The key to excellent static and functional results in cleft lip and palate (clp) reconstructive surgery lies in detailed knowledge of the normal and pathological labial, palatal, and velopharyngeal anatomy. The current therapeutic concepts of clp treatment aim at functional reconstruction or imitation of normal anatomy, as expounded by Kriens (1967), Randall et al. (1974), Delaire (1975), Millard (1976), Kernahan et al. (1984), Koch et al. (1998), and Bitter (2000). Earlier anatomical studies of clp and normal anatomy were performed by macroscopic dissection (Burkitt and Lightoller, 1928; Whillis, 1930; Mullen, 1931; Veau, 1932; Koerner, 1942; Lee, 1946; King, 1954; Slaughter et al., 1960; Fara et al., 1965; Fara and Smahel, 1967; Braithwaite and Maurice, 1968; Novoselov and Lavrentiev, 1969; Pennisi et al., 1969; Fara and Dvorak, 1970; Brescia, 1971; Dickson, 1972, 1975; Dickson and Dickson, 1972; Latham, 1973; Latham and Deaton, 1976; Azzam and Kuehn, 1977; Klueber and Langdon, 1979; Lewin et al., 1980; Atkins et al., 1982; Doyle et al., 1983; Boorman and Sommerlad, 1985a, 1985b; Maher, 1986; Swarts et al., 1986; DeMey et al., 1989; Kuehn and Kahane, 1990; Zhang, 1990; Huang et al., 1997a, 1997b, 1998a, 1998b; Ishijima et al., 2002). All peripheral structures were ablated during the preparation, which was intended to avoid by dissection from different directions (Dickson and Dickson, 1972). Nonablative visualization required either virtual serial sectioning by CT or MRI (Kleinheinz and Joos, 2001), sonography (Martin et al., 2000; Munshi et al., 2000), or physical serial sectioning (Latham and Deaton, 1976; Latham et al., 1980; Swarts et al., 1986, Swarts and Rood, 1990; Spauwen et al., 1991; Ishijima et al., 2002; MinGuo et al., 2004).
Three-dimensional modeling of anatomy and histology generally permits optimal visualization of complex structures and offers new aspects to surgical approaches for potentially higher efficacy, safety, and less operative trauma. Virtual sectioning techniques (e.g., CT, MRI, sonography) yield far inferior microanatomical resolution, however, compared to physical serial sectioning. Discrete anatomical structures cannot be delineated by virtual sectioning techniques because of the ambiguous relationship between signal intensity and tissue composition (Aritan et al., 1997; MingGuo et al., 2004). Building physical laminate 3D reconstructions from serial sections, first achieved by Born (1876, 1893), is a traditional standard technique in anatomy (e.g., Gaupp, 1893; Kuhn, 1971; Fischer, 1989; Haas and Fischer, 1997). Since Born's work, several laborious manual and graphical techniques have been proposed to reconstruct objects from serial sections (Gaunt and Gaunt, 1978). Earlier 3D visualization of midfacial and Eustachian tube anatomy stacked acrylic glass or wax blueprints from outlined structures of interest (Siegel and Todhunter, 1979; Latham et al., 1980; Swarts and Rood, 1990; Spauwen et al., 1991). However, 3D reconstruction from serial sections for accurate computer models of anatomical structures, with complex shape and mixed tissues, can be very difficult (Haas and Fischer, 1997; MingGuo et al., 2004). The quality of a reconstruction from histological sections depends on several factors: tissue changes during preservation and embedding; tissue distortion from serial sectioning by microtome or wire saw; tissue staining contrast and optical reproduction for on-screen digitization or alternative reproductive media; accuracy of data acquisition, including the vertical dimension; reliability of section alignment; and quality of surface generation (manual drawing or computer-based visualization). No matter how much error is tolerated, the reconstruction must yield a result adequately exact and detailed to comply with the objectives of faithfully representing the spatial and volume relationships for a realistic mesh generation and finite-element analysis (FEA), deformation calculation, and education (i.e., optimum comprehensibility). These issues have been reviewed elsewhere (Prothero and Prothero, 1986; Brändle, 1989; Deverell et al., 1989, 1993; Hara et al., 1989; Laan et al., 1989; Lozanoff and Deptuch 1991; McLean and Prothero, 1991; Ongaro et al., 1991; Lozanoff, 1992; Marco and Leith, 1992; Moss, 1992; Vuillemin et al., 1992; Clarke et al., 1993; Hibbard et al., 1993). The use of computers in 3D reconstruction began more than 20 years ago (Gaunt and Gaunt, 1978; Huijsmans et al., 1986) and fundamental algorithms have been outlined (Johnson and Capowski, 1985; Moss, 1992; Kvasnicka and Thiele, 1995).
