IEEE THan&acZLom on NacZe^oA Sciznce., Vo£.

WS-27, Jane 1974

A Tutorial on ART
(Algebraic Reconstruction Techniques)

Richard Gordon
Mathematical Research Branch
National Institute of Arthritis, Metabolism and Digestive Diseases
Bethesda, Maryland 20014, USA

I. Introduction

Algebredc Reconstruction Techniques (ART) were introduced by Gordon, Bender & Herman (19T0) for solving the
problem of three dimensional reconstruction from projections in electron microscopy and radiology. This is a
decenvolution prcbl^ of a particular type: azi estimate of a function in a higher dimensional space is decon-
volved from its experimentally measured projections to a lower dimensional space. For instance, an x-ray photo-
graph represents the projection of the three-dimensional distribution of x-ray densities within the body onto a
two-dimensional plane. A finite number of such photographs taken at different angles allows us to reconstruct
an estimate of the original 3-D densities. ('Density refers to optical density.)
The ART algorithms for solving this problem have a simple intuitive basis. £ach projected density is thrown
back across the higher dimensional region frcm ^ence it came, with repeated corrections to bring each projection
of the estimate into agreement with the corresponding measured projection.
In order to discuss the ART algorithms, one must first carefully consider the representation of space in
dig!tea cosnputers (Sections II and III). SuS^sequent sections wIjJl survey the original ART algorithm (Section
IV), convergence criteria (Section V), variations on the ART algorithm (Section VI), reliability of reccmstruc-
tions (Section VII) and computing efficiency (Section VIII) ·
All symbols used are sunmarized in Table 1.

II· The Representation of Space

In digital cossputers it is necessary to represent continuous space in a discrete fcushion. Many truncated
basis sets have been used to represent the higher dimensional space in reconstruction problems (Gordon & Herman,
197^)· However, the ART algorithms €tre formulated in terns of a particular kind of basis set: one which divides
the reconstruction space into a finite number of nonoverlapping elements or subregions. The unknown density dis-
tribution is approQdmated by the values assigned to each element by the reconstruction algorithm.
Let us assume that the unknown density function is identically zero outside a finite region $1. Let the
reconstruction space Ä be divided into η nonoverlapping elements (Figure l). Ideally we might want this division
to be as fine as possible. The minimum fineness of the division offtis linked to oiur computer representation of
the projections to the lower dimensional space.
Let us assume that each projection, which will be of finite extent, is also divided into nonoverlapping
elements. The maximum size of the projection elements should be dictated by the presumed spatial resolution in
the projection. (This resolution is determined by the physics of the radiation tised.) It is ccxonon practice to
use a spacing between elements which is half of the presumed resolution. (Finer division may be warranted if one
is attempting to achieve superresolution by deconvoluting the spread functicm.)
Let j=l,...,m, represent all the projection elements of the available projections taken together. For
each projected element 9^ there is a corresponding subregion g^ in ñ of which is the projection. The exact
shape of depends on the paths of the radiation through ft. will be referred to as the passage for radiation
falling on .
Let r represent a point inftand f(r) our unknown density function. Then

f f(i) dr « p^ j-l,...,m (1)

^^j
where pj is the e3φeΓimβntal measiurement of the J*^ projection element of f(?). The i^proximation sign (pt)

78
Figure 1. Illustration of the geometiy of recon­
struction trcm projections. Ihe large square is a
two-dimensional reconstruction space ft. It is
divided into η small squares or reconstruction
elements fl^, l-l, · · · ^m, each of ifhich is to be
assigned a value f^. The cezitrolds are indicated

ñ
by dots. Radiation i J B p l n g e s on the film or detector 1 2
•
at projection element giving the measurement p^. m

yÁ
This radiation has traversed the ctoject throu^ the • •
passage which is shaded. In this particular
case the radiation travels in a parallel ray at •
angle θ and the ray width i s chosen s c that one
/Λ • •
0
centrold is encountered per row of elements 9,^ by
η-2 n-1
gj, except for the last row. If we were to dis­ 1
regard this edge effect, then C j would be identi­
cally 1. The shaded parts of the small squares
represent Π $%^· Their areas are the w^^.
For a given squareft^,u^^ = 1 if its centrold r^
is in the shaded region g^, » O i f not.

FILM

indicates that the measurement process is not perfect. Equations 1 are the fundamental equations from which all
reconstruction methods for detezmlning f (r) begin.
The passage through Β has no necessazy geometric relationship to the elements of ft. Let be the 1^^
element of ft. Then we may define the region Sj Πft^^as the intersection of passage with the reconstruction
element ft^. This allows us to rewrite Equation 1 as
η
> ff(í)aí (2)

Our goal is to obtain an approxlnatlon for the tinknown function f (r) by assigning an estimate f^ of its
value to each region 9.^. The best estimate would result when is the average value of f (?) over the subregion

f ( r ) dr

(3)
•'ft

This ideal outcome cannot be attained due to limitations on the amount and quality of data as well as the recon­
struction algorithms themselves.
Since we are ignorant of how the function f(?) varies within the elementft^,we do not know its value in
Πftj^»However, we may make the assxaiption that if f^^ is the average value of f(r) overft^then the integral
ÜX r ( $ ) over Sj Π ft¿ xnay be estimated by the geometric fraction

79
 ί, dr

mul-tiplied by t^i
w iel,...,n; >1,·.·,ιη (5)

this further approxlmation. Equations 2 beccme a set of simultaneous linear equations in the unknowns f^:

η
NT*
Ρj « > Wj^^ f^ á»l,...,m (6)

Although these look like an ordinary set of linear equations, they sure distinguished by a number of features:
1) The matrix {w^^} is quite sparse, since from the geometry of projections 0 =» 0, (the mull set),
for most pairs (i,d)- (That is, osly a relatively few of the are nonzero. Each passage Sj
encounters only a relatively few of the R^.)
2) The size of the matrix {w. .} can be enormous. In typical applications η starts at 2500 and with little
*J 6 q
aznbition η can easily reach 10 . In some cases η » 10^. The number of projection elements m ranges
from 500 to 10^ and 10*^ in corresponding cases. Thus the matrix size nm ranges frcm 750,000 to 10^"^ or

3) The equations are ordinarily highly underdetermined, i.e., m < < n.

