Quantitative microscopy
Transcription
Quantitative microscopy
Quantitative microscopy Ilan Hammel DSc One picture is worth ten thousand words Problems of Quantitative Microscopy The amount of material that is actually examined in the microscope is often a tiny fraction of the whole object of interest. If a 100x magnification oil immersion lens is used a 3 dimensional (3D) sampling 'brick' with dimensions of about 70µm x 90µm x 10µm can be used for counting cells. For a brain of volume 500 cm3 each brick thus represents a fraction of about 1/8,000,000,000th (an eight billionth) of the whole brain. 1 How can a meaningful and representative estimate of the total number of neurons in the brain be made from this tiny fraction of the brain? BY USING STEREOLOGICAL METHODS Stereology is related to the areas of stochastic geometry and spatial statistics. It can be described as geometric sampling theory or a set of sampling methods for quantifying geometric parameters such as the number of cells, the length of a root, the area of a particular region in an agricultural field, or the volume of an organ. Mandarim -De-Lacerda CA. Stereological tools in biomedical research. An Acad Bras Cienc.;75:469-86. 2003. Problems of Quantitative Microscopy The act of taking a cross-section through an object, such as a polished metallurgical section or a thin histological section, causes the feature of interest to be seen with reduced dimensionality. An unavoidable reduction in dimensionality (from Dn to Dn-1) has been introduced by the sectioning process. 2 Sectioning features in a 3D space with a plane showing the area intersection with a volume, the line intersection with a surface, and the point intersection with a linear feature. Compression caused by the knife action in microtomy direction. Paraffin - or similar embedding media like Paraplast or Paraplast Plus - causes an intrinsic distortion of structures: the shrinkage is caused by fixation, dehydration and paraffin infiltration. In paraffin embedded material shrinkage can be around 25% and compression around 10% relative to the fresh material. Vo /Vf = (lo /l f) 3 3 Geometric Probability - Stereology • Both of these problems are simply resolved in practice by applying a branch of applied mathematics and probability theory known as stereology. • It is the science of obtaining 3D information from 2D sections. • It was founded in the early 1960's by a group of Mathematicians, Biologists and Material Scientists who had a common interest in quantifying 3D structure from 2D cross-sections. • The founders of this new inter-disciplinary science realised that there was a common generic problem. The 3D features they were interested in could not be directly quantified from cross-sections. • The insight provided by the founders of stereology was that they realised that the features seen on a section were related to the features in 3D space in a well-defined statistical way. Sampling • The problem of drawing sound conclusions about a geometrical feature from microscopical observations has a direct analogy in the world of survey sampling . For example, consider the task confronted by opinion pollsters. • By taking a small sample (perhaps 500 people) from a large population (6,000,000), it is possible for opinion pollsters to give surprisingly accurate predictions of the outcome of elections. Correct Sampling = All parts of the structure or population have to have the same chance to contribute to the sample. Accurate forecasts rely upon two factors. 1. The sample must be representative of the population. 