum of the deviations of the two numbers whose difference was taken. This means that in effect the 300 Chapter 12 Figure 12.3. Unfolding the linear intercept distribution from a mixture of sphere sizes. Lines project the histogram bin heights onto a vertical axis and the increments given the relative proportion of each sphere size. (For color representation see the attached CD-ROM.) uncertainty of the estimated number of smaller size spheres grows rapidly and depends strongly on the number of larger size spheres and the precision with which they are determined. Hence the need to obtain large numbers of interception counts, since the standard deviation of any counting data is simply the square root of the number counted. In other words if you count one hundred events the one-sigma uncertainty –— is ÷100 = 10 out of 100 or 10%, and to reduce this to 1% precision would require counting 10,000 events. Plane Intersections Linear intercepts were primarily used when manual measurements were required to obtain data, because they could be performed by drawing (random) lines on images of sections and measuring the length of the intersections. With modern computer-based instruments, it is usually easier to measure the intersection areas made by the sampling plane with the objects. It is very difficult to use plane probes that must be isotropic, uniform and random (IUR) in a real object because once a single plane section has been taken, other planes that might intersect that one cannot be generated. However, for specimens which are themselves IUR any plane is as good as another, and examination of plane sections is a very common approach. Every section through a sphere is a circle as shown schematically in Figure 12.4. Figure 12.5 shows an image of pores in an enamel coating on steel. The pores may be expected to be spherical as they result from the evolution of gas bubbles in the firing of the enamel, and all of the intersections are observed to be circular. The distribution of the sizes of circles can be calculated analytically. The probability of obtaining a circle of radius d ± dd (corresponding to counts that will fall into one bin of a histogram) from a sphere of diameter D is equal to the vertical thickness of a slice of the sphere with that diameter. The shape of the distribution is Unfolding Size Distributions 301 a b c d e Figure 12.4. Different size circles are produced by sectioning of a sphere. (For color representation see the attached CD-ROM.) P(dcircle) = d/(D · (D2 - d 2)1/2) (12.1) as shown in Figure 12.6. As for the linear intercept case, if there are several different sizes of spheres present in the sample a particular size of intercept circle could result from the intersection of the plane with several different sphere sizes. The resulting measurement 302 Chapter 12 Figure 12.5. Example of sectioning of spheres: bubbles in a fired enamel coating. histogram would show the superposition of data from the different sizes in proportion to their relative abundance and to their size (it is more likely for the plane to strike a large sphere than a small one). This distribution might be unfolded sequentially and graphically as shown before for the linear intercepts, since the largest circles could only come from the largest spheres. Then as before, a corresponding number of intercepts would be subtracted from each smaller bin and the process repeated until all sizes are accounted for. This is not a very practical approach because of the tendency for statistical Figure 12.6. The distribution of planar intercepts through a sphere is proportional to the vertical thickness dz of a slice of the sphere covering a range of circle sizes dr. Unfolding Size Distributions 303 variation to concentrate in the smaller size ranges, and instead a simultaneous solution is generally used. For the case of spheres, this method was first described by Saltykov (1967) and has been subsequently refined by many others, with an excellent summary paper by Cruz-Orive (1976) that includes tables of coefficients suitable for general use. The method is simply to solve a set of simultaneous equations in which the independent variables are the number of circles of various sizes (measured on the images and recorded as a histogram) using the coefficient matrix, to obtain the numbers of spheres of each size present in the material. The matrix is determined by calculating as discussed above, from geometric probability, the