If) The rank of the matrix {v^j} is unknown.
5) The matrix {w^j} is nonnegative since w^^ ^ 0.
6) The data values p^ are ordinarily nonnegative.
7) The unknown function f (r) is ordinarily assumed to be nonnegative, so that one desires a solution for
which f^ ^ 0·
8) The errors in the data may cause the equations to be inconsistent.
9) Statistical effects of noise in the data should be analyzed, if possible.
10) The approximations l>y i^ilch we arrive at Equations 6 introduce systematic errors which have yet to be
analyzed.

III. Efficient Representation of Space

Because of the enormity of the matrix {w^^] we must go to great lengths to reduce the size of the reconstruc­
tion problem.
In the interest of retaining detail in o\u? estimate of the unknown density f (r), the division offtshould
be made as fine as possible. On the other hand, the size of the matrix {v^j} directly depends on the nuniber of
elements into whichftis divided, so that the division offtshould be as coarse as possible. The size of the
projection elements, however, limits the coarseness of the division of ft. How this limit comes in may best be
seen by considering a further approximation.
Let r^ be the centroid of the element ft,. Assume eachft^is convex, so that r^ € A¿ (where the symbol ^
means "is an element of the set of points"). Define the matrix

u . ^1
1 0 if not
if not

Under certain geometric conditions, discussed below, it is reasonable to approximate w^^j f^ by u^^ f^^, so that

Σ "id ^i d-l,...,in (8)

1=X

80
 For each passage Sj we have succeeded in approxixaating the integral of f(r) in Equation 1

r f(?) dj (9)

by the sum of the estimates f^ at a finite set of points within gj. For a given passage g^ the number of points
it intersects ^

^5 " Σ J-l,...,in (10)

1»1
must be nonzero. l!his limits the coarseness of the division of ft.
The shape of the passages gj does not enter into the formulation of the reconstruction problem in teims of
Equations 1. However, in the case of projections obtained from radiation transmitted through the regionft,we
may call each passage a ra^^ since it goes more or less in a straight line. Ifftis a convex region of
d-dimensional space (divided into η eloaentsft^all of the same size), then the mean distance acrossftwill be
approximately n^^ ^ (where the unit of len^h is the linear dimension of eachft.).The centroide r. within S .
1 1 d
should be distributed €ú.ong its length in order to accurately represent the density distribution along . For
those rays Sj which roughly coincide with a diameter offt,we will then require > n^/^, which sets a more
stringent requirement on the coarseness of the division of ft.
The centroide contained in the passage S^. are st^posed to represent the integral of f (r) within g^:
η η
p^ « J f(r) dr = y J f(r) dr « V w^^ f^ J.l,...,m (ll)
*j i»l *j ^ ^1 1^1
However, the \ise of the matrix [u^jj aaans that the centroide actually represent

7 ^ij ^1 - y ¡ ^í^) ^ j=l,...m (12)

(The symbol ^ means "such that".) In general the stm of integrals in Eqixation 12 will differ from the sum in
Equation 1 1 . We may formulate an approximate geometric correction c^ for this discrepancy

L J f(?) ai « y J d? ϋ=.1,...,ιη (13)

ΣΗ
irtiere η

i9r.€§á
Σ I dr

i^ch siii^lifies to

J.a?
" d=.l,...,Bi (15)

dr
«1
provided that all elementsft^^have the same size. From Equations 1 1 , 12 and 13 it follows that:
η

" Σ ^id á-i.-.ni (16)

i=»l

81
 This is equivalent to the geometric approximation that

i=l,...,m; d-1.
\i «d «id (17)

imich can only be true on the average.

It is sometimes possible to choose the projection and reconstruction elements in such a way that c^ is a
constant Independent of j. For example, consider the reconstruction spaceftto be a two-dimensional square
divided into η equal squaresft^,1=»ΐ,,,.η, «u-ra^ed on a grid. A projection of parallel rays at angle Θ (for
which the are strips acrossft),project onto a one-dimensional line (Figure l). Choose a ray width

r^ = min( |slne|, |cose| ) (l8)

In this case the strip g^ intersects one centrold r^ per row of elements of ft (except for edge effects). The
area of Sj per row of elements offtis equal to the area of one elementft^^.Thus In Equation 15 we find c^ β 1 .
This justifies the use of Equation 8.
When the geometry requires that c^ not be a constant, we may nevertheless preprocess the data so that

Ρ ^
'd "5*^ « y \j j»l,...,m (19)
^ 1=1

Thus it is not necessary to \ise the weights w^^ with, for example, divergent rays, for ertlich c^ depends on j.
We may rewrite Equation 19 as

J-1,. (20)

IV. The Original ART Algorithm

Gordon, Bender & Herman (1970) formulated the ART algorithm in terms of Equation 20, under the implicit
assunptlon that c^ s l. (The justification, using Equation ΐθ, was found later.) Since the algorithm is itera­
tive, let fj designate the q*^ estimate for f^. Let

jel,...,m (21)

th
be the value of the projection of the q**" estimate. Let f be the mean density of the object in ft. Then the ART
algorithm is

(22)

(23)
j j

We may express this in words as follows. The iterative process is started with all reconstruction elements set
to a constant (f^a?). In each iteration the difference between the actual data for a projection element and the

sum of the reconstruction elements representing it (Equation 2 1 ) is calculated ( τ**· - ~ ) · The correction is

evenly divided amongst the ( N ^ reconstniction elements ( i a r ^ ^ S j ) and added to them. If the correction is
negative, it may happen that the calculated density for a reconstruction element becomes negative, in which case
it is set to zero (max operator, guaranteeing fj ^ O). Each projection element is considered in turn

