Journal of Microscopy, Val. 1 7 0 , Pt 3 , June 1 9 9 3 , pp. 2 2 1 -228.
`Received 2 January 1 9 9 2 ; accepted 1 7 December 1 9 9 2
`
`Methods for large data volumes from confocal
`scanning laser microscopy of lung
`
`E. H. OLDMIXON* & K. C A R L S S O N t
`*Departments of Medicine, Brown University and Memorial Hospital of Rhode Island,
`Pawtucket. R1 0 2 8 6 0 . U.S.A.
`tPhysics IV, Royal institute of Technology. Stockholm S-10044, Sweden
`
`Key words. Confocal scanning laser microscopy, dog lung parenchyma, intensity
`correction, volume fusion, three-dimensional reconstruction, alveolar ductal unit.
`
`Summary
`
`Confocal scanning laser microscopy makes it possible to
`obtain series of optical sections in precise registration.
`Certain studies of lung parenchyma, however, require both
`the fine resolution obtainable with high-numerical-aperture
`(NA) objectives and the extensive fields of view that usually
`would be achieved only with low-NA objectives. This
`article presents a technique that resolves this conflict by
`using a sequence of operations: (i) to correct intensity
`variations on individual sections due to non-uniform
`illumination/detection characteristics of the microscope;
`(ii) to correct intensity variations between successive
`sections in a series due to, for example, depth-related
`absorption or step changes in detector sensitivity; (iii) to
`adjust adjacent, overlapping stacks of sections to a common
`intensity level; and (iv) to fuse a group of such overlapping
`stacks into a single series of larger sections. This resulting
`stack may contain, for example, a complete cross-section of
`an alveolar ductal unit about 500 pm or more in diameter
`at about 1-pm pixel resolution.
`
`Introduction
`
`This paper differs from much of the image-registration
`literature in addressing the problem of registering two or
`more images of the same scene obtained by different
`imaging techniques (e.g. fusion of multimodality images) or
`registration of images of the same subject obtained by one
`method but at different
`times and under different
`conditions, say of position and illumination. Here we
`explore a procedure for enlarging the volume from a
`biological sample which can be visualized using a confocal
`microscope.
`The method was developed to enable us to study certain
`architectural aspects of the dominant subunit of mamma-
`lian lung parenchyma, the alveolar ductal unit. These units
`are approximately cylindrical, with a relatively long,
`
`0 1993 The Royal Microscopical Society
`
`central lumen around which are placed alveoli opening
`onto the lumen. In fully inflated, canine lung parenchyma.
`the ductal unit diameter, measured at 90' to the lumen axis
`from a back wall of an alveolus on one side to a back wall
`on the other, is 500pm or more. One research goal is to
`measure quantitatively how the parenchymal surface.
`essential both in gas exchange and lung mechanical
`behaviour, is disposed within the alveolar ductal units as
`a function of radial distance from the lumen axis.
`Parenchymal tissue was stained generally with a
`fluorescent dye, while the volumes previously occupied by
`air or blood remained non-fluorescent. To locate and
`measure the parenchymal surface with sufficient accu-
`racy, it was necessary to record the fluorescence intensity in
`1 x 1 x 1-pm volume elements throughout a volume large
`enough to contain the complete cross-section of an alveolar
`ductal unit and over a length of the ductal unit as long as
`obtainable, preferably at least 200 pm. The required data
`volume would
`then have dimensions of
`at
`least
`500 x 500 x 200 elements, orthogonally arrayed. We
`chose to examine ductal units whose axes are normal to
`the surface of the sample block and thus also parallel to the
`depth coordinate; this is the dimension of the data volume
`equalling approximately 200 elements.
`Using a confocal scanning laser microscope (CSLM), it is
`possible to collect a square array of fluorescence intensities
`(e.g. 256 x 256, 512 x 512, 1024 x 1024) at a single
`depth, then to descend to a greater depth and collect
`another square array, aligned precisely with the preceding
`one. For reasons explained later, however, each such
`square array of volume elements is physically too small to
`record the fluorescence pattern across an entire alveolar
`ductal unit cross-section. In practice, therefore, the whole
`data volume has to be built up by recording from 4 to 9
`adjacent series of 200 sections each, each series recording
`the entire available depth but only a fraction of the cross-
`section of the alveolar ductal unit. The lateral displace-
`
`221
`
`

`
`222 E . H. OLDMIXON & K . CARLSSON
`
`Fig. 1. Composite picture of four 256 x 256-pixel images which.