2. The questions that are asked should generate truthful answers. If either of these principles are ignored then the results of an opinion poll are likely to be unreliable, that is to say, the results will have a systematic error or bias. 4 Uniform Random Sample Each and every member of the population must have an equal chance of being selected for the sample before the sampling begins. If the opinion pollster took a sample of people coming out of the Stock Exchange he/she would almost definitely get a markedly different result than If he/she had sampled people coming out of an unemployment advice centre. The samples are not representative because the selection of individuals is not of uniform random. A uniform random sample of a population is said to be free from selection bias. In order to ensure that the questions get truthful answers from the sampled individuals: 1. the interviews and questionnaires should not use leading questions 2. adopt a moralistic high tone 3. should be carried out in a neutral manner. The quantification of micro-structures from samples taken from a larger object can be considered as a spatial variant of the opinion pollsters problem. • The population in this context is the totality of the geometrical feature of interest within the object (i.e. volume, surface area, length or number). • The sample is the subset of material that is actually imaged in the microscope. • The questions that can be asked take the form of geometrical probes such as points, lines, planes and volumes. 5 To ensure that the conclusions drawn from a microscopic analysis are representative of the whole object Care Must Be Used In Taking A (Spatial) Sample And Asking Truthful (Geometrical) Questions. A representative sample of the object of interest can be obtained By taking a uniform random sample from the object (e.g lung). • This means that each and every portion of the object should have an equal chance of being seen in the microscope field of view before the sampling begins. • In practice this sampling will often have a nested structure. The Nested Structure of Tissue Sampling • The object will be cut exhaustively into thick slabs. • A random sample of these are further sectioned into rectangular blocks. • A random sub-sample of the blocks are embedded in wax, resin or plastic. • A random series of sections are taken with a microtome and mounted on microscope slides. • Randomly positioned fields of view on the slides are then imaged and used for analysis. 6 Geometrical probes • Geometrical probes (points, lines, planes or volumes) are 'thrown' into the object of interest. • The number of intersections is dependent both on the amount of the probe and the amount of the feature of interest. • Providing that either the set of probes or set of features are randomised properly there is a well defined statistical relationship between the amount of the feature and the number of intersections that are produced. • These statistical relationships represent the classical ratio estimators of stereology Geometrical probes The probability that a particular probe hits (or intersects) a feature is dependent on both the type of probe and feature. • A point thrown at random within a reference space hits features within the reference space with a probability proportional to the volume of the features. • A line of isotropic random orientation thrown into a reference space will hit features with a probability proportional to the surface area of the features. In practice: • It is difficult to microscopically probe a 3D object directly with a set of point or line probes. • Therefore, the material is sectioned first and a suitable test system of points, lines, sample frames etc. is randomly superimposed on the planar sections. This probing of a section is equivalent to direct probing of 3D space and is the essence of the stereological approach. 7 Calculation method for determining the chances (frequency) of obtaining different-sized transections from transecting a sphere in size class k of diameter 2R with many parallel planes (R=4r/ π). Bias in stereological measurements - Holmes effect Finite section thickness introduces two opposite sources of bias into conventional stereological measurements. • These have been recognized for a long time [Holmes, A. (1927) "Petrographic methods and calculations" Murby, London]. • Various correction procedures have been derived (Cruz-Orive, 1983). Bias in stereological measurements Holmes effect Correction procedures are not used widely. • Because they require knowing the section thickness. • Making some assumptions about feature shape and uniformity (also isotropy of orientation). • and/or performing more measurement work to obtain results. • In practice, sections with no more than 3µm thick are useful to most of stereological procedures (an exception to the use of thin sections is the disector's method, where thick section over 20µm could be used). 