82
(jel-Hnod^q). The calculation is repeated a number of cycles (K) for the whole set of projection elements until
reasonaPble convergence it attained.
The mean density of the object is

i may 1>e eetlmated from a single projection Fj^ (such as a single x-ray photograph) provided the in projection
Pj^ partition the x'econstructlon space At

rS. If j-j»
R » Π and g. ñ»., - - for all j,j'3S.,S.,6Pfc (25)
j^S,€PÍ J ^ l0 ifnot i i ^

({Sj€Pj^l partitions ft if evezy point ϊη^ is in exactly one of the g^.) Then

3 f^ (26)
' d? Γ d?
a -"β
The approximation may be improved by averaging over the estimates frcm each Ρ :

κ
f = ^ f^ (27)
fc=l

where κ is the number of projections.

The starting values f^ are often chosen to be identically zero, instead of ?. If we order the projection
elements such that all of those from one projection come before those of the next, then f^ will be the same
in either case after the index j has gone throufiji the first projection P^^.
Alternatively, one may use a rough algorithm such as the sumnation method (reviewed iu Gordon & Herman,
197^) to produce starting values:

j 3 SjH it^'^

If each projection P^^,tesl,...,ACforms a partition offt,then there are exactly κ terms in each summation. The
advantage of such an initial estimate is that the sequence in q, {fj, should converge more rapidly when {fj} is
near the final result. However, with underdetermined Equations 20, any distortions introduced by such an initial
estimate may be retained. (The sumnation or "back projection" method often produces "ghosts". See Crowther,
DeRosier & íQug, 19T0, and Bellman, Bender, Gordon & Rowe, 1971.)

V. Convergence Criteria

It is necessary to determine when an iterative algorithm has converged to a solution ^ich is optimal
according to some criterion. Various criteria for convergence have been devised.
Gordon, Bender & Herman (1970) proposed three measures for the convergence of the f the discrepancy
between the measured and calculated projection elements

83
 (29)

3-1 *' *'

the nonuQlfomlty or variancg η

and the entropy

V« = y ( t ¡ . f)2 (50)

α Σ ( ^ ) ' » ( ^ ) ( «
tends to zero, to a minimum and to a minimum with Increasing q.
A test pattern is a known function f^(r) which is digitized over the reconstruction elementsft.as f f?].
th 1 *· Ι·*
^q ^ Γ /^α .t.2
The Euclidean distance between the test pattern and (tl q- qfestimate is
its (32)

Gilbert (1972a) showed that by the criterion ö^//V^ ART begins to converge with Increasing q but then
diverges provided Equations 8 are inconsistent. The absolute values Gilbert obtained for this criterion were
greater than necessary, because he did not use the variable ray widths of Equation l 8 , which I formulated and
were later spelled ^^xt in Herman & Rowland ( 1 9 7 1 ) · The divergence of ART with Inconsistent data is nevertheless
real. Bierefore, the computation should be stopped before divergence begins. Herman, Lent & Rowland ( 1 9 7 3 )
ezoplrically foxmd that the minimum value of 5^//V^ coincided within one iteration with the stopping criterion

|V^*^.V^| < V^/lOO (33)

These difficulties with inconsistent data have led to some of the variations of the ART algorithm.

VI. Variations on ART

A. Generalized ART
We may speak of a generalized ART algorithm A as any iterative function which finds new values for the
reconstruction elements intersecting a passage fraa their old values:

f f^ = A(fJ, {f^, 3 ίΐ^,π ^^m^ Pj) 15 ft^n m

The sum of the new reconstruction elements should be closer to the value of the projection element:

|pf^-Pj| ^ |pJ-Pj| j=l4mod^q; q:K),...,Βη-Ι (55)

In general, A will have an eacplicit dependence on the fJ under consideration as indicated in Equation JU.
For exaogple, Gordon, Bender & Herman ( 1 9 7 0 ) suggest a series
00

f f^ = maxf 0 , 7 A/f?)' 1 (36)

i«0
where the coefficients A^ are constrained by the relation
09

^ ^0 * ^ * X A, iff)' » (57)

8**
Only one of the is detemined by Equation 37· ART with c^ a 1 is the special case

A^^ » 1 (38)
A^ « 0 for i ^ 2

Β· Multiplicative ART
Multiplicative ART (Gordon, Bender & Herman, 19T0) is specified by

i
irtilch corresponds to
= O

A^ - 0 for i ^ 2

in Equation 3^· (ihe saiae algorithm was found independently by Schmidlin, 1972.)
The choice between additiv» ART (Equation 23) and multiplicative ART depends on the physics of the radiation
used. For transmitted radiation, the form of the reconstructed object should be independent of an additive con­
stant. S\ich a constant may result frcm variable e^osure in an x-ray, variable development of the film, or an
intervening filter, such as the amorphous material increasingly deposited on a sample by the electron beam in
electron microscopy. Except for the nonllnearity of the max operator, independence from an additive constant is
accooiplished ty Equation 23, additive ART. However, in nuclear medicine emitted radiation is measured. The
count may change rapidly with short-lived radioactive elements but the proportions troa one region to another
determine the form. In this case a multiplicative algorithm, such as Equation 39, will preserve the fozm inde­
pendent of tiro variation between projections. (It may, of course, be possible to slowly normalize the project
tions In some cases before reconstruction.)
Multiplicative ART must be started with some f^ > 0. It has an advantage over additive ART that once
f^ 3 0, then it remains so in all subsequent iterations. Clearly all f^ ^ ^l^^j ^ ^ should, in general, be
set to zero and remain so vhen ρ^ s 0 (cf. Budinger & Gullberg, 197^). These reconstruction elements demark the
subregion of ñ containing the object being reconstructed.