Performing virtual 3D reconstruction on a computer has desirable advantages over laminate model building. Virtual reconstructions are not merely visual representations of objects, but they can be measured and analyzed accurately (landmark coordinates, distances, surface area, volume). They can be stored, duplicated, transferred between colleagues, or provided in accessible databases (i.e., the World Wide Web). They can be cut and dissected electronically in any plane and the muscular dynamics calculated by FEA. Automatic digitization or surface computing should facilitate rapid reconstruction of a large series of specimens (Lozanoff et al., 1994). Stereolithography allows output of virtual reconstructions as physical models at any time, if desired. The geometric basics, problems, and requirements of computer-based 3D reconstruction in anatomy have been known for quite some time (Christiansen and Sederberg, 1978; Boissonat, 1988; Lozanoff and Deptuch, 1991; Gerard et al., 1993; Arnold et al., 1996; Gan et al., 2002). A variety of computer applications have been developed by various groups in parallel. However, no software has become a standard and the packages widely differ in concept (voxel- or polymesh-based), platform (Windows, Mac, UNIX/Linux), implemented features, and availability (free distribution to exclusive product design rental software) (Haas and Fischer, 1997). Software generation was paralleled by a burst in hardware speed and decrease in price.
The present study was designed to compare two software solutions for 3D reconstruction, based on histological serial sections. The systems in this article are demonstrated for two techniques of computed 3D reconstruction on an average hardware basis.
MATERIALS AND METHODS
Eighteen aborted fetuses, coded A to R, were included in this study after thorough informed parental consent. The gestational ages ranged between the 11th and the 32nd week. Six fetuses bore clp; the other 12 fetuses had normal anatomy and thus served as controls (Table 1). All of the fetuses were fixed with 4% formaldehyde solution. The midfacial region, including the cranial base, was isolated from the infraorbital rim to the lower lip or chin, with a maximum 20 mm section thickness (analogue to the cuvette height). Twelve randomized specimens (four, or 33%, bearing clp) were infiltrated with 2% celloidin (Merck, Darmstadt, Germany) for 2 days, 4% for 2 days, and 8% for 1 week (Dickson and Dickson, 1972; Shibahara and Sando, 1988; Swarts and Rood, 1990; Cohen et al., 1993, 1994; Heinsen et al., 2000; Hannecke et al., 2001; Ishijima et al., 2002). Spontaneous evaporation concentrated the celloidin solution to approximately 16%. Chloroform hardening was for 2–3 days; a second hardening was performed in 70% alcohol. Tissue sections (60 μm thick) were cut by sliding microtome (Polycut S; Leica, Heidelberg, Germany) in transverse (10) or coronal (2) plane and stored in 70% alcohol. Hematoxylin and eosin (H&E; Merck) staining was performed according to routine protocol (Romeis, 1989), embedding with Caedax (Merck) that required 4–6 weeks to dry. Six (33%) randomized specimens (two, or 33%, bore clp) were impregnated with a modified plastination technique according to Fritsch (1989, 1996) and Fritsch and Hegemann (1991). The undecalcified midface was dehydrated 4 weeks by freeze substitution, 2 weeks degreasing and forced impregnation with BiodurE12 (Hagens, Heidelberg, Germany), and epoxy resin embedment. Biodur required 1 week to harden and was sectioned by wire saw (Well W. Ebner, Mannheim, Germany) in transverse (five) and coronal (one) plane. The thickness of the sections was approximately 120 μm, but thickness became less because of saw path loss and ultra-burnishing, leaving section thickness of approximately 90 μm. After mounting and polishing, the sections were stained with azure II/methylene blue for 3 min at 90°C and counterstained with basic fuchsin at the same temperature (AMF) (Fritsch, 1989). Preparation of the sections, independent of histological method, took 3 months to preserve tissue integrity. The sections were observed using a stereomicroscope (STEMI SV 11; Zeiss, Jena, Germany; magnification, 6 to 66×) by microsegmentation (Universal Forschungsmikroskop Leica, Wetzlar, Germany; magnification, 35 to 1,000×). Every section's vertical thickness was calculated over the main tissues: nerve, muscle, bone, cartilage, and connective tissue, and a thickness correction factor was derived from a vertical test calibration (marked glass with exactly 0.1 mm thickness; Table 1). Horizontal