frequency distribution for the number of circles in each size class due to three dimensional objects of each size. This can be written as NAi = a¢ij · NVj (12.2) where the subscript i ranges over all size classes of circles in the measured histogram and the subscript k covers the corresponding sizes of spheres in the (unknown) distribution of three-dimensional objects. The equations are more readily solved by using the inverse matrix a, so that NVj = (1/d)a ij · NAi (12.3) where d is the size of the bin used in the histogram. A typical a matrix of coefficients is shown in Table 12.1. As an example of the use of this method, Figure 12.7 shows images of nodular graphite in a cast iron. The sections are all at least approximately circular and we will make the (typical) assumption that the nodules are spherical. Figure 12.8a shows a distribution of circle sizes obtained from many fields of view covering the cross section of the enamel. Converting these NA data using the coefficients in Table 12.1 produces the plot of sphere sizes shown in Figure 12.8b. Other Shapes The “sphere unfolding” method is easily programmed into a spreadsheet, and is often misused when the measured objects are not truly spheres. Of course, similar matrices for a’ can be computed for other shapes of objects, and inverted to produce a matrices. In fact, this has been done for an extensive set of convex shapes including polyhedra, cylinders, etc. (see for example Wasén & Warren, 1990). It is well to remember, however, that the results are only as useful as the assumption that the three-dimensional shape of the objects is known and is the same for all objects present. Another shape model that has been applied to the “nearly spherical” shapes that arise in biological materials in particular (where membranes tend to produce smooth boundaries as opposed to polyhedra) is that of ellipsoids. These may either be prolate (generated by revolving an ellipse around its major axis) or oblate (generated by revolving the ellipse around its minor axis), and can be used to approximate a variety of convex shapes. NV(1) NV(2) NV(3) NV(4) NV(5) NV(6) NV(7) NV(8) NV(9) NV10) NV(11) NV(12) NV(13) NV(14) NV(15) NV(1) NV(2) NV(3) NV(4) NV(5) NV(6) NV(7) NV(8) NV(9) NV10) NV(11) NV(12) NV(13) NV(14) NV(15) Alpha matrix for sphere unfolding (Cruz-Orive) NA(2) NA(3) NA(4) NA(5) NA(1) 0.26491 -0.19269 0.01015 -0.01636 -0.00538 0.27472 -0.19973 0.01067 -0.01691 0.28571 -0.20761 0.01128 0.29814 -0.21649 0.31235 Alpha matrix for sphere unfolding (Saltykov) NA(1) NA(2) NA(3) NA(4) NA(5) 0.1857 -0.0750 -0.0261 -0.0132 -0.0080 0.1925 -0.0776 -0.0270 -0.0136 0.2000 -0.0804 -0.0280 0.2085 -0.0836 0.2182 NA(6) -0.00481 -0.00549 -0.01751 0.01200 -0.22663 0.32880 NA(6) -0.0054 -0.0083 -0.0140 -0.0290 -0.0872 0.2294 NA(7) -0.00327 -0.00491 -0.00560 -0.01818 0.01287 -0.23834 0.34816 NA(7) -0.0039 -0.0055 -0.0085 -0.0146 -0.0301 -0.0913 0.2425 NA(8) -0.00250 -0.00330 -0.00501 -0.00571 -0.01893 0.01393 -0.25208 0.37139 NA(8) -0.0028 -0.0039 -0.0056 -0.0088 -0.0151 -0.0319 -0.0961 0.2582 NA(9) -0.00189 -0.00250 -0.00332 -0.00509 -0.00579 -0.01977 0.01527 -0.26850 0.40000 NA(9) -0.0022 -0.0029 -0.0040 -0.0057 -0.0090 -0.0155 -0.0329 -0.1016 0.2773 NA(10) -0.00145 -0.00186 -0.00248 -0.00332 -0.00516 -0.00584 -0.02071 0.01704 -0.28863 0.43644 NA(10) -0.0016 -0.0022 -0.0028 -0.0041 -0.0058 -0.0091 -0.0163 -0.0346 -0.1081 0.3015 NA(11) -0.00109 -0.00139 -0.00180 -0.00242 -0.00327 -0.00518 -0.00582 -0.02176 0.01947 -0.31409 0.48507 NA(11) -0.0013 -0.0016 -0.0021 -0.0028 -0.0040 -0.0059 -0.0094 -0.0168 -0.0366 -0.1161 0.3333 NA(12) -0.00080 -0.00101 -0.00129 -0.00169 -0.00230 -0.00315 -0.00512 -0.00565 -0.02293 0.02308 -0.34778 0.55470 NA(12) -0.0009 -0.0012 -0.0016 -0.002 -0.0027 -0.0038 -0.0058 -0.0095 -0.0174 -0.0386 -0.1260 0.3779 Table 12.1. Alpha matrices to unfold sphere size distribution from measured circle sizes NA(13) -0.00055 -0.00069 -0.00087 -0.00113 -0.00150 -0.00208 -0.00288 -0.00488 -0.00516 -0.02416 0.02903 -0.39550 0.66667 NA(13) -0.0007 -0.0007 -0.0010 -0.0013 -0.0018 -0.0026 -0.0037 -0.0057 -0.0093 -0.0178 -0.0408 -0.1382 0.4472 NA(14) -0.00033 -0.00040 -0.00051 -0.00066 -0.00087 -0.00117 -0.00167 -0.00234 -0.00427 -0.00393 -0.02528 0.04087 -0.47183 0.89443 NA(14) -0.0004 -0.0006 -0.0006 -0.0009 -0.0010 -0.0015 -0.0021 -0.0031 -0.0051 -0.0087 -0.0171 -0.0420 -0.1529 0.5774 NA(15) -0.00013 -0.00016 -0.00020 -0.00026 -0.00034 -0.00045 -0.00062 -0.00094 -0.00126 -0.00298 -0.00048 -0.02799 0.08217 -0.68328 2.00000 NA(15) -0.0001 -0.0002 -0.0003 -0.0004 -0.0005 -0.0006 -0.0009 -0.0013 -0.0020 -0.0033 -0.0061 -0.0130 -0.0360 -0.1547 1.0000 Unfolding Size Distributions 305 a b Figure 12.7. Images of graphite nodules in cast iron. a b Figure 12.8. a) The distribution of measured circle sizes from the cast iron in Figure 12.7, and a) the calculated distribution of sphere sizes that generated it. (For color representation see the attached CD-ROM.) 306 Chapter 12 All of the sections made by a plane probe passing through an ellipsoid produce elliptical profiles (Figure 12.9). These can be characterized by to dimensions, for instance the lengths of the major and minor axes. DeHoff (1962) presented a method for unfolding the size distribution of the ellipsoids, which are assumed to have a constant shape (axial ratio q) but varying size (major axis length), from the size distribution of the ellipses. This is a modification of the sphere method in which NVj = (K (q ) d ) ◊ Â a ij N Ai (12.4) d is the size increment in the histogram, and the additional term K(q) is a shape dependent factor that depends on the axial ratio of the ellipsoid q and whether the ellipsoids are prolate or oblate; this factor shown in Figure 12.10 for both cases. Note that it is necessary to decide independently whether the generating objects are oblate or prolate, since either can produce the same elliptical profiles on the plane section image. If this is not known a priori, it can sometimes be determined by examining the sections themselves. If the generating particles are prolate, then the diameters of the most equiaxed sections will be close in size to the width of those with the highest aspect ratio, while if the particles are oblate then the most a b c d Figure 12.9. Ellipses formed by planar sections with prolate (a..d) and oblate (e..g) ellipsoids. (For color representation see the attached CD-ROM.) Unfolding Size Distributions 307 e f g Figure 12.9. Continued a b Figure 12.10. K shape factors for prolate (a) and oblate (b) ellipsoids as a function of the axial ratio q of the generating ellipse. 308 Chapter 12 Figure 12.11. Elliptical profiles from prolate and oblate ellipsoids. (For color representation see the attached CD-ROM.) equiaxed sections will be close to the length of those with the highest aspect ratio (Figure 12.11). In either case, the axial ratio q is taken as the aspect ratio of the most elongated profiles and used to obtain k from the graph. It is in principle possible to extend the unfolding method of analysis to convex features that do not all have the same shape. The profiles observed in the plane section can be measured to determine both a size and shape (e.g. an area and eccentricity) and a two-dimensional histogram of counts constructed (Figure 12.12). From these data, calculation of both the size and shape of the generating ellipsoids or other objects should be possible. The calculation would be of the form NVij = Â a ijkl ◊ N Akl (12.5) where the subscripts i, j cover the size and shape of the three-dimensional objects and k, l cover the size and shape of the intersection profiles. Calculating the fourdimensional a’ matrix and inverting it to obtain a is a straightforward extension of the method shown before, although selecting the correct three-dimensional shape and its variation is not trivial. The major difficulty with this approach is a statistical one. The number of counts recorded in many of the bins in the two-dimensional NA histogram will be sparse, especially for the more extreme sizes and shapes. These are important to the solution of the equations and propagate a substantial error into the final result. Figure 12.12. An example of a two-dimensional histogram of size and shape of intersections made by a plane probe. (For color representation see the attached CD-ROM.) Unfolding Size Distributions 309 Shape information from the profiles can be useful even when a single shape of generating object is present. Ohser & Nippe (1997) have shown that measuring the shapes of intersections with cubic particles (which are polygons with from 3 to 6 sides) can significantly improve the determination of the size distribution of the generating cubes. Simpler Methods The unfolding of linear intercepts is one of the simplest methods available, in use long before modern computers made the solution of more complicated sets of equations easy. It can be performed