C. Unconstrained ART
In order to have a version of ART which can be analyzed by the methods of linecu: algebra, Herman, Lent &
Rowland (1973) Introduced unconstrained ART for which

irtiere Nj is redefined by η
Nj « ^ w^j J-l,...,m (if2)
ial
When the matrix {v^} is replaced by {u^} using Equation 17, Equation k2 becomes equivalent to Equation 10.
equation kl is the same cus the iterative matrix inversion method of Kacmarz (1937)·
Unconstrained ART has been proven to converge to the unique solution {f^} minimizing the variance
η
V - ^ (fj-f)2 (1.5)

itbea Eqnatlooe 6 are consietenfc. On data inconsistent with Equations 6 unconstrained ABI converges cyclically.
i.e., the will be the same after each cycle, for large K.

85
 The algorithm tised in the EMI-Scanner (Hounsfield, 1972) is equivalent to unconstrained ART though the
geometric correction factor Cj is calculated in a manner different from Equation 15*
Unconstrained ABT, of course, permits soltxtions with some f^ < 0, i.e., negative densities.

D. Constrained ART
Fully constrained AI^ is defined by

f f ^ = min[ F, max[0, fj + ^ijíPj-Pj)/»^! ] W

resulting In 0 ¿ f^ i F. This constraint may be useful when the object being reconstructed is known to have
densities less than F. Herman, Lent & Rowland (1973) chose F = 1 which precludes solutions in integers.
Partially constrained ART (another name for ART, Equation 23) and fully constrained ART have been proven to
converge to a solution of Equations 6 vñien the equations are consistent. However, the solution is not naceseariiy
the one of minimum variance (Equation U3). On Inconsistent data, partially and fully constrained ART are con-
jectured to converge cyclically (Herman, Lent & Rowland, 1973)·

£· ARTS
ART2 (Herman, Lent & Rowland, 1973) uses an Inteimedlate estimator f J

with
f f^ » max(0, ψ^) (kS)

in the partially constrained version or

ff^ = mint F, max(0, tf^) ] (hj)

in fully constrained ART2. p^ is still defined in terms of the f n o t f^:

η

i«l
The inteimedlate estimator may become negative at one Iteration and return to zero or a positive value at a
later Iteration. This is an inq^rovement over ART for which a positive correction always leads to a positive f^.
ART2 and fully constredned ART2 have been shown to converge to the solution of minimum variance (Equation
k3) whai Eolations 6 are ccmsistent.

E. ART3
ART3 (Herman, 197^) vas designed to find a solution of Equations 6 within a preset error tolerance which may
vary from one projection element to another:

X
η
^r^J ^ ''ij ^i ^ V^j (^9)
1=1
The algorithm is (Johnson et al., 1973)

where
if |P3-pJ|>2€j (Sla)
γ. 2(pj-pJ.€j) if €j<Pj-pJ<aej (51b)
^ ] 2(pj-pJ^€j) if € 3 > P j - p J » -eCj (51C)
if |p¿-pj|*€j (51d)

86
EqufitioQ 6 nay be thought of as defining a hyperplane In n-diinenslonal space and Inguallty k9 a *hyperslab*.
Condition 31a results in a projection of the n-dimensional point {f^} onto the hyperplane of Equation 6 and is
equivalent to unconstrained ABT (Equation kl) · A point vithin a distance of the hyperslab is reflected about
the nearest face of the hyperslab by Conditions 51b or ^Ic. If point {f^} is already vithin the hyperslab, no
correction is made (Condition 31d).
ABT3 nay be shown to converge to a solution of Inequality k9 in a finite nunber of iterations, provided a
solution exists*

Q* ART with Binary Constraint

An ART method for binary patterns (f^ » 0 or 1, i^l, •••,n) based on ART2 and ART3 has been presented by
Herman (1973).
Gilbert (l972b) has given a method, applicable to any iterative algorithm, which yields a two-valaed {f^}.
The densities are divided into two groups after a few iterations, and each f^ is set to the mean of its group*
The procedure is repeated in subsequent iterations*
This manner of introducing a nonlinear constraint into a reconstruction may be of widespread usefulness.
As an example, the known densities of tissues might be Incorporated in x-ray reconstructions (H* Blum, personal
communication) ·

H. ART with a Pamping Factor

The introduction of a damping factor Δ as in

f f l . f?*^Vy(p^.pJ)/Kj (52)

leads only to a partial correction if 0 < Δ < 1. Sweeney & Vest (1973) fo\ind that 0.5 improved the
reconstruction. If the projection elements Sj are considered in random order within a cycle, both rapid con­
vergence and tolerance to noise (inconsistent equations) are obtained (S. A. Johnson, personal C G o m u n i c a t i o n ) ·

I. Minimizing Local Discontinuity

Kashyap & Mittal (1973) have introduced the notion that a desirable solution to Equations 6 is one minimiz­
ing the variance (Equation k3) plus a measure of the local discontinuity

Σ
η
J(a) - v + a^(f^ " " 7 " ^
^i»^ (53)
1=1 i»
where the space ñ is two-dimensional and i* ranges over the ei^t elementsft^^iclosest toft^(on a square lattice).
Such a function explicitly takes into account the fact that the {f^} represents an object whose density should
vary in a piecewise ccmtlnuous fashion. Equation 53 may be generalized for any d-dlmensional reconstruction
space R: η η

J(a) . ν + «Χ(ίΐΤ ^ii» ^1·)^

Pil

where is a matrix describing the weights to be attributed to each of the near neighbors of the reconstruc­
tion elementft^.({η^^^ι} is a sparse matrix: each row contcdns at most 3^ -> 1 nonzero elements if the are
arranged in a d-dimensional eubic lattice.)
Althou^ Kashyap & Mittal (1973) invert the resulting matrix directly (which requires an enormous amount of
CGoputing time). Equation 53 msy be incorporated into ABT-type algorithms for τβρΙά Iterative solution.
The second part of J(QE) would be minimized if

y
η
f^, =0 i«l,...,n (55)
^i - "^ii*
i^l
which may singly be regarded as η homogeneous equations of the form of Equations 6. These may be solved simul-