scaling used a reference scale of known diameter, included in every picture. Each section was scanned by analogue (OM 1 Olympus, Japan; Zeiss 456005 adapter x2.5; Ektachrome 64 film, E6 reference processing, Paris, France) or digital camera (2,272 × 1,704 pixels, 72 dpi, 24 bit picture depth, True Color RGB, uncompressed RAW format, by Canon Powershot G3, LADC 58B adapter, and 250D 58 mm macrolens, Canon, Japan; on a backlighted phototable). The color slides were scanned (2,736 × 2,112 pixel, 1,200 dpi, 24 bits picture depth, True Color RGB, uncompressed, 17 MB average; Polaroid SprintScan 4000, Waltham, MA) and visualized in Photoshop 6.0 software as bitmaps (bmp; Adobe, San Jose, CA). The digital camera photos were downloaded (ZoomBrowser EX, Canon) and transformed to tagged image format (12 MB average TIFF).
Table 1. Table of specimen data: gestational age and diagnosis, embedding process, number of histological sections, section thickness and artifact per section
approximate gestational age (Hansmann, 1976; Voigt and Schumacher, 1985)
embedding, staining method
number of histological sections
planimetric percentage of disruption artifact per section
planimetric percentage of compression artifact per section
planimetric percentage of staining artifact per section
planimetric percentage of drying artifact per section
total artifact areas
celloidin & HE
celloidin & HE
celloidin & HE
celloidin & HE
bilateral cleft lip alveolus and palate
Biodur E12, azur 11, methylenblue, fuchsin
left cleft lip alveolus and palate
celloidin & HE
celloidin & HE
celloidin & HE
celloidin & HE
celloidin & HE
celloidin & HE
right cleft lip alveolus and palate
Biodur E12, azur 11, methylenblue, fuchsin
left cleft lip alveolus and palate
celloidin & HE
Biodur E12, azur 11, methylenblue, fuchsin
Biodur E12, azur 11, methylenblue, fuchsin
Biodur E12, azur 11, methylenblue fuchsin
Biodur E12, azur 11, methylenblue, fuchsin
celloidin & HE
The first segmentation was performed using SURFdriver 3.5.6 software (University of Honolulu, Honolulu HI) (Lozanoff and Diewert, 1989; Lozanoff and Deptuch, 1991; Lozanoff et al., 1994). Segmentation was done at variable magnification from 10 to 20×, and vertical setoff correction, surface generation, and dynamic visualization were used. Hardware was an Intel Pentium 2 CPU, 512 MB RAM, NVIDIA G-force 4 TI (64 MB) workstation. Differentiation of the following structures of interest was performed: orbicularis oris muscle (MOO, peripheral portion pMOO, marginal portion mMOO) zygomaticus major (MZMA), zygomaticus minor (MZMI), quadratus labii superioris (MQLS): levator labii superioris caput angularis, caput infraorbitalis and caput zygomaticum, caninus and levator labii superioris alaeque nasi, depressor anguli oris (MDAO), risorius (MR), depressor labii inferioris (MDLI), mentalis (MM), nasalis (MN), dilator nares (MDN), levator septi nasi (MLSN), buccinator (B), platysma (P), levator veli palatini (LVP), tensor veli palatini (TVP), musculus uvulae (MU), palatopharyngeus (PP), constrictor pharyngeus superius (CPS), salphingopharyngeus (SP), palatoglossus muscle (PG), fascia veli palatini, glands. Bone: atlas, axis, mandible, maxilla, nasal bone, occipital bone, palatine, sphenoidale, temporal bone, zygomatic bone, vomer. Cartilage: alar, triangular and septal, Eustachian tube. Vessels, arteries: facial artery (AF), ascending palatine (APA), palatine descending (APD), major and minor palatine (AP), pharyngeal ascending (APHA), recurrent pharyngeal artery (APR) (Huang et al., 1998b). Veins: plexus venosus pharyngeus (VPP), V. buccalis (VB), V. facialis (VF), V. jugularis (VJ), Vv. labiales (VL), plexus venosus palati, and venae palatinae (VP). Nerves: nervus facialis (NF), N. trigeminus (NT), with subunits: N. mandibularis (NMD), N. maxillaris (NMA), and in part: N. auriculotemporalis (NAT), N. glossopharyngeus (NG). The structures for 3D reconstruction were identified and labeled on a section-by-section basis, creating contours by mouse click. Automatic labeling was imprecise. The software reproduced a maximum of 9 independent structures, with 16 substructures (n = 144). Three-dimensional vertical orientation was performed manually, using an internal marker, such as the vertebral column and the cranial base. Contours were manually superimposed over the different tissue types along visible color lines of tissue morphology separation. For the surface reconstruction, no skew correction or vertex stripping, 75% stringency, and nonmetric smoothing was used. Skew correction did not result in significantly better