8 Section Thicknes Small method Unbiased estimation of the Area of a 2D object The area (A) of a 2D object can be estimated without bias by randomly translating a point grid over the object in both the x and y direction. A=P(loln)/Mag2 The 2D nucleator Area = π li2 The 2D nucleator estimates the area of any object irrespective of its size, shape, and orientation by measuring the distance between a "central point" in the object and the intersections between the object boundary and the four radiating test lines. 9 The 2D nucleator Area = π li2 Calculation of Linear Magnification One key piece of information required in stereological work is the magnification (Mag) of the images used for measurements. The best method for determining the final linear magnification of an image is to capture an image of a calibrated graticule at the same magnification used for the images. 1mm Unbiased estimation of 2D Boundary length in the plane (The method is a result of the famous Buffons needle problem.) Boundary length is estimated by randomly translating and rotating a grid of parallel lines, of grid spacing D, over the object of interest. The number of intersections between the object boundary and the set of test lines is counted (I) and the boundary length is then estimated from B = (π / 2) I D 10 Buffons needle problem LA = LINEAR ELEMENTS IN A PLANE LA = LINEAR ELEMENTS IN A PLANE The average number of intersections formed by an array of test lines which are uniformly distributed in both position and orientation is: 2π d λ dφ dλ π /2 dλ d N L (φ) = ∫ cos( φ) = [ cos(φ ) d φ] = 0 L2 2π 2π L2 ∫0 π L2 2dλ 2λ = NL=∫ λ π L 2 π L2 LA = (π/2) NL = (π/2) I D 11 B =(π/2) (A/L) I Unbiased estimation of root length in the plane B = (π / 2)( I / 2 ) D = (π / 4) I D Mag = 2.35 I = 41 D = 6cm / 2.35 = 2.56 B = 82.4 cm Mag = 25 I = 71 D = 4mm / 25 = 160µm B = (π / 2) I D 12 12 14 20 12 13 B = 17.8 mm Unbiased estimation of 2D Objects per Unit Area (Na) 3DUDPHWHU 9ROXP HGHQVLW \ 6XUIDFHGHQVLW \ / HQJW KGHQVLW \ 1 XPEHUGHQVLW \ 9ROXP H 6XUIDFH / HQJW K 1 XPEHU &RQQHFWLYLW\ 3URILOHGHQVLW \ ,QWHUVHFWLRQGHQVLW \ ' LP HQVLRQ L0=L3/L3 L-1=L2/L3 L-2=L1/L3 L-3=L0/L3 L3 L2 L1 L0 L0 -2 L =L0/L2 L-1=L0/L1 8QLW V 1RQH µm-1 µm-2 µm-3 µm3 µm2 µm 1RQH 1RQH µm-2 µm-1 1 RW DWLRQ Vv Sv Lv Nv 9 6 / 1 Χ QA IL Conversions Area Conversion of mm 2 to m2 1 mm 2 = (1 × 10 -3 m)2 = (10 -3 m)2 = 10 -6 m 2 Conversion of µm 2 to m 2 1 mm 2 = (1 × 10 -6 m)2 = (10 -6 m)2 = 10 -12 m 2 Volume Conversion of µm 3 to m 3 1 µm 3 = (1 × 10 -6 m)3 = 10 -18 m 3 Reciprocal Length (L 2/L 3=L -1 e.g. surface density) Conversion of µm -1 to m -1 1 µm -1 = (1 × 10 -6 m)-1 = (10 -6 m)-1 = 1/(10-6 m) = 10 6 m -1 13 Design-based and model-based stereology • Model-based - assumptions of randomness are made about the structure (probes are fixed, structure random). • Design-based - no assumptions are made, the sampling uses carefully randomised probes (structure fixed, probes random). Design-based and model-based stereology • Material science problems are often cast in a model-based setting. • Biological problems have increasingly been seen as ideal for design-based methods. • Neither of these approaches relies upon unrealistic model assumptions of sphericity or uses ill conditioned unfolding methods. DELESSE'S PRINCIPLE Unbiased estimation of Volume Fraction – Vv • The volume fraction of a phase within a 3D reference volume can be estimated without bias by randomly translating a point grid over a uniform random cross-sectional image in both the x and y direction. • The number of points hitting the reference space (Pref) and the number of points hitting the phase of interest Y , (PY) are counted and volume fraction of Y is estimated from Vv (Y,ref) = 14 Σ AY / ΣAref = ΣLY / ΣLref = ΣPY / Σ Pref The