87
 taneouely with Bquatioos 6 by any of the ABT algorithms tolerant to inconsistent data* (Equations 55 have the
unique solution f^ e 0 if the matrix i^j^t " \±t) nonsingular, which depends on the boundary conditions.)
Equations 6 and 55 together form a system of η ·*- m overdetendned equations. Each Equation 55 geometrically
represents a "passage** confined to a local group of reconstractlom elements.
A continuum approach leads to another fozmnlation of an ABS-like algorithm. Tha limit of Equation 5^ as
the subdivision offtgets finer:

J(a) - f [f(r).fj^ dr + α Γ í^^f (r)]^ dr (56)

''ft "^ft

transfoxms the reconstruction problem into a problem in the calculus of variations (Gk>RDQN & Kashyagp, 197^)·
This opens the reconstruction problem to solution by numerical methods for partial differential equations.
For example, the matrix {^j^^j^t} i-s equivalent to the finite difference schemes used to solve partial dif­
ferential equations. If the general ABT algorithm of Equation is restricted to

f f ^ « fj + A({fJ,}, pj) i 5 Ä^n ?i 0 (57)

we can see that it represents the Solerían numerical approximation to the partial differential equation

. A(f(?,q), fpj]) (58)

Appropriate forms for A are given by Gordon & Kashyap (1974).

J. Monte Carlo Algorithm

The first Monte Carlo algorithm of Gordon & Herman (l971a,b) solved Equations 20 in integers. For each
projection element corrections were made by adding or subtracting 1 at random to the 3 Π ^ 0 with
the constraint that f^ 2^ 0. The statistical mean of this process is equivalent to ART. Smooth reconstructions
were obtained by averaging a number of Monte Carlo solutione. The lijnit of such averaging yields the same
reconstruction as ART. (Frank, 1973, mislabels ART as a Monte Carlo method.)

VII. The Reliability of ART Reconstructions

Resolution criteria and the performance of reconstruction algorithms according to these criteria have been
reviewed by Gordon & Hersaan (157^). In general, these criteria produce a single number, such as the δ of
Equation 32, whose relationship to the visual quality of the reccastruction is only heuristic. Thus, I will only
discuss two methods for obtaining visual measures for the reliability of reconstruction algorithms.
Equations 6 are ordinarily highly underdetermined, so that there exists an Infinity of solutions. luay
deterministic algorithm locates only one of these solutions. Ohless the noise in the data is extreme, the
original object is also amongst the solutions. If we had before us sample solutions from the whole space of
solutions, these might have seme features in common with one another. Other features might be shared only by
certain subsets of the solutions. The former will be called real features, since they must be part of the
original object. The latter features are suspect, since they may or may not be Included in the original object
(Gordon, 1973)·
It is only necessary to demonstrate two solutions, one with and one without a given feature, to label that
feature as suspect. However, without seeing saaq^les from the whole space of solutions, we can at best form a
heuristic guide to the real featuires.
Given any iterative algorithm, we may start it off with an arbitrary function {f^ and then iterate to
convergence. Each different (f^ may generate a different solirtion to Equations 6. If we had a systematic way
of generating the {f^, then we could explore the space of solutions.
Gordon (1973) proposed starting this exploration with the **ordlnary" initial value f^ =» constant, say « 0
for an additive algorithm. Let {fj^(O)} designate the reconstruction so obtained. This reconstruction will

88
 ccmtaln v a r i o u s featiires v i s i b l e t o t h e eye. Xhese features may b e removed b y s e t t i n g t h e corresponding t^(0)
w i t h i n t h e feature t o another value, say 0 o r t h e surrounding d e n s i t y . The r e s u l t i n g f u n c t i o n , {f^l, nay now b e
used ae the s t a r t i n g f u n c t i o n , t o obtain a new reconstruction £ f j ^ ( { f ^ } ) } . I f the feature i s not contained i n t h e
new reconstruction, then i t i s suspect. If it does appear, we g a i n some confidence i n i t s r e a l i t y , since i t s
reappearance was apparently forced by t h e data. This notion i s strengthened b y t h e empirical observation that
manipulations o f t h i s type made on a l o c a l r e g i o n o f a reconstruction have only minor consequences f o r other
regione. (This phenomenon occurs i n s p i t e o f t h e f a c t that t h e passages cross the r e c o n s t r u c t i o n space ft, so
t h a t regions far apart are mingled i n t h e data { p ^ } . )
Many reconstructions can accumulate from t h e various features seen i n { f j ^ ( O ) } . Three operators oxt given
f o r ccobining the solutions i n a way which attempts t o v i s u a l l y preserve t h e i r common featvures (Gtordon, 1973).
Barbleri (197^) makes deliberate use o f the nonlinear constraint f^ ^ 0 t o produce a f u n c t i o n which may
v i s u a l i z e t h e regions i n which t h e reconstruction i s u n r e l i a b l e . Let

^ - Ul^.^n h (59)

Then the complement o f [t^ i s defined as*

^1 " ^ " ^i i=»l,...,n (60)

Thus Equations 6 may be r e w r i t t e n as

η η

"Σ 1-1
''iJ '
1»1 1=1
"Σ ^ ' (
^Σ ""υ) · 1=1
(61)

I d i l c h defines the complementary p r o j e c t i o n data { p ^ } . Let { f ^ } and ( f ^ ) be solutions b y some a l g o r i t h m of

Equations 6 and the conglementary equations ^

Ρj « ^ ^ij ^1 J»l,...,m (62)

i-1
Define

m f^-^f^ 1-1, . . . , n (63)

If t h e algorithm were l i n e a r , we would ei^pect s F, 1 * 1 , . · · , η . I f i t i s not l i n e a r , {ω^} w i l l depart from

t h e constant F . Barbleri (197^) suggests t h a t those subregions o f R i n which t h e depart s i g n i f i c a n t l y frcni F
are less r e l i a i b l y reconstructed than the o t h e r s . This conjecture warrants f u r t h e r t e s t i n g and t h e o r e t i c a l
justification.