surface morphology. Stringency, a built-in algorithm of lining up consecutive slices, aligned minimal alignment problems not covered by manual superpositioning. Vertex stripping reduced the number of labeling points for faster surfacing; however, it was not used to maintain all details. The object editor after surface reconstruction (Boissonat, 1988) allowed activation of the nine independent structures, each of which could be individually set in shininess and opacity. Options were rotation-limited to x- and z-directions, like a QuickTime VR (Nieder et al., 2000) object, or free rotation in all three axes, or stereoscopic rotation giving 3D depth when a pair of red/cyan glasses were used. Objects could be visualized either as wire frame, simple box, or as full-framed, with every polygon outlined. Render quality could be chosen for wire frame, no texture mapping, full texture mapping, visual cropping, contrast, and background color. Subtle irregularities, such as zigzag contours, were smoothed with the built-in nonmetric smoothing algorithm. Missing sections due to shredding or folding on the microtome were considered by interpolation over the missing vertical standard section height from the preceding to the subsequent transection for continuous surface generation.
The second segmentation was performed with SeViSe software, implemented by the authors (Department of Computer Science VII, University of Dortmund) (Dohrmann et al., 2004), on the digital photographs, converted to TIFF. Hardware equipment was a 2.4 GHz Pentium IV, 800 MHz bus, i875 motherboard, 1 GB RAM, and NVIDIA G-force 4 FX5600 (254 MB) workstation. The SeViSe software system consists currently of three components: a module for segmentation of the histological sections, a unit for registration, and a component for visualization. The module for FEA of anatomical reconstruction is currently under development. SeViSe was created by the author group, without commercial association (no financial interest or stock ownership, thus not posing or creating a conflict of interest with the study). The sections were segmented and manually labeled on a section-by-section basis creating contours by mouse click. Automatic labeling was imprecise and therefore not used. To reconstruct a reliable 3D model, careful matching and registration of the sections is essential. The stages in this matching approach are divided in a full- and a partial-polygon-matching algorithm (Dohrmann et al., 2004). The algorithm calculates the optimal transformations translation and rotation between two corresponding point sets (form segmentation) and determines the optimal scaling by an adapted point-matching algorithm (Arun et al., 1987). Additionally, the point matching is applied to match the polygons from two adjacent cross sections (Umeyama, 1991). After matching, registration is required to handle small differences in translation, rotation, and scaling, which may be substituted by manually introduced matching point pairs. The scaling between the cross-sections is determined by means of a caliper that is placed in every image, with the same adjusted distance. A specially developed image registration algorithm is used to calculate the optimum translation and rotation between any two adjacent cross-sections (Fuchs et al., 1977; Dohrmann et al., 2004). Thereafter, the images are transferred pixel by pixel proportional to their thickness into a 3D array, properly in space. This regular 3D model is obtained by resampling the sequence of images in 3D space, using an interpolation algorithm to complete gaps. If a section was lost due to shredding or folding on the microtome, the preceding and subsequent transactions were joined for continuous surface generation to account for the missing section with averaged slice thickness. The resulting data set (approximately 2,200 × 1,700 × 500 pixels) is visualized by using direct volume rendering techniques at interactive frame rates. A 3D voxel model with a clipping box superimposed was constructed to reveal a view into the relevant structures. To visualize the segment information of the histological slices in 3D space, the anatomical segments are also interpolated and then stored as isosurfaces. All visualization is performed by combined volume and polygon rendering, offered by the Volume Graphics Library (VGL; www.volumegraphics.de). A user-friendly interface is provided by the software, which allows interactive parameterization of the 3D model. In addition, the software facilitates a shareable 3D viewing architecture to change between 3D und 2D mode.