`together, record the entire cross-section at this depth of an alveo-
`lar ductal unit in dog lung parenchyma. In the text, the upper left
`image is I, upper right is m, lower right is n and lower left is 0; the
`entire image area (512 x 512 pixels) shown here is p. Each quad-
`rant image's content duplicates some content in the neighbouring
`images; thus, 1 shares content with m and o etc. These images are
`raw data fr3m the CSLM, except that for visibility in the figure the
`intensities of dim pixels were increased. The images I, m, n, o and p
`all are square, in reality; the vertical elongation and slight shearing
`seen were introduced by the video printer.
`
`these section series are not as yet
`ments between
`controllable at pixel dimensions, but have uncertainties of
`tens of micrometres. To enable subsequent lateral regis-
`tration of the whole series, each series is made to contain
`some of the same tissue features as its neighbours (see
`Fig. 1).
`Four problems are encountered when progressing from
`the separate section series to the single data volume
`containing the desired ductal unit:
`(1) illumination/detection characteristics of the microscope
`are not identical at every pixel location within a field of
`view;
`
`(2) the same structure is usually not recorded at the same
`intensity from one section to the next, an effect due to any
`of several causes, such as increasing absorption or
`scattering of incident and emitted light with depth, (mild)
`bleaching, or changes (made between sections) in instru-
`mental illumination and/or detection sensitivity;
`(3) the section series have to be adjusted to a common
`intensity level;
`
`(4) the section series containing different portions of the
`desired ductal unit cross-section must be fused into a single
`series containing a complete cross-section of the subject.
`
`Methods
`
`Specimen preparation
`
`Figure 1 shows dog lung parenchymal tissue. The lobe was
`obtained by approved procedure, inflated to a transpul-
`monary pressure of 30 cm HzO, then chemically and
`mechanically
`fixed by perfusion with glutaraldehyde
`[2.5% in sodium phosphate buffer, 0.075 M , pH 7.4, 350
`mOsm, 3% dextran (average molecular weight 72,000 Da)],
`tannic acid (4% w/v in sodium cacodylate buffer, 0.2 M, pH
`7.4, 350 mOsm, 52.5 mM sodium chloride, 3% dextran) and
`ethanol (linear gradient from 10% in H20 to 100%. 500 ml
`total), sequentially. Samples were cut with snapped, double-
`edged razor blades, immersed at once in absolute ethanol
`and dehydrated with gentle flow under vacuum overnight.
`Samples were submerged with gentle swirling in Lucifer
`Yellow CH fluorescent stain (0.001% in absolute ethanol) in
`vials (kept dark) for 24 h, rinsed with absolute ethanol (two
`changes, 30 min each), then processed through propylene
`oxide to Spurr's epoxy, to be cured at 55°C for 36 h. Each
`block was milled or lathed flat on one face, cutting into
`embedded tissue, sectioned with a glass knife and glued by
`the corners of this face to a flat-faced, ring-shaped plastic
`adaptor. This new facet was machined parallel to the first,
`then sectioned with a glass knife. The finished sample was
`carefully released from the adaptor, touched up and
`attached with a piece of adhesive putty to a microscope
`slide by either sectioned facet, exposing a smooth, level
`tissue section. Tissue prepared by this method is believed to
`retain in viva size and shape (Oldmixon et al., 1985).
`
`Confocal microscopy
`
`A CSLM developed at the Royal Institute of Technology in
`Sweden (Carlsson et al., 1985: Carlsson & Liljeborg. 1989)
`produces digitized images in square format, with up to
`1024 x 1024 square pixels
`in orthogonal rows and
`columns. Series of images obtained successively deeper
`within a sample are precisely aligned. Routinely, resolution
`in the plane of the section is closely comparable to that of
`the same microscope used in conventional transmission
`mode. The images, acquired in digital form, are suited to
`digital
`image processing and analysis and may be
`reconstructed in three dimensions.
`A characteristic of this microscope that is especially
`significant for this project is its low geometric distortion
`within each section, maximally 1.7% (Carlsson, 1990). We
`have elected to view this distortion as insignificant and have
`not attempted to compensate for it.