Barbleri (197^) f a r t h e r si¿sgosts a m o d i f i c a t i o n o f ART-type algorithms. A f t e r calculating b o t h f ^ and f J

i n the- ordliisry » i C T i ^ r , ene adds tf? each

- | - ( F - -f) m
irtxere
«J - tl*il (65)
I n some sense t h i s procedure provides a step b y step balance and seems t o y i e l d a cleaner r e c o n s t r u c t i o n .
Other v i s u a l c r i t e r i a f o r t h e r e l i a b i l i t y o f reconstructions w i l l come from f u t u r e p s y c h o f b y s i e a l eiqperi-
ments on t h e p r o b a b i l i t y o f detecting f e a t u r e s and d o u b l e - b l i n d e3q>eriments on t h e accuracy o f medical diagnoses
based on reconstructions.

Vni. E f f i c i e n c y i n Computing

Becanse o f the size o f the reconstruction problem, i t i s necessary t o pay s t r i c t a t t e n t i o n t o confuting

efficiency. I w i l l propose some general g u i d e l i n e s .
Equations 19 allow a substantial Increase i n confuting speed over Equations 6. Since the u ^ a r e a l l 0*s or

89
 Vs, it is not necessary to perfom the miltiplication f^. One need only detennine which reconstruction
elements have centroids lying in g^, as in Equation 20* Moreover, the data can he scaled so that Pj/^j is an
Integer. We then seek a soltttion to Equations 20 for which the f^ are all integers. (£q[uations 20 are then
Diophantlne equations.)
A single term in the suomation in Equations 19 or 20 then requires only an integer addition in a cooputer,
whereas a term of Equations 6 requires both a floating point multipllcaticm and a floating point addition.
Moreover, there are many more nonzero v ^ than nonzero u^^. Floating point multiplicatlooi or addition takes 3
to 7 times longer on cco^uters than integer addition. (Thus algorithms based on Equations 19 or 20 can be more
than 10 times faster than those based on Equations 6.
A number of other devices may be used to reduce the size of the reconstruction problem or increase its
coD^nxtaticnal speed. For instance, the nuniber of projections can be minimized consistent with the desired
* resolution' in the reconstruction space. (See Gordon, Bender & Herman, 1970, for sequences of reconstructions
with ir^reasisg naabers o f projections, and Goxdm & Herman, 197^, for a review of resolution criteria.)
Ώιο skillful use of machine language will also reduce ces^uting time. The innermost loop of an iterative
algorithm, such as ART, may be written in machine language and repeated in consecutive memory locations. This
stack of instructions is then entered at the appropriate point so that the number remaining is exactly the
n u B i b e r needed for a given projection element. A stack makes it unnecessary to increment an index, whléh can save
ifO^ coo^uting time for the ABT algorithm.
Finally, special purpose c o o q D u t e r s may be considered. For instance, the Array Transform Processor by
Raytheon can speed an iterative algorithm a hundredfold over a coc^aráble general purpose cooq^er. Optical or
electronic analog or h^rid devices mig^it achieve even greater speeds. Ihis is enphasized by the fact that the
multi-view tomograph device of Grant (1972) can achieve a full three dimensional x - r a y reconstruction by optical
means in a time cooparable to that required by the EMI-Scanner to reconstruct a few planes.
It is is^ortant to cootpare the capabilities of the JSIC-Scanner with the multi-view tomograph. The ^ d -
Scanner achieves accurate values for the densities f^, but is restricted to 50 χ 50 reconstruction elenents
across the head. (5 nsn X 3 nm for eachft^^.The latest model allows 1.5 ma χ 1.5 πη.) The 20-view tomograph has
a modizlatlon transfer function (iflCF) ^Ailch falls off to 10^ between 1.3 and 6.2 cycles/nm (Meyer, 1969, Figure
13)· This roughly corresponds to an optical resolution of 0.15 to 0.5 nm (Bibezman, 1973, Table 2.1). If we
were to consider reconstruction elements of these dimensions, we would need 700 χ 700 χ 700 to 3700 χ 3700 Χ 3700
for the equivalent spatial resolution or 3 X 10*^ to 5 X elements. Because the tomographic reconstruction
algorithm is rough (Gordon & Herman, 197^), the individual f^ are inaccurate. There is in general a trade-off
between spatial and density resolution; however^ the optimum balance has not yet been schieved.

IX· Conclusión

There are four major classes o f reconstruction algorithms: the sunmiation methods (of which clcuisical
tomography is a prime exan^e), the convolution method, the Fourier methods and series expansion methods (Gordon
& Herman, 197^)· ART algorithms bel<mg to the Xaitter category.
All ART algorithms use the same basis set, the division o f the reconstruction space into nonoverlapping
elements. All share the property of iteratively atteo^ting t o match the (weighted o r unweighted) simis o f
appropriate reconstruction elements with the corresponding projection data.
We can expect nimierous versions o f the KBT algorithms to b e developed i n the future because of their
intuitive sin^licity and wide margin for variation. Future ART algorithms will undoubtedly include versions for
use o n parallel ecoiputere and analog hardware liaplementations. We may also look forward to a more thorough
mathematical analysis o f linear ART algorithms and those idiich are linear with simple inequality constraints.

Acknowledgements

I would like t o thank Leslie Gordon, J. Z. Hearon, Helmut V. B. Hirsch and Rosalind Marimont for critical
reading o f the manuscript and Marian Caesar for careful t^setting.