SURFdriver crashed frequently under Windows NT and 2000 operating systems. The menu for manual segmentation was intuitive, but segmentation mode was imprecise and therefore not used (Fig. 1a), just like the (manual) object adjustment automatic (Fig. 1b). The SURFdriver software used a head-specific individual average vertical slice thickness (0.07–0.09 mm) and free best-fit manual superpositioning of sections in two dimensions. Osseous structures such as the cranial base, cervical spine, and the pterygoid process were used in congruence because they showed the least deformation from sectioning artifacts. Although a subjective assessment, only the whole section could be moved, rather than individual structures within a whole section. The quality of the reconstructed images was distinct and the spatial positions and complicated adjacent relationships of various structures could be shown in direct viewing when displayed (Fig. 2). The computerized 3D reconstruction model clearly displayed muscles and delineated their spatial relationship. For example, the 3D model contained the muscles and fasciae of the Eustachian tube and delineated the relationships between LVP, TVP, hamulus, CPS, and MU, which are the most important structures encountered during cleft palate surgery. However, vessels and nerves could only become partially aligned because of the lack of clarity of those structures in consecutive sections. Structures of secondary interest could not be included because of inadequate landmarks, unclear connections, and program crashes. Smoothing and stringency were not quantified, but nonetheless were used to create aesthetic structure surfaces, with imprecise spatial resolution.
The 3D reconstruction data of the cleft lip, palate, and Eustachian tube generated by SURFdriver software could be exported in data exchange file (DXF) standard format and transferred to computer-aided design (CAD) systems for creating rapid prototyping models. All reconstructed structures could be displayed respectively, in groups or as a whole, and could be rotated in 3D space or rotated continuously at different speeds. The time spent in displaying or rotating one reconstructed image composed from 50 sections was 1.5 sec. Any diameter and angle of the structures could be conveniently measured. Various features, such as transparency control, specific object selection, animation, 3D animation as a QuickTime VR (Nieder et al., 2000), and a variety of manipulation modes, made visualization of the complex clp anatomy feasible. However, those attributes were counterbalanced by frequent program hangups.