`
`

`
`An argon laser was used for illumination at 473 nm.
`with typically 0.06-0.1 mW actually transmitted to the
`specimen.
`
`Apparent depth resolution and field of view
`
`The depth resolution of the CSLM is such that an image
`obtained with a 63 x oil-immersion, planapochromatic
`objective (NA 1.4) resembIes subjectively a physically
`microtomed section of the same sample about 1.3 pm
`thick; with a 4 0 x oil-immersion, planapochromatic
`is to a 2.0-2.5-pm-thick
`objective, the resemblance
`physical section. The respective objectives produce images
`which cover nominal 161 x 161- or 256 x 256-pm2 fields
`of view. A 25 x objective (NA 0.64) provided a larger field of
`view, but unfortunately the optical sections were subjec-
`tively equivalent to 7-10-pm-thick physical sections.
`
`Limits on depth of observation
`
`In theory at least, sections may be obtained from the surface
`of the specimen down to a depth equal to the working
`distance of
`the objective. In practice, the maximum
`productive working depth may be decreased drastically by
`an optically dense specimen, which may
`introduce
`attenuation and degradation of resolution. If the refractive
`index of the specimen, n, differs from that, nl, of the medium
`interposed between the objective and the specimen (air,
`immersion oil, etc.) then the actual depth, dl, from which
`the image is obtained will not be the apparent depth, d,
`within the sample measured by the mechanical vertical
`displacement of the stage relative to the position at which
`the optics focus on the top of the specimen, but will be given
`by (Carlsson, 1991)
`
`increased by
`The effective working distance may be
`observing the specimen without a coverslip, if the physical
`properties of the specimen permit; the imaging depth may
`thus be increased to about 2 70 pm with the 63 x objective
`and deeper with the 40x objective. With epoxy-embedded
`lung samples, 200pm is currently the limiting depth, in
`practice.
`
`Sizes of physical and data volumes
`
`Combining the preceding considerations, the physical
`volumes that can be sampled from lung are 1 6 1 x 1 6 1 x
`200pm when using the 63x objective and 256 x 256x
`200 pm using the 40x objective. Converting these physical
`volumes to data volumes, however, must be influenced by
`such factors as the requisite resolution, anticipated pro-
`cessing techniques, and available hardware and software.
`
`LARGE C S L M DATA VOLIIMES 22 3
`
`In the present application, in-plane pixel dimensions of
`1 x 1 pm would be satisfactory to retain connective tissue
`cable profiles as groups of four or more distinctively high-
`intensity pixels. The 40 x (1.2 5 x auxiliary lens) objective
`filled this requirement, giving 256 x 256 pixel images with
`pixel dimensions of 1.0 x 1.0 pm2.
`intensity correction
`Experience with section series
`algorithms (the second problem, above) has shown that
`vertical or depth spacing between sections should be less
`than half the apparent corresponding physical section
`thickness, here, 2,512 = 1.25pm or less. Quite con-
`veniently, this spacing was chosen as 1.0 pm, giving voxel
`dimensions of 1.0 x 1.0 x 1.0 pm.
`The physical view volume, therefore, could be captured in
`200 sections each of 256 x 256 pixels. Each voxel, when
`recorded, is a 1-byte (8-bit) number, so the data volume
`from one stack would occupy 13.1 Mb. During processing.
`voxels are stored as 2-byte data, bringing the size to
`26.2 Mb per stack. Anticipating having to fuse together 4.
`5, 6, or even 9 stacks to portray the ductal unit segment,
`even a single instance of the data volume could eventually
`occupy 2 3 6 Mb.
`
`Computing hardware and software
`
`The growing availability of fast processors, generous
`memory and graphical interfaces makes it almost futile to
`specify a minimal system, but the work described here has
`been done comfortably with the following: Sun Microsys-
`tems (Sunnyvale, CA) 3/150 six-slot VME chassis with
`16 MHz Motorola 68020 processor, 19" grey -scale monitor,
`4-Mb memory, 160-Mb SCSI disk, 1/4" cartridge tape (QIC-
`24, 57Mb capacity), 1.2-Gb CDC (Minneapolis. MN) disk
`with Interphase (Austin. TX) ESMD controller, Vicom
`Systems (Fremont, CA) VME image processing peripheral
`(six VME slots, 1 5 Vicabus slots) with point processor,
`ensemble processor, 4-Mb image memory, 1-Mb display
`memory, display subsystem with 19" RGB monitor, and Sky
`(Lowell, MA) Warrior vector processor.