90
 Table 1 . i^ynbols i subscript for projection elements (Equation l)
k subscriprt for projections (Equation 25)
A eeneralized ART-Uke reeurslve ftmction
Κ *"e?^enf S^%^uaif«^%ff^^ projection
(Equation 34)
κ number of proáectlons (Equation 2?)
i*^ tern of generalized ART-like algoidtha
i siabscript on series (Equation 36)
lehlch can be represented by a series expansion
m n u z E b e r of projection elements (Equation 1)
(Equation 36)
max max(a,b) » a i f a > b , « b i f b ^ a
AKC iU^ébraic Reconstruction Technique (s) ml η
min(a,b) a a i f a < b , = b i f b á a
α might for the local discontinuity in J(a)
mod mod^qs-remainder when q is divided by m
(Equation 53)
η number of reconstruction elements ft,
geoinetrlc correction for the representation
(Equation 2 )
of the integral over a passage g. by the sum
Ν. nuniber of centroids r^ Intersected by
of the values at the centroide r^ contained
(Equation 10) or its generalization (Equation
In the passage (Equations 13, lU-)
k2)
d dimension of the reconstruction space ft
weight for the contribution of reconstruction
d? an area or volume element of the reconstruct
elementft,,toft,in J(a) (Equation 5^)
-blon spaceft(if dpi2, dr Β dxdy; if d » 3,
0 the empty set
dr a dzdydz; etc.. Equation 1)
density measurement of the .th
J projection
mean discrepancy between measured projection
element (Equation l)
elements p^ and their estimate at the q*^
geometrically corrected density of the J th
iteration, pj (Equation 29)
projection element (Eolation 19)
6« Euclidean distance between a test pattern
complement of p^ (Equation 61)
£f^ and its reconstruction at the q*^
iteratic»! (Equation 32) density of the J*^ projection of {f^
« 1 if i = i«, =» O if not (Equation 21)
"11·
Δ damping factor (Equation 52) the set of passages g. corresponding to the
th ^
••is an element of" k projection (Equation 25)
€
error margin for the J th projection elenent projection element
(Equation k9) q iteration counter, incremented each time an
••such that" iterative algorithm goes from one projection
f unknown density function (Equation 1 ) element to the next (Equation 21)
f mean of the unknown density function f over a point in the reconstruction space ft
the reconstruction spaceft(Equation 26) (Equation 1)
estimate for the density function f in the centrold of reconstruction element ft,
reconstruction elementft.(Equation 3) (Equation 7)
*fch
"θ width of a parallel ray at angle θ (Equation
q estimate for f^ (Equation 21) OS)
ft reconstruction space (the domain of f)
inteimedlate estimator for fJ (Equation h3) ft,1 ^ ^ reconstruction element (Equation 2)
ccngolement of f^ (Equation 6o) entropy of {f^ (Equation 31)
ii^ passage throughfttraversed by radiation
^J
conqpleoent of f^ (Equation 65) reaching projection element (Equation 1)
known density function or test pattern angle (Equation l8)
(Equation 32) indicates intersection of centrold r, with
""a
overage of f* over the reconstruction element passage gj (Equation 7 )
(Equation 32) V variance of {f,} (Equation h3)
maximum density (Equation hk) variance of {:^} (Equation 30)
correction factor for ART3 (Equations 50, 51)
fraction of reconstruction elementft,inter­
''id
subscript for reconstruction elements sected by passage (Equation h)
(Equation 2)
ω, sum of f, and f, (Equation 63)

91
 aym of fJ and fJ (Equation 65) R. Gordon (1974). A comprehensive bibliography on
reconstruction from projections. Methods of
t} Information in Medicine, in press.