SeViSe crashed sometimes when frequent section changes were performed during the segmenting process (e.g., more than 10 changes in 1 min). The reason for the crashes was due to platform independence. However, this could be nearly overcome by introducing exception handling. The software permits selection of cursor size in the segmenting process and has no autosegmenting function yet, but it is under development. The segmenting menu was likewise intuitive (Fig. 3a). Because an unlimited number of structures could be integrated, we segmented up to 30 independent structures per specimen by an average of m = 90 polygons/layer. We used 1 mm diameter cursor size to avoid covering structures of interest. A reverse calculation of artifacts was performed by matching of internal landmarks such as the cranial base, cervical spine, and longus colli muscle. The autoalignment function searched for an integrated reference object and automatically scaled each section in horizontal plane (Fig. 3b). The 3D image was distinct and precise, providing details that were more elaborate than in SURFdriver. Comparatively, the whole array of anatomic structures was seen; specifically, the spatial positions and complicated adjacent relationships of important structures could be shown in direct viewing when displayed, including nerves and vessels (Fig. 5). Zigzag surfaces from insufficient rendering were not seen. The complete structures of interest as listed in the Materials and Methods section, not only the most important components, could be reconstructed (Figs. 4 to 7). The 3D reconstruction data could also be exported in DXF/virtual reality modeling language (VRML) format and transferred to CAD systems or automatic mesh generators for creating a highly accurate rapid prototyping model as well as perform FEA prospectively. All reconstructed structures could be displayed either independently or as complete conformation and could be rotated in 3D space continuously at different velocity. Compared to SURFdriver, the visual feedback of SeViSe is provided via real-time volume and polygon rendering. The time spent in loading a set of 50 slides was < 2 sec for a single slide (average load). Displaying or rotating of a 3D model, including a model created from 50 slides, is close to real time, using powerful hardware acceleration (depending on the graphic card). Any diameter and angle of the structures can be conveniently measured. Various features, such as transparency control (variable opacity), specific object selection, animation, and a variety of manipulation modes, make visualization of the complex clp anatomy possible. The spatial resolution of the model has a scale of 1:0.027, with one pixel equivalent to 0.027 mm.
Our study compared two software algorithms and their 3D virtual anatomical reconstruction because exact 3D micromorphological reconstruction may improve surgical planning and approaches for clp surgery. Transverse and frontal serial sections of the midface of 18 fetal specimens (11th to 32nd gestational week) were used for two manual segmentation approaches. The first approach used bitmap images and either Windows-based or Mac-based SURFdriver commercial software that allowed manual contour matching, surface generation with average slice thickness, 3D triangulation, and real-time interactive virtual 3D reconstruction viewing. The second manual segmentation approach used tagged image format and platform-independent prototypical SeViSe software developed by the authors. While both approaches provided useful results, software crashes and erroneous landmarks were more problematic using SURFdriver software compared to SeViSe software. Most importantly, the SeViSe software provided the most detailed and exact 3D reconstructions.
It is not yet possible to display detailed spatial relationships of the newborn's lip, palate, and pharynx in 3D images generated by virtual sectioning from CT or MRI imaging (Kleinheinz and Joos, 2001; MinGuo et al., 2004) because of limited resolution. Discrete anatomical structures cannot be delineated due to ambiguity between signal intensity and tissue composition (Aritan et al., 1997; MingGuo et al., 2004). To improve the safety and efficacy of clp reconstructive surgery, it is necessary to create geometrically accurate and detailed computer models with higher resolution and better soft tissue differentiation than those available with virtual sectioning. Both requirements were tested in the present study using aborted fetuses because their developmental anatomy resembles that which the clp surgeon faces in the operated infant at surgery. Adult cadavers, even when originally afflicted with clp, are nowadays operated on during their lifetime and therefore have a pathological condition that is very different from a young infant with clp. For our study, it was not possible to attain a bigger collection of well-preserved nonmacerated fetuses without additional malformations of identical gestational age. This is to be understood against the background that in Germany, isolated clp is not a medical or legal indication for abortion.
Our purpose was therefore intended to gain an overview of the complete fetal period. Scaled exact 3D reproduction is mandatory for the mathematical reproduction of natural and pathological muscular anatomy (Sundsten and Prothero, 1983). A 3D physical replica of clp anatomy, including age-dependent and interindividual variation, had not been constructed. Due to the workload for each reconstruction, only isolated models of the soft palate, uvula, Eustachian tube, or temporal bone have been constructed before (Latham, 1976; Long et al., 1976; Siegel and Todhunter, 1979; Latham et al., 1980; Sundsten and Prothero, 1983; Kernahan et al., 1984; Swarts and Rood, 1990; Rudé et al., 1994; Namnoun et al., 1997; Berry et al., 1999; Kober et al., 2000; Ishijima et al., 2002; MingGuo et al., 2004). Case studies, while providing generalizations about structural relationships, overlook interindividual and developmental variation.
