3 o u ~ I of Microscopy, Vol. 138, P i 2, May 1985, pp. 221-227.
`Recaved 12 May 1984; accepied 23 January 1985
`
`Digital image tiles:
`a method for the processing of large sections
`
`by DENNIS A. SILAGE and JOAN GEL*, Department of Electrical Engrmm'ng, Temple
`University, Philadelphia, PA 19122, U.S.A., and *Department of Pathology, Mount Sinai
`Medical School, New York, NY 10029, U.S.A.
`
`KEY WORDS. Video image analyser, video image processing, digital image processing,
`morphometry, reconstruction.
`
`SUMMARY
`Segmentation of large areas of light microscopic slides into N by N fields, and each of
`these fields into M digital image tiles, allows the scanning, storage and digital processing of
`large images. Any of the original N2 fields or composites of M adjacent tiles can be recalled
`to the video display for analysis. Developed procedures for use on a microscope equipped
`with a precision scanning stage allow registration of the image coordinates (X-Y) for any
`original or composite field and the alignment of one of these fields along the depth (Z) axis
`by means of external, machined fiducial marks in serial sections. To facilitate work
`whenever unavoidable, we have incorporated methods for digital image panning and
`zooming (changes of magnification) and discuss their use and implications.
`
`INTRODUCTION
`We are currently engaged in the development of a computer-based video image
`processing system for reconstructions in light microscopy to facilitate the quantitative study
`of those non-randomly distributed structures which are not accessible to conventional
`morphometry. Our approach differs from others published in the literature (Street & Mize,
`1983) in that we store digital images of serial sections which are later recalled to a video
`display monitor overlaid with an interactive touch sensitive screen peripheral (Silage & Gil,
`1984) for selective tracing and labelling.
`We have identified several major interrelated difficulties which also occur in other
`applications of computer-based video image processing in light microscopy, particularly
`when a scanning stage is involved: (1) the image coordinates (X-Y) must be precise, (2) a
`reliable serial section (Z) alignment procedure is required, and (3) it is nearly unavoidable
`because of their size or number that some profiles will not be fully contained in an
`individual microscopic field under study. In serial histological reconstructions this difficulty
`is indeed serious because no profile or part of a profile of the structure of interest may be
`discounted. While this could in part be corrected by reducing the magnification, there are
`practical limitations to the available choices. This article describes how we have devised a
`procedure for dividing the histological section into a mosaic of small square digital image
`'tiles'. Furthermore, .appropriate procedures for serial section alignment are also described.
`
`@ 1985 The Royal Microscopical Society
`
`22 1
`
`