`Microscope control and image acquisition/archiving
`visualization software accompanies the CSLM; the com-
`bination is available from Molecular Dynamics (Sunnyvale,
`CA) as a commercial product. Vicom Systems VIPs
`software, particularly
`the interactive command inter-
`preter, was used for examining pixel values on displayed
`images and for correcting
`image sequences by
`the
`compensating image, for example. The desired operations
`can be programmed and the Unix awk utility used to create
`a filed sequence of commands retrievable from within the
`command interpreter. While this approach does not operate
`the hardware at its maximum speed, the time lost is small
`and the rapidity and flexibility of programming are more
`than compensatory. A cycle of correction by the com-
`pensating image takes about 3 s per serial section.
`
`

`
`Generating the compensating image using the local thin-
`plate spline program (Franke. 1982) in Fortran requires
`over 1 h on our system. We have not attempted to vectorize
`the code.
`Intensity adjustment section by section through a stack
`makes full use of both Vicom and Sky. Computation of
`adjustment factors takes about 5 s per serial section, and
`subsequent intensity adjustment about 3 s. As most of the
`time is spent transferring images to and from disk, we
`suspect that the Unix filing system creates a bottleneck here.
`Translating and fusing four 256 x 256-pixel sections
`followed by writing out the 512 x 512 result, for entire
`sequences, is handled by the Vicom command interpreter
`reading an awk-generated command file.
`
`Choices of pixel and section spacing
`
`For a variety of reasons concerned with the amount of
`installed memory, the underlying architecture of processing
`boards and display systems, and associated software, the
`image size most convenient for processing and analysis on
`our system is 512 x 512 pixels. Fortunately, this size often
`suffices to hold a complete cross-sectional view of our
`subject, the ductal unit.
`A ductal unit is, roughly, a cylinder c. 500pm in
`diameter. A cross-section of a ductal unit centred in a
`5 12 x 5 12 pixel image with some of the surrounding tissue
`for context could have pixels somewhat less than or equal to
`1.Opm on a side. The connective tissue cables which we
`wish to detect within the ductal unit typically have
`minimum caliper diameters of several micrometres or
`more; they often fully occupy groups of four or more pixels
`and thus are readily detectable.
`As mentioned above, pixels 1.0 pm2 can be produced by
`using a 40x objective to scan 256 x 256 pixel images
`(which are c. 2 56 x 2 5 6 pm in size). Since four such images
`arranged as quadrants cover a square c. 512 pm on a side,
`it is feasible in many cases to move each image toward the
`centre of the composite field, so that each overlaps its
`neighbours, and still cover the cross-section of the ductal
`unit and the surrounding context. Such a group of four
`series is used for illustration in this article, but other ductal
`units may require six (2 x 3) or nine (3 x 3) series to cover
`the ductal unit cross-section.
`
`Approaches to solutions of problems encountered in
`fusing multiple section series into single data volume
`
`Compensating for non-uniform illumination/detection within
`a section
`
`CSLMs which scan by deflecting a beam angularly neither
`illuminate a field of view perfectly evenly nor detect light
`emitted from each pixel position with equal efficiency. We
`
`compensated for this systematic unevenness in each section
`before proceeding further.
`The combined illumination and detection behaviour of
`the system can be determined empirically by scanning
`certain glasses which fluoresce uniformly and brightly upon
`their first exposure to laser light (e.g. an orange filter, Schott
`No. OF530). Intensity gradients seen systematically on
`individual scanned images are then due to the illumination/
`detection characteristics of the entire system. We sup-
`pressed high-frequency photon noise and extracted the
`gradients by averaging together eight images each scanned
`from fresh regions of the glass, using the same objective.
`pixel spacing and pixel number as for the section images.
`The plane of focus was 5-15 pm below the glass surface.