We used the plastination and celloidin technique to provide cross-sectional images of higher resolution than those obtained from CT or MRI scanning, or sonography, with better soft tissue visualization. Both techniques were used to compare detail recognition and artifact correction not merely to apply one embedding and staining method but to be able to evaluate software algorithms in different situations. Because detail was lost with the plastination approach (Landes et al., 2005), the celloidin-hematoxylin-eosin method was relied on for the present study. The transverse cutting plane was preferred because it followed the mainly anterioposterior course of the soft palate muscles. Frontal sectioned specimens were preferred for reproduction of the facial perioral anatomy. Specimens embedded with plastination or celloidin provided essentially 2D slices, eliminating some of the ambiguity inherent when viewing the cut surface of specimens prepared using a serial milling technique. Also, using these approaches allowed the specimens to be kept for further study, which is not possible with a serial milling technique, which destroys the specimen (Sittel et al., 1997; Fish et al., 2001; Steinke, 2001). Preparations containing cranial bone and different tissue types are reported to be dimensionally stable following fixation, dehydration, and embedding [± 5% change according to McLean and Prothero (1991) and MingGuo et al. (2004)]. Entius et al. (1997) correlated sheet-plastinated slices to computed tomography and MRI images to find good correlation. Anatomical data with the highest spatial resolution can be derived from these sets. Three-dimensional information is retrieved from manual or automatic tracings of the structures, which are vertically superpositioned.
In the 1970s, outlines on opaque paper, wax, plaster (Gaunt and Gaunt, 1978) or plexiglas (Latham, 1976; Long et al., 1976; Siegel and Todhunter, 1979; Latham et al., 1980; Sundsten and Prothero, 1983; Swarts and Rood, 1990; Rudé et al., 1994) were superimposed. Computer-supported attempts began digitizing structural edges, expressing them as line reconstructions, although without surface information. During the 1980s, surface modeling and solid modeling evolved. The first attempts reconstructed an object as a shell, recognizing landmarks representing streams of landmarks, defining a viewing space by mapping a window and its coordinate contents through a scaling function. This function consisted of magnification and tissue thickness values. Geometric mapping algorithms were used to map coordinates for each contour. SURFdriver initially used that approach (Lozanoff and Diewert, 1989). SURFdriver software was programmed in the 1990s and version 4.0 will shortly appear. SURFdriver is distributed at a reasonable price and found frequent application in various 3D reconstruction projects (Spaendonck et al., 2000; Ribeiro et al., 2002; Decraemer et al., 2003). Using ready-to-use anatomical reconstruction software that needs no programming allows immediate use.
Solid modeling defines each object as a volume, after defining a viewing space. The data input is defined as voxels. Boissonat (1988) attempted to merge these two concepts (Lozanoff and Deptuch, 1991) using volumes themselves constructed around contours. Intersection planes between contours and volumes were identified and surface tiles defined (Lozanoff and Diewert, 1989; Lozanoff, et al., 1994). FEA was used in midfacial growth of mice (Lozanoff and Diewert, 1989; Lozanoff et al., 1994; Ma and Lozanoff, 1996). SURFdriver encompasses data acquisition and display with additional export functions for surface application. Although the final 3D reconstruction was distinct and accurate to our perception, weaknesses of SURFdriver were repeated instability under Windows NT and 2000 operating systems at data input. The instability could not be eliminated using multiple debugging functions. Shutdowns occurred frequently. However, once data input was complete during 3D viewing, shutdowns did not occur. Colors and transparencies were limited in number. Rendering textures have been added recently using new surface maps. The smoothing, stringency, and skew correction functions crop details and cannot be adjusted or quantified. Otherwise, the program appears comprehensible for the inexperienced user. The number of structures that can be incorporated is considerable, however limited. No individual slice thickness can be included. The biggest problem, apart from the regular shutdowns, were frequent reconstruction errors that included random landmarks appearing in the reconstruction that were not in the original data set. The authors attribute differences in segment contours from slice to slice to be the cause of that problem (SURFdriver manual; Scott Lozanoff, personal communication). Multiple manual corrections, therefore, had to be performed that finally limited the number and precision of reconstructed structures.