`
`222 Dennis A. Silage and Joan Gil
`
`I N S T R U M E N T A T I O N
`We used an image analyser system described by Silag & Gil (1982, 1984). Leitz
`Orthoplan microscope equipped with a servomechanical scanning stage (Model 050-1 12)
`which moves 2.5 p m per increment, 500 increments per second is interfaced to a Digital
`Equipment Corporation (DEC) PDP-11124 minicomputer system. The PDP-11/24 has a
`dual 10.4 megabyte cartridge disk (RL02) mass storage device and is configured under the
`DEC RSX-1IM operating system. The minicomputer system includes the DEC VS-11
`video graphical display system which, in addition to its conventional use for high resolution
`graphics, can combine and synchronize one external video source. The video frame store,
`capable of digitizing and storing a single analogue video frame in an internal digital
`memory, is a Colorado Video model 274C. A high resolution black- and-white analogue
`video scene, derived from a Panasonic WV1800 video camera mounted on the microscope,
`is digitized by the video frame store to a maximum of 480 vertical by 512 horizontal picture
`elements (pixels) with 64 grey level resolution (6 bit) per pixel. The video frame store also
`can recompose an analogue video frame from its internal digital memory for display on a
`video monitor.
`
`HISTOLOGICAL SECTION A L I G N M E N T A N D T R A N S L A T I O N
`The precise alignment of serial histological sections and the proper translation of image
`field coordinates is crucial. Before aligning the sections in our instruments we have devised
`a number of routines to examine servo-mechanical accuracy. The registration of the
`movement of the servo-mechanical scanning stage and the high resolution video system is
`verified by imaging a target with an orthogonal grid (histological slide with a microscopic
`graticule) in real time combined with another identical orthogonal grid generated by the
`video graphical display system of the computer. The registration of the two grids is
`observed while the microscope stage is manually translated over its range of 2 cm.
`The stage is originally equipped with a manual joystick which provides the horizontal
`and vertical servomotors with translational command signals. Automatic scanning is
`achieved by simulating these command signals by digital signals from the minicomputer.
`We calibrated the signals required to translate the stage exactly 200 pixels horizontally and
`vertically by using a stage micrometer.
`The alignment between consecutive sections in the depth or Z axis direction is based on
`the ability to combine two high resolution, black-and-white video signals in a video mixer
`circuit of our own design (Silage, 1982). A low objective power digital image of fiducial
`marks of the preceding histological section is recalled from archival disk storage into the
`video frame store. Fiducial marks are needed because we have observed in the lung that
`alignment over natural internal structures, such as the pleura, leads to an easily
`recognizable 'sideways shift'.
`External markers have been used in embryological three-dimensional reconstruction
`since the late nineteenth century (Burston & Thurley, 1957; Ware & LoPresti, 1975;
`Walmsley, 1983). We developed in our laboratory techniques for embedding large blocks of
`tissue ( 2 ~ 2 ~ 0 . 5 cm) in a hard plastic material (Quetol 651) intended for sectioning in a
`
`rotary microtome equipped with Ralph knife. This plastic material, unlike paraffin, is
`capable of being machined prior to sectioning to provide easily discernible fiducial marks
`(McNiff & Gil, 1985) (we use an F80 bit with a diameter of 0.0135 inches (0.343 mm), at
`less than 2 mm intervals in a triangle). The holes are nearly circumferential on the section
`and are easily discerned if they have been touched by a highlighter pen prior to covering.
`Because of the dispersion and size of the fiducial marks we use the lowest objective
`power to produce an image that is nearly 2 . 5 ~ 2 . 5 mm. The current histological section is
`then micromanipulated until the two analogue video signals, one from the video frame store
`and the other from the real-time video camera mounted on the microscope, are aligned.
`
`

`
`Digital image tiles
`
`223
`
`RETR I EURL +
`
`Fig. 1. The segmentation of an image into N Z fields and each field into M digital tiles is depicted for N 2 = 4
`and M = 4 . The original four fields (Fl-F4) are each composed of four tiles (TI-T4). A video display
`retrieved from digital storage may be a combination of any four adjacent tiles. A single digital tile can be
`zoomed or any adjacent set of four fields can be panned to comprise the display.
`
`IMAGE S E G M E N T A T I O N A N D COORDINATES
`Our goal was to first achieve the segmentation of a large scene defined as a preselected
`area of interest in a histological slide into N by N fields for digital storage and the
`subsequent segmentation of each field into M digital tiles to facilitate their recall in
`appropriate groups to form a composite field (Fig. 1). A digital field is that segment of the
`image which, at a given magnification, will comprise the video display without manipula-
`tion. Our approach is to regard each digital field as a 'window' of 400 horizontal by 400
`vertical pixels. Table 1 lists the measured size of the digital field and the resolution of the
`digital image for several dry objectives available on the Leitz Orthoplan microscope in our
`laboratory. We have chosen for our current application to define a digital tile as a quadrant
`(M=4) of the complete but composite digital field, which implies that the tile is exactly 200
`by 200 pixels, but alternate definitions of a tile (for example, 100 by 100 pixels with M = 16
`or 80 by 80 pixels with M=25 such tiles comprising a complete field) are feasible.
`
`Table 1. Measured size of a digital field and the resolution of the digital image
`for the dry objectives in our laboratory.
`
`Object
`2,5
`6.3
`16
`
`Field (mmlside)
`2.48
`1.15
`0.395
`
`Resolution (pixelsicm)
`1603
`3478
`10127
`
`

`
`224 Dennis A. Silage and Joan Gil
`
`Fig. 2A. Four selected digital image tiles at the confluence of four originally adjacent digital fields, each
`acquired and stored separately. Compared this display with that of Figs. 2B, 2C and 2D.
`
`Fig. 2B. The original display of the same area of Fig. 2A before digitization and segmentation.
`
`After alignment and coordinate determination a program controls the scanning of the
`histological section to acquire and store a complete image of N by N fields, consisting of
`MN2 digital tiles with an identification of their coordinate origin. The acquisition and disk
`storage of a single digital field after the scanning stage has been automatically positioned
`requires 32 s on our instrument.
`During subsequent image analysis the retrieval from disk storage and video display of a
`single digital field, as depicted in Fig. 1 and shown in Fig. 2A, again requires 32 s.
`
`