Delesse's principle is based on homogenous structures. In stereological viewpoint in "isotropic and uniform random, IUR, sections". Vv = Volume Fraction f(z) is the probability of finding a test plane at a position between z and z+dz Vv = Volume Fraction If the test planes are uniformally distributed in position, so that the probability of finding a test plane in any interval is the same as that for any interval, then f(z)dz = dz / L L dz L A(z) dz V = 3 =V 3 v L L A A = ∫0 A A(z ) L = ∫0 15 Unbiased estimation of Volume Fraction - Vv The number of points hitting the reference space (light and dark grey) is 15 and the number hitting Y (light grey phase only ) is 6 (5+2x0.5) , thus the volume fraction of Y in the reference volume is estimated to be, 1 2 4 3 5 Vv = 6 / 15 = 0.40 A test system of cycloids Reference Volume Many of the methods used in stereology provide estimates of the amount of a feature per unit reference volume i.e. densities . For example: Volume density, Vv, the volume proportion of the phase of interest within a reference volume. Surface density, Sv, the area of an interface within a unit reference volume. Length density, Lv, the length of a linear feature within a unit reference volume. Numerical density, Nv, the number of discrete objects in a unit reference volume. 16 'Reference Trap' Final data should be presented per animal and not as density e.g. Hypertrophy vs Hyperplasia. Sv = 6 (2r)2 / (2r)3 = 3 / r Sv = 4πr2 / (4πr 3/3) = 3 / r Sv = Surface Density = Surface to volume ratio SV = Surface Density = Surface to volume ratio π 2π dS sin(Φ )dΦ dΘ dS d NL = ∫ ∫ cos(Θ ) = 0 0 L3 4π 2L3 The number of intersections dS S = formed all surface area in the N L = ∫∫S 3 2L 2L3 structure is the sum: The total extent of surface area per unit volume = SV = 2 NL 17 SV = Surface density is directly related to the number of interactions (Ii) formed with test lines of length Lt : SV= 2 Ii / Lt COHERENT SQUARE GRID SV = ΣIi / d ΣP i COHERENT MULTIPURPOSE GRID SV = 4 ΣIi / z ΣP i Intercept length and grain size • Surfaces within real specimens can have very large amounts of area occupying a relatively small volume. • The mean linear intercept λ of a structure is often a useful measure of the scale of that structure, and is related to the surface-to-volume ratio of the features, since λ = 4VV / SV= 4V / S Chandreshakar (1943) showed that for a random distribution of stars in space • • 18 The mean nearest neighbor distance is L = 0.554•Nv-1/3 where Nv is the number of points (stars) per unit volume. For small features on a 2D plane the similar relationship is L = 0.5•NA-1/2 where NA is the number per unit area. Normalized count: PA= 13 counts / 1600 µm2 = 0.0081 counts/µm2 Geometric property LV = 2 P A 2 • 0.0081 = 0.016 ( µm/µm3 ) = 16 ( km/m3 ) Disector NA = NV (D + t) t = slice width A = Lookup plane area NA2 – NA1 = NV (t 2 – t1) (NA 2 – NA 1) A = Nv (t 2 – t1) A NV = N / V (slice) 19 Increased Islet Volume but Unchanged Islet Number in ob/ob Mice. Bock T, Pakkenberg B, Buschard K. Diabetes. 52:1716-22, 2003. The figure illustrates the sampling of histological sections. The pancreas was exhaustively sectioned, and both the sections marked as primary and the reference sections were used in the stereological investigations. Each sampled primary section was systematically investigated as illustrated in the top of the figure. This was performed in two sessions during which either the counting frame (lower left in the figure) was attached to the table to count the islets as described in the text, or the point-counting grid (lower right in the figure) was attached to the table to count the total volume of islets, the total volume of pancreas, and the volume-weighted mean islet volume In the primary section, islets denoted B, C, and D are defined as within the counting frame because they do not touch the (full) exclusion lines, and they are either completely within the frame or touch the (dashed) inclusion lines. The islets denoted A, E, and F are defined as being outside the frame because they are either completely outside the frame or