Our SeViSe software system was designed for use on a standard workstation and to provide simultaneously and interactively the anatomy, with all of the important 3D structural spatial relationships, which are also useful for both presurgical planning and intraoperative orientation. The quality of the reconstructed images was distinct and precise and could serve as a standard for models created with other techniques, including interindividual and growth/age-related deviation. A spatial resolution of 0.027 mm was reached, comparable to the 0.008 mm resolution provided by a microtomographic method by Spaendonck et al. (2000) that, however, does not reproduce soft tissues. Global deformation is confined to superposition of multiple structures on a section-by-section basis (Verbeek, 1996). Isolated structure deformation is biased by whole section realignment. Rudé et al. (1994) used maximum congruence for realignment, as described by Prothero and Prothero (1986) and Hibbard et al. (1992, 1993). Haas and Fischer (1997) likewise noted overall distortion of sections during processing and corrected this error in product design software by performing defined nonproportionate scaling of the contours. The authors likewise noted a relative shift of organs (blur) during histological processing so that exact alignment of consecutive sections was not possible. Insufficient adjustment resulted in surface twists, that is, noncorresponding curve points lie in morphologically noncorresponding parts of the contours. A second realignment and deformation backward calculation will prospectively be performed from MRI sections; however, MRI sections still suffer from inadequate tissue differentiation. The absolute orientation of each frame is established by a stochastic analysis based on anatomical landmarks. Episcopic pictures taken from the microtome block and alignment algorithm can be used alternatively (Laan et al., 1989; Lozanoff, 1992).
Variations in average section thickness, such as jumps in the microtome blade due to changes in tissue composition, cannot be considered by SURFdriver. SeViSe does consider those variations. Rendering and smoothing yield a defined average systematic error of 2.7% (minimum, 1.3%; maximum, 4.7%) in SeViSe (Dohrmann et al., 2004), which remained unclear in SURFdriver. Nonetheless, the attained quality in both reconstructions is far higher than in patient 3D CT or MRI (Kleinheinz and Joos, 2001; Kuehn et al., 2001).
Fine muscle detail cannot become segmented fiber by fiber, especially in the highly complex perioral region (Zhang, 1990), by either software. Large tissue volumes could become segmented at higher magnification, with overlapping optical section series using confocal laser scanning microscopy (Karen et al., 2003). However, both software solutions did not succeed in automated reliable structure rendering. While the software segmented all areas of identical color and transparence, semiautomatic copying of a manual segmentation to the next section's slightly altered contour was not possible. Nor were external fiducial markers used (Brändle, 1989; Hara et al., 1989; McLean and Prothero, 1991; Streicher et al., 1997) for realignment and deformation calculation. Preliminary artifact assessment (Landes et al., 2005) showed isolated structures to be differently deformed than the whole anatomic conformation. Therefore, external fiducials yield merely approximate deformation information because each tissue is deformed with different elasticity.
SURFdriver is distributed for Windows and Mac. SeViSe is platform-independent (Windows, Linux, etc.), distributed under the educational license of Qt (Trolltech, Norway), and can become readily completed or modified for different ventures. However, the software requires considerable programming knowledge for adaptation to a specific project. The data generated by both software solutions can be deliberately handled with standard platforms, with frequent turning, modifying, and moving and considerable surface shading and coloring. Even reconstruction of branching structures or foramina is maintained when alternative software packages are used for interactive data processing (Haas and Fischer, 1997). Export in DXF or STL format to CAD systems will create high-accuracy rapid prototypes (Kermer et al., 1998). A rapid prototyping model could valuably serve for resident education and various surgical exercises to be carried out on the 3D stereophysical model. Direct vicinity of structures in the computer model enables mesh generation and FEA of muscular influence and deformation of passive tissues. This will be performed as the next step of development of SeViSe. After simulated standard operations, the postoperative dynamics will be compared to the actual case results to assess the outcome on physiological postoperative biomechanics. Developmental growth visualization also will be established by image fusion of models derived at different gestational ages (Diewert et al., 1993).
The authors thank Dr. Karl Meller, Department of Anatomy and Cytology, Ruhr University, Bochum, Germany, for his help in generating the celloidin H&E sections as well as Dr. Heinrich Mueller, Department of Computer Science VII, Dortmund University, Dortmund, Germany, for his support in generating the SeViSe software.