`Since emission intensity from the fluorescent glass falls
`I'apidly with repeated scans, laser power and detector
`sensitivity (photomultiplier voltage) were tested at distant
`locations on the glass before scanning the final calibration
`images. Since the fields to be scanned should not be
`illuminated beforehand, the stage must be moved just prior
`to scanning; there is, therefore, a slight chance that a
`previously exposed region would be encountered. The
`scanned images were scrutinized for suspect, dim areas,
`usually straight-sided, angular and crossing the image
`border; such flawed images were discarded and not included
`in the averaging.
`The averaged image, however, still contained random.
`high-frequency variations. We inspected the pixel values
`(now in 16-bit representation) in 8 x 8-pixel blocks in
`about 50 regions across the entire image and recorded the
`characteristic pixel intensity of each region. The object was
`to specify benchmarks to pass to a surface interpolation
`program which would produce a standard image with first-
`order continuity, while preserving all the important features
`of the illumination/detection behaviour. Additional pixel
`values were specified at the corners and along the sides:
`here values were chosen by extrapolating visually smooth
`curves through groups of pixel values at interior locations
`in the neighbourhoods of the desired border pixels. The local
`thin-plate spline surface interpolation method of Franke
`(1982) produced a 256 x 256 array of pixel values which
`when displayed appeared to be a noise-free rendering of
`the averaged image. The final compensating image C (see
`Fig. 2) was generated:
`C(i, j) = m * k/S(i, j ) ,
`where i, j are indices over the ranges of pixel locations in
`image rows and columns, respectively, C is the compensat-
`ing image, S the interpolated intensity of a standard image,
`m is the maximum representable pixel intensity (here
`+ 0.9995) and k is the lowest value in image S.
`Each scanned section was multiplied by image C to
`produce a new, compensated image in signed 16-bit
`representation.
`
`

`
`L A R G E CSLM D A T A VOLUMES 2 2 5
`
`reports no important changes to the intensities. In practice,
`this algorithm determines the factor needed to change the
`intensity of a section to that of its predecessor, and the
`factor applied to any section is the cumulative product of
`the preceding factors. This might result in some sections
`being brightened undesirably to the point of saturation and
`may be avoided by reducing the brightness of the first
`section by 15-20% or by scaling the factors so that the
`maximal pixel intensity in the adjusted series will be below
`saturation. Both variations have been used successfully.
`It may be of interest that once the images have been
`loaded from disk the computing system calculates the factor
`relating image-pair intensities in about 1 s. Mathematically,
`the code computes the eigenvector, but does so using the
`Vicom image processor to define regions of interest, add
`images, multiply them, compute their fractional powers,
`etc., taking a 'parallel processing' approach. These routines
`may be requested from the first author.
`Note that this serial-section intensity-correction algorithm
`succeeds or fails depending on changes in image content, the
`characteristic dimensions of fluorescent structures and voxel
`dimensions, particularly the depth increment. In our
`experience, it tends to fail on typical lung parenchymal
`image series if (other settings remaining the same) the depth
`increment is increased above 2.0 pm. (A failure is signalled by
`the appearance of saturated regions which spread and fuse.
`often eventually to obliterate the images.)
`
`Calculating displacements needed to form a composite image
`
`Consider a series of four 256 x 256-pixel images (I, m, n, o)
`collected in the pattern shown in Fig. 1. The four
`256 x 256-pixel
`images, arranged
`to abut, form a
`512 x 512-pixel image, p. The origin of each image is
`considered conventionally to be its upper left corner
`expressed as (row = 0, column = 0). Thus, the origin of
`image I is (0, O),, coincident with the origin of image p. The
`origin of image m on image p, though, is (0, 256),, or n,
`(256, 256), and of o, (256, O),.
`Relative displacements between I , m, n and o were
`determined by inspecting each neighbouring, overlapping
`pair of stacks along their depth, and looking for bright.
`distinctive, well-defined, shared features. Connective tissue
`cables shared by 1 and m (or by m and n, n and o or o and I)
`are often best. The pixel intensities in and around such
`features visible on two sections from the two different series
`were recorded on graph paper. Although the displacement
`between stacks will seldom be an integral number of pixel
`rows or columns so that the two pixel intensity patterns
`might be identical, the edge, centroid or maximal brightness
`can usually be located in both instances of the feature. The
`row and column displacements needed to move the landmark
`of the feature on image m to coincide with that on image 1
`were expressed in terms of coordinates on image p.