touch the exclusion lines. In the reference section shown at the bottom of the figure, islets B and D are still present, and therefore they are not counted, whereas islet C is not present, and therefore it is counted 20 Disector - NV of convex features in three dimensional space. This count: N = 6 Relationship: NV = N / V0 Geometric property: NV = 6 / 32 µm3 = 0.19 µm –3 = 1.9x1011 per cm 3 Why Start with 100-200 Counts and 10-20 Sections? • The 100-200 counts stems from the fact that the sampling with disector probes by itself must be viewed as independent random sampling. • At the moment, there is no mathematical basis for calculating the variance of a two or three dimensional systematic random sample and one must resort to the formula for independent sampling. Thus, the basis for suggesting that 10-20 sections be used is primarily empirical. Bonaventura Francesco Cavalieri Born: 1598 in Milan, Duchy of Milan, Habsburg Empire (now Italy) Died: 30 Nov 1647 in Bologna, Papal States (now Italy) • Bonaventura Cavalieri joined the religious order Jesuati in Milan in 1615 while he was still a boy. • In 1616 he was transferred to the Jesuit monastery in Pisa. His interest in mathematics was stimulated by Euclid's works and after meeting Galileo, considered himself a disciple of the astronomer. • The meeting with Galileo was set up by Cardinal Federico Borromeo who saw clearly the genius in Cavalieri while he was at the monastery in Milan. 21 • In Pisa, Cavalieri was taught mathematics by Benedetto Castelli, a lecturer in mathematics at the University of Pisa. He taught Cavalieri geometry and he showed such promise that Cavalieri sometimes took over Castelli's lectures at the university. • Cavalieri applied for the chair of mathematics in Bologna in 1619 but was not successful since he was considered too young for a position of this seniority. He also failed to get the chair of mathematics at Pisa when Castelli left for Rome. • In 1621 Cavalieri became a deacon and assistant to Cardinal Federico Borromeo at the monastery in Milan. He taught theology there until 1623 when he became prior of St Peter's at Lodi. Aft er three years at Lodi he went to the Jesuit monastery in Parma, where he was to spend another three years. • In 1629 Cavalieri was appointed to the chair of mathematics at Bologna but by this time he had already developed a method of indivisibles which became a factor in the development of the integral calculus. Cavalieri’s estimator The honorary application of Cavalieri’s name to the estimator has led to some confusion, at least linguistically, with regard to his contribution to the stereologic method, as manifest in usages such as Cavalieri’s method and Cavalieri’s principle. The theorems of Cavalieri deal with ratios of volumes of solids based on all the planes, the indivisibles, through the solid, not a finite number of sections of known thickness (divisibles). The use of slices to estimate the volumes of regular solids was in use long before Cavalieri’s work. The major virtue of Cavalieri’s analysis, as Sterio points out, was the generalization of the use of indivisibles to the ratios of volumes of any two solids . Thus it seems best to designate the stereologic method as the Cavalieri estimator, thereby honoring him without incorrectly attributing the method or the principle to him. This estimator of volume, named in honor of Cavalieri, is the product of the sum of the areas of parallel sections cut through a particle and the thickness of the individual sections, ( ΣAi )*∆x, where Ai is the profile area of the i-th section and ∆x is the mean section thickness. 