using other distributions than the linear one that applies to spheres, of course. If the actual shape of the objects present is known, the exact distribution could be generated as discussed in the preceding chapter and used for the unfolding. A relatively compact and general approach has been proposed for the case of other convex shapes by several investigators (Weibel & Gomez, 1962; DeHoff, 1964). Using the usual nomenclature of NV (number per unit volume), NA (number per unit area), VV (volume fraction), PL (intersections per length of line) and PP (fraction of points counted), the relationships proposed are NV = (K b) ◊ N A 32 VV 12 (12.6) and NV = 2g ◊ NA PL PP (12.7) Note than both procedures require determining the total volume fraction of the particles (VV or PP) as well as the number of objects per unit area of section plane, and one requires determining the mean linear intercept (the inverse of PL). The parameter K takes into account the variation of the size of features from their mean, and is often assumed to have a value between 1.0 and 1.1 for many real structures that constitute a single population (e.g., cell organelles). Figure 12.13 shows K as a function of the coefficient of variation of the sizes of the objects. Figure 12.13. The shape distribution parameter K as a function of the relative standard deviation of the mean diameter. 310 Chapter 12 Table 12.2. Shape coefficients for particle counting Shape Sphere Prolate Ellipsoid 2 : 1 Oblate Ellipsoid 2 : 1 Cube Octahedron Icosahedron Dodekahedron Tetrakaidekahedron g 6.0 7.048 7.094 9.0 8.613 6.912 6.221 7.115 b 1.38 1.58 1.55 1.84 1.86 1.55 1.55 1.55 Notice that the size distribution of the intercept lengths is not used here. Instead, the fact that the mean value of the intercept length of a line probe through a sphere is just (p/4) times the sphere diameter is used (and adjusted for other shape objects). b and g are shape factors that depend upon an assumed shape for the threedimensional objects. They vary only slightly over a wide range of convex shapes, as indicated in Table 12.2 and Figure 12.14. This method represents a very low cost way to estimate the number of objects present in three dimensions based on straightforward measurements taken on plane sections. However, it is presently giving way to more exact and unbiased techniques described under the topic of the “new” stereology. Lamellae Linear probes are also useful for measurement of other types of structures. One of the advantages of linear probes is the ease with which they can be produced with uniform, random and isotropic orientations (which as noted above is very difficult for plane probes). Drawing random lines onto a plane section can be used to measure intercept lengths, either manually or using a computer. Lamellar structures occur in many situations, including eutectic or other layered structures in materials, sedimentary layers in rocks, membranes in tissue, Figure 12.14. The shape coefficient b for ellipsoids and cylinders of varying axial ratio. Unfolding Size Distributions 311 Figure 12.15. A plane section through a layered structure produces an apparent layer thickness that is generally larger than the true perpendicular dimension. (For color representation see the attached CD-ROM.) and so on. Because these structures will not in most cases lie in a single known orientation perpendicular to the section plane, the apparent spacing of the layers will be larger than the true perpendicular spacing (Figure 12.15). There is no guarantee that on any particular section plane the true spacing will be revealed. Random (IUR) linear intercepts through a layer of true thickness t produce a distribution of intercept lengths as shown in Figure 12.16. If the data are replotted as a histogram of 1/l, the inverse of the measured intercept length, a much simpler distribution is obtained (Gundersen, 1978). The mean value of this Figure 12.16. Plots of the intercept length distribution for linear intercepts through a layer. (For color representation see the attached CD-ROM.) 312 Chapter 12 triangular distribution is just (2/3) (1/t) so the true layer thickness can be calculated as t = 3 2 (1 lm ) (12.8) where lm is the mean measured intercept length. Furthermore, if the layers are not all the same thickness but have a range of thickness values, the distribution can be unfolded using the same graphical technique as shown above for linear intercepts through spheres.