`
`Fig. 2. Compensation image for systematic intensity gradients within
`a field of view (see text). The image consists of 256 x 256 pixels, with
`a minimum value of 0.77490 at row 103, column 139 (row 1, col-
`umn 1 is the upper left corner), and maximum value of 0.99609 at
`row 256, column 1. The image was generated by performing a local
`thin-plate spline interpolation among over 50 landmarks selected
`manually from averaged and filtered images taken of a fluorescent
`glass standard, then computing reciprocal image and scaling.
`
`Equalizing intensities among images of a series
`
`During the time needed to scan a section series through the
`available depth range,
`the overall section
`intensity
`decreases due to the factors mentioned above. If this
`decrease jeopardizes the number of significant figures in
`the fluorescence intensity data, detector sensitivity may be
`adjusted between section scans so as to keep the maximum
`recorded intensity just below saturation. Since we do not
`know beforehand the optimal adjustment, even such
`section series may be improved by post-acquisition correc-
`tion of overall section intensities within each series. Our
`method of correcting uneven intensities among images in a
`depth and/or time series may be described simply. For each
`pixel location, the intensity on a section is plotted against
`the intensity on its immediate predecessor, forming a two-
`dimensional scatter plot or histogram. Any pixel pair in
`which one of the two intensities is very low (0.0-0.06 of
`full scale) or very high (1.0 of full scale) is deleted. This
`eliminates airspace with noise and saturated pixels,
`respectively. The slope of the best fit line (least sum of
`squared minimal distances from points to the line) through
`the remaining points is calculated, and the section is divided
`by the slope. The process is repeated through the stack,
`three times. On the third and final pass, the process typically
`
`

`
`226 E. H. OLDMIXON & K . CARLSSON
`
`This manual procedure is facilitated by displaying both
`sections on a monitor and magnifying to inspect the
`individual pixels. Usually there is no uncertainty by the
`operator as to the relative displacement; when there is, it is
`usually between two possible displacements differing by one
`pixel. In practice, when the matches made at 3-5 levels
`throughout the pair of stacks are reviewed, the proper
`displacement becomes clear.
`It should be possible to find those best matches by a
`computable criterion, perhaps using an algorithm similar to
`the intensity adjustment algorithm but varying displace-
`ment along rows and columns instead of intensity. We have
`not implemented such a procedure, however. The present
`manual matching procedure takes several hours of work
`per volume data set. The process was repeated for image
`series 1 and o; the displacement expressed the translations
`along rows and columns needed to move image o to
`coincide with image 1. Next, the process was repeated for n
`and m, then n and o; in these cases, the displacements were
`expressed as those needed to move n to coincide with m (or
`o) before m (or o) was translated to coincide with I. If we
`notate the row and column translations needed to move
`features on image b to coincide with those on image a as
`i
`b -+ a (r, c), then the translations described above are rn -+
`(r, c), o -+ 1 (r, c), n + m (r, c) and n -+ o (r, c). Note that
`a -+ b (r, C) = -b -+ a (r, C) = b -+ a (-r, -c).
`The four displacements (I -+ m -+ n + o -+ 1) should
`make a closed loop:
`rn --+ 1 (r, c) + n +m (r, c) + o -+ n (r, c) + I + o (r, c)
`
`If not, then the error has probably been introduced from
`two sources: (i) instrumental distortions of the fields of view
`and (ii) non-cancelling errors in selecting integral pixel
`translations to approximate actual translations between
`fields of view. Our experience has been that 1 -+ I (r, c)
`values are of the order of 1 or 2 pixels, a disparity readily
`eliminated by slightly changing one or more of the four
`component translations. The fact that the translation loop is
`nearly closed gives confidence in both the instrument and
`the stack-fusing technique.
`The preceding steps may leave the centre of the ductal
`unit's profile well away from the centre of image p; since the
`data volume may be conveniently rotated about its centre
`when making three-dimensional reconstructions, we try to
`bring the centre of the ductal unit to the centre of the data
`volume. The translation needed to shift the fused image p so
`that the tissue of interest is centred in the final 5 12 x 5 12
`image q is denoted by p -+ q (r, c).
`The final translations are
`
`Fig. 3. An example of four 256 x 256-pixel images (1, m. n, o)
`translated and fused to form 512 x 512-pixel image q (see text).