22 Threedimensional surface reconstruction of a spheroid test object. Duerstock et al. J. of Microscopy 210 , 13 8-148, 2003. A computer generated plot showing the number of particles required to be sectioned to reach a 95% confidence interval of ±5% of the mean volume for varying dimensions of oblate ellipsoids, varying standard deviations of the volumes in the parent populations and varying ratios of the section thickness to the major axis. 200 mean volume = 1. ± . H HF H H 160 B 1:2:2 J 1:4:4 1:6:6 1:8:8 1:10:10 H 120 H P 8 0 Number of measured particles to re ach criterion 4 0 H F H F FJ START F J J F F F H HH HJJ JJ H H H J J H F H J H J J JJ J J J J H H HH F FF H H H H H J F HH J JJJ H F HHH F J JJ J H FH F F H J J H FHH J J F JHFF H HHHH J FJHH J JJJJJ JF H H H JFFF HHHFJJJ HF JJ J J JJ BBB H BJFHBJJ F F B F J JHHFH FF HH J HH FHHH JJJ FF J H H JJ H HJ J J JJJJJJ H FHFFF BBH HHH JHHH BJJJ BB FHHHH BBB F JJ H HHH JHJJJJJ F FHFH JJ H JJ J H J JJ F F F 0 0 200 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 RANDOM NUMBER GENERATOR GENERA TE AT RAN DOM A SPHEROID HH H H F F RANDO M NUMBER GEN ERA TOR CHOOSE A CUTTING PLANE 0.5 mean volume = 1. ± . F H F P H J HF H P P P H H F P H J H H PP F P H H JJ H J F P P P P P H P J FP PPP FH H F H F F J P PF F PP J J PF P F J P P F F PP P P P PP H B F P PF H H J J J P HP P HF J P P P P PF F F H J B J PH JP H BP P F HP P P F B J B BHJ B BB B J P HFFP PP B FJ B PP H B F HJJ P H F F F F B JP P P HFJF P P FFPPP JJF H H J P B H H J J JP H P H J P JJ FF BP HBPP HJPP H BP HBH P J B J P J B J B J P PPP P P FP B PBFP B J HH P PH HBP F B H PJBPHP B H F FF J JJ B B B P HF HJ B P F B F P P 160 P 120 8 0 SLICE SYSTEMATICALL Y THE SPHEROID 4 0 CALCULATE THE SPHEROI D VOLUM E 0 0 200 160 120 8 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 mean volume = 1. ± . JH F P P P F FFFP F F P P P H HHFP FH HH P H HH J B F P P H P H P J F H F FJ J P P F F P P J P H J J F F P F H P P B J F P P J J F J P P J JH J H H J J P H H H J H J F P FP H P P H FF FP P FF JH F J BBP B FJF B J J F HF JF J H F JH J J P H H JFP P JHF H F J HP H P H J J J JB BF F P H FF B FH B PF H H P P PF JJ P P JF H H J BP B JB P FF P P J J H JJ HF J J P J HH F J JJ H B BP H B B P P B H P B FJ F H F P H FJ J H PF JP FJP B FFP P H H JP J JJ H F BPP H J F PP F JF HH P FF J P F H F HHH F P J B HF H HJBB BPF H PHPP JJ P PF P F F JHFF J F HB HP J J HH H H F J J J H BJ F F H B F F P J B F HJ J H F J P B H 0.5 J J YES REPEAT ? NO 4 0 GOTO STATISTICS AND GRAPHICS SUBRO UTINES 0 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 Section thickness / Major axis SERIAL S E C T I O N I N G (e.g. MRI, CONFOCAL) ALONG EITHER MAJOR AXIS OR MINOR AXIS Ellipsoid, axial ratio 1 : 3 : 3 minor axis = 0.596 unit ength l major axis = 1.79 unit length 0.5 0.5 MIN OR A XIS INTERVAL CONFIDENCE DO Section thicknes = 0.2-0.4 Along the major Axis 0.4 0.3 0.3 0.2 95% NOT RECOMMENDED Many serial sections MAJOR AXIS 0.4 . ± . . ± . . ± . . ± . 0.2 0.1 0.1 0 0 0 50 100 150 200 0 50 100 150 200 100 150 200 Ellipsoid, axial ratio 1 : 1 : 3 minor axis = 0.860 unit ength l major axis = 2.58 unit length 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 50 100 150 200 NUMBER OF 23 SECTI ON TH C I K N ES S 0 50 SAMPLED PARTICLES Nucleator and Star Volume vn* = 4π/3 li3 Correlation between mitochondria volume fraction and star volume during pancreatic acinar cell regranulation 24 Surface-weighted star volume Gittes, F. (1990) Estimating mean particle volume and number from random sections by sampling profile boundaries. J. Microsc . 158: 1-18. v s* = 2π/3 li3 CV = σ/ vN = (vV / vN –1)1/2 Evaluation of the measure of disperssion (CV = σ/vN) demonstrate that even with large number of measurements the precision of CV estimation is poor. Number of sampled profiles AUTORADIOGRAPHY 3-5 minutes 25 15-30 minutes 2 hours Defective cytoplasmic granule formation. II. Differences in patterns of radiolabeling of secretory granules in beige versus normal mouse pancreatic acinar cells after [3H]-glycine administration in vivo. Hammel I, Dvorak AM, Fox P, Shimoni E and Galli SJ. Cell Tissue Res. 1998; 293:445-452. Data can be presented as Silver Grain Density or Percent of Labeling Correlation of size with “age” 26 Brain map Densitometry Autoradiography We are now at the end of our journey, and the two ideas worth repeating are: 1. $200 is a lot cheaper than $100,000. A little education can save a lot. 2. I am very proud to be part of that rapidly growing branch of microscopy known as “quantitative microscopy". It is the future of structural biology. 27