`
`Digital image tiles
`
`225
`
`Fig. 2C. Four times digital magnification (zoom) of the lower left digital tile of Fig. 2A.
`
`Fig. 2D. One quarter size display (pan) of the four original digital fields in this example.
`
`Compare this with Fig. 2B, which shows the original display of the same area before
`digitization and segmentation. Differences in the background were emphasized to visualize
`the location and coordinate alignment of the digital tiles.
`While any original or composite digital field at the preselected magnification can be
`retrieved, the instrument also has the capability of panning and zooming the digital image.
`in zooming, a digital tile is magnified by pixel replication, that is each pixel is initially
`reproduced an even number of times as a square array of new but identical pixels. Although
`
`

`
`226
`
`Dennis A. Silage and Joan Gil
`
`the resultant digital image has abrupt controls, it can be smoothed by a spatial low-pass
`filtering template, analogous to the processing that occurs in gradient edge detection (Hall,
`1979; Rosenfeld & Kak, 1982). After the retrieval of a single digital tile in 8 s, the zoom
`operation executes in an additional 26 s on our instrument.
`A digital image is panned by pixel averaging, that is the grey scale values of an even
`number of pixels in a square array are averaged to produce a single, new grey level and
`pixel. A prerequisite of the pan operation is the recall of four digital fields from disk storage
`which requires 128 s. The pan operation then executes in an additional 14 s.
`
`DISCUSSION
`Our development provides the investigator with flexibility in locating and digitally
`processing structures in large histological sections. The image segmentation and coordinates
`allows the digital storage and retrieval of a composite scene much larger than that possible
`at the magnification chosen. The method of digital image tiles allows us to select the field
`where the tracing of the whole structure is feasible. When attempting partial tracings on
`different fields does not seem practical, we have the possibility of panning of the image.
`Finally, zooming will be helpful in the unavoidable instances where small structures are to
`be traced.
`Although digital panning and zooming are additional options intended only for
`occasional use, the possibility of digital changes of magnification raises questions of
`considerable theoretical interest. In particular, problems can arise due to the so-called 'coast
`of Britain' effect, which in fact means that for rugged outlines, boundary length
`measurements depend on the magnification (Paumgartner et al., 1981; Rigaut, 1984). In
`panning the reduction of magnification evidently results in some loss of resolution;
`zooming, on the contrary, does not increase the resolution, while the digital image
`processing procedure yields an artificial roughness of its own. This complex problem would
`require analysis along the principles of fractal geometry (Mandelbrot, 1982). The possibility
`of tracing and measuring at different magnifications is offered only as a last resort and it
`cannot be recommended for routine use until the effects of any digital image smoothing
`template with respect. to linear resolution have been analysed.
`The procedures developed here for section alignment, image segmentation, and
`coordinate measurement are being applied to the serial reconstruction of lung parenchymal
`tissue (Silage & Gil, 1982). In this image analysis the serial contours of these non-randomly
`distributed structures are more readily identified when several histological sections are
`reliably aligned and available for recall and display. The use of an interactive touch sensitive
`screen peripheral for tracing and labelling facilitates the analysis. The details and
`documentation of our instrument are readily available on request.
`
`A C K N O W L E D G M E N T S
`The authors were supported by the National Heart, Lung and Blood Institute grant
`HL-26676. We acknowledge the assistance of Miss Judith M. McNiff with histology and
`photography and Mr Daniel C. Barrett in the preparation of the manuscript.
`
`R E F E R E N C E S
`Burston, W.R. & Thurley, K. (1957) A technique for the orientation of serial histological sections. J. Anat.
`91, 409412.
`Hall, E.L. (1979) Computer Image Processing and Recognitton. Academic Press, New York.
`Mandelbrot, B.B. (1982) The Fractal Geometry of Nature. Freeman, San Francisco.
`McNiff, J. & Gil, J. (1985) Serial sections of Quetol-embedded osmium-fixed lung with external reference
`.
`marks for alignment. Stain Technol. 60, 38-42.
`Paumgartner, D., Losa, G. & Weibel, E.R. (1981) Resolution effect on the stereological estimation of surface
`and volume and its interpretation in terms of fractal dimension. J. Microsc. 121, 51-63.
`
`

`
`Digital image tiles
`
`227
`
`Rigaut, J.P. (1984) An empirical formulation relating boundary length to resolution in specimens showing
`'non-ideally fractal' dimensions. J. Microsc. 133, 