`Th translation operation has moved images I, o and n toward the
`centre of image q. h a g e m has not been shifted in this example;
`m -t q (r = 0, c = 0). The portions of I, n and o which were trans-
`lated across the original boundaries were discarded where they fell
`into a portion of the neighbouring image after translation, but kept
`in two places where the content of image I lies in an 'unoccupied'
`portion of the original o domain and where o does the same in n.
`
`o + q (r, C) = 0 -+ 1 (r, C) + P -+ 4 (rr c).
`The remaining question is how to deal with the regions of
`overlap among the images. A simple approach, which is
`usually satisfactory, is to retain each pixel of an image which
`remains within the original 256 x 2 56 domain of that
`image, and to retain a pixel shifted out only if its new location
`is unoccupied by a pixel from any other image (see Fig. 3).
`
`Equating intensities among stacks
`
`Before creating image q, the relative brightnesses of I , rn, n
`and o must be inspected and probably modified. Although
`
`Fig. 4. The final result; a surface-shaded reconstruction of canine
`lung alveolar ductal unit constructed from stacks containing the
`sections shown in Fig. 1. Tissue not belonging to this ductal unit
`has been removed.
`
`

`
`the intensities within stacks I, m, n and o have ostensibly
`been equalized, there is no guarantee that their intensities
`are equal, stack to stack. Having decided upon the
`translations, however, we have also defined the over-
`lapping regions of corresponding sections in each series.
`Taking the I-m pairs as one example, we generated a
`histogram from stack 1 of the tissue intensities in the region
`of overlap with stack m, then the same for stack m. The
`factor used to bring the tissue to a common intensity was
`determined by inspection. (The eigenvector method used to
`adjust pairwise intensities within a stack is also applicable.)
`The factors from m to I, n to m, o to n and I to o were found,
`then adjusted as required to bring the final intensities of all
`stacks to the desired level. Note that the product of these
`four factors should be unity.
`fusing four 256 x
`Figure 4 shows the result of
`256 x 192-voxel stacks into a 512 x 512 x 192-voxel
`stack by this process.
`
`Applications to lung structure
`
`Larger view volumes composed in the manner described
`prove useful in the investigation of lung parenchymal
`structure in several respects. A ductal unit is packed among
`other ductal units, and few portraits of one exist. The
`boundary of the composed ductal unit can be determined
`readily, and tissue beyond the boundary then deleted, with
`the result that a reconstruction of
`the edited volume
`shows the normally embedded ductal unit now isolated for
`study (Fig. 4). Qualitatively, it may be informative to see
`an anatomic structure which otherwise might remain an
`abstraction. The composed volumes promise to be useful
`in quantitative studies as well. The positions of tissue
`components (connective tissue condensations, for example)
`may be marked on the sections, since the internal structure
`of the tissue is visible there. It is also feasible to extract the
`voxels at the air-tissue interface, record their locations and
`determine the local orientation of the interface. The ductal
`unit axis may be established by viewing the entire volume
`and individual sections in turn. It is then possible, for
`example, to discern the radial distribution of connective
`tissue and air-tissue interface about the duct axis.
`
`Generality of the method and relationship to other
`techniques
`
`Techniques developed for the combination of the Royal
`Institute of Technology's confocal microscope and lung
`parenchymal samples should be applicable to other image
`acquisition methods and other samples, if the following
`qualifications are met.
`
`(1) The instrumental illumination/detection pattern within
`the field of view must be characterizable and reproducible.
`
`L A R G E CSLM DATA VOLUMES 2 2 7
`
`intensity equalization
`(2) To perform post-collection
`throughout a section series after unknown alterations
`to the instrument illumination/detection levels between
`section, the sample should contain
`three-dimensional
`structures which will appear in the same pixel locations
`on successive sections and which have the same intrinsic
`fluorescence in both instances. In the lung, examples
`of these structures which serve well with 1 x 1 x 1 pm
`voxels are connective tissue cables, nuclei and basement
`membranes.
`(3) Geometric distortion of the field of view has not been
`addressed, as it is judged not to be a problem with this
`CSLM. If distortion correction must be done, it should
`probably be performed after the within-section and within-
`series intensity adjustments have been made.
`
`The alignment of neighbouring section series as done
`with lung samples depends on small, bright, distinctive
`image