η "intersection of" H. Gordon & R. Bender (l97l) · Hew three-dimensional

Algebraic Reconstruction Techniques (ART). Proc. 29th
Annual Meeting Electron Microscopy Society of America,
Laplaclan, + In two dlmenslone, Boston, 9-13 August, 82-83.
hx by
etc. (Equation 56) R. Gordon, R. Benders G. T. Herman (1970). Algebraic
reconstruction techniques (ART) for three-dimensional
electron microscopy and x-ray ^otography. J. Theor.
Biol. 22, 471-481.
References and Bibliography on ART-llke Algorithms R. Gordon & G. T. Herman (l971a). Reconstruction of
pictures from their projections. Quarterly Bulletin
of the Center for Theoretical Biology, State tbilver-
Μ· Barbleri (197^). A criterion to evaluate three sity of Hew York at Buffalo hjil), 71Ί51*
dimensional reconstructions from projections of
unlmown structures. J. Theor. Biol., in press. R. Gordon & G. T. Herman (1971b)· Reconstruction of
pictures trom their projections. Comm. A.C.M. 14(12)
S. Η· Bellman, R. Bender, R. Gordon & J. E. Rowe, Jr. 759-768. (Figs. 6 and 7 are interchanged.)
(1971) · ART is science, being a defense of algebraic
reconstruction techniques for three-dimensional R. Gordon & G. T. Herman (1974). Three dimensional
electron microscopy. J. Theor. Biol. ^ 205-216. reconstruction frcm projections: A review of
algorithms. Int. Rev. Cytol. 38, in press.
R. Bender, S. H. Bellman & R. Gordon (1970). ART and
the ribosome: a preliminary report on the three- R. Gordon & R. L. Kashyc^ (1974). Variational formu­
dimensional structure of individual ribosomes deter­ lations of the problem of reconstruction from
mined by an algebrcdc reconstniction technique. projections, in preparation.
J. Theor. Biol. 22, 483-487.
R« Gordon, J. E. Rowe, Jr. & R. Bender (1971). ART:
L. M. Blberman (1973)· Perception of Displayed a possible replacement for x-ray crystallography at
Information. Plenum Press, New York and London. moderate resolution. First European Biophysics
Congress, Baden near Vienna, September 14-17,
T. F. Budinger & G. T. Gullberg (1974). Three- Ε XVIIl/4, 4i*l-445.
dimensional reconstruction in nuclear medicine by
iterative least-squares and Fourier techniques. D. G. Grant (1972). Tomosynthesis: A three-
Lawrence Berkeley Laboratory, University of California,
LBL-2146, 133 ΡΡ· dimensional radiographic imaging technique. IEEE
Trans. Bio-med. Eng. BM19(l)^ 20-28.
Correspondent (1971)· The ART of the possible. J. FT Greenleaf, S. A. Johnson, S. L. Lee, G. T.
Nature New Biology 2^, 131. Herman & E. M. Wood (1974). Algebraic reconstruction
of spatial distributions of acoustic absorption within
R. A. Crcwther, D. J. DeRosier & A. IQug (1970). The tissue from their two-dimensional acoustic projections.
reconstruction of a three-dimensional structure from Acoustic Holography_5, 591-603·
projections and its application to electron microscopy.
Proc. Roy. Soc. Lond. A 317, 319-340. G. Τ· Herman (1972). Two direct methods for reeonstruct-
ing pictures from their projections: a comparative
R. A. Crowther & A. Klug (197Ι)· ART and science, or, study. Computer Gn^hics and Snage Processing 1(2)^
conditions for 3-I> reconstruction from electron 123-144. "~
mieroscppe images. J. Theor. Biol. 32, 199-203.
G. T. Herman (1974). Reconstruction of binary
J. Frank (1973)· Cooiputer processing of electron patterns from a few projections. International
micrographs. In; Electron MlcroscopiccCL Techniques Computing Syn^osium, Davos, Switzerland, 4-7 Septenber
in Biology, J. Koehler, ed.. Springer Verlag, 1973^ editors A. Guenther, et al., North Holland Publ.
Heidelberg, pp. 215-274. Co., pp. 371-379·
P. Gilbert (l972a). Iterative methods for the G. T. Herman (1974). A relaxation method for recon­
reconstruction of three-dimensional objects from structing objects from noisy x-rays, submitted to
projections. J. Theor. Biol. 36, 105-117. Mathematical Progranming.
Ε P. C.Gilbert (l972b). An iterative method for three- G. T. Herman, A. Lent & S. Rowland (1973)· ART:
dimensional reconstruction from electron micrographs. Mathematics and applications. A report on the mathe­
Proc. Fifth European Congr. on Electron Microscopy, matical foundations and on the applicability to reeLL
pp. 602-603· data of the Algebraic Reconstruction Techniques.
J. Theor. Biol. ί*2(ΐ), 1-32.
R. Gordon (19Γ3)· Artifacts in reconstructions made
from a few projections. Proceedings of the First G. T. Herman & S. Rowland (1971)· Resolution in ART:
International Joint Conference on Pattern Recognition, an experimental investigation of the resolving power
Oct. 30 to Nov. 1, Washington, D.C., IEEE Computer of an algebraic picture reconstruction technique.
Society, Publications Office, 9017 Reseda Boulevard, J. Theor. Biol. 33, 213-223·
Suite 212, Northridge, CaliTorkda 91324, pp. 275-285.

92
G. Τ. Herman & 5· Rowland (l97if). Three methods for First International Joint Conference on Pattern
reconstructing objects from x-rays: a comparative Recognition, Oct* 30 to Nov* 1, Washington, B.C.,
study* Coo^uter Graphics and Image Processing, in IEEE Computer Society, Publications Office, 9017
press. Reseda Boulevard, Suite 212, Northridge, California
9132i^, pp. 286-292.
G. N* Hounsfield (19T2). A method of and apparatus
for examination of a body by radiation such as χ or R. A. Meyer (1969). 3-I> radiography - modulation
gamma radiation. The Patent Office, London, Patent transfer function. Memo MRT-1-127, h Decaaber to
Specification 1283915'· D. G. Grant, Applied Physics Laboratory, Johns Hopkins
Iftiiversity, Silver Spring, Maryland, 25 pp.
S. A. Johnson, R* A. Robb, J* P. Greenleaf, E. L*
Ritman, G. T. Herman, S. L. Lee & E. H. Wood (1973). P. Schmidlin (1972). Iterative separation of sections
Temporal and spatial χ ray projective reconstruction of in tomographic scintigrams. Nucl. Med. (Stuttg.)
a beating heart, preprint. Mayo Clinic, Rochester, 11, 1-16.
Minnesota*
P. R. Smith, T. M. Peters & R. H. T. Bates (1973).
S* A. Johnson, R* A. Robb, J* F. Greenleaf, E. L* Image reconstruction from finite numbers of projections.
Ritman, S. L* Leo, G. T. Herman, R. E. Stum & E. H* J. Fnys. A: Math*, Nucl. Gen* 6, 36I-382.
Wood (197^). Tbe problem of accurate measurement of
left ventricular shape and dimensions from multiplane D. W. Sweeney & C. M. Vest (1973). Reconstruction
roentgenographic data. Proceedings of the Second of three-dimensional refractive index fields frcm
Workshop on Contractile Behavior of the Heart, Utrecht, multidirectional interferometric data. Applied
6-7 September, 1973. European J. Cardiol*, pp. 137- Optics 12(11), 26k9'266k.
155.
A. Todd-Pokropek (1973). Tomography and the
S* Kaczmarz (1937). Angenaeherte Axifloesimg von reconstruction of images from their projections.
Systemen linearer Gleichungen. Bulletin International Third International Conference on Data Handling and
des Sciences et des Lettres, Classe des Sciences Image Processing in Scintigraphy, Massachusetts
Mathematiques et Naturelles, Serie A: Sciences Institute of Technology, 6-9 June, 1973, in press.
Mathematiques, 355-357-
B. K. Vainshtein (1973)· (Three-dimensional electron
R. L. Kashyap & M. C. Mittal (1973). Picture microscopy of biological macromolecules.) Uspekhl
reconstruction from projections. Proceedings of the Fizlcheskikh Nauk 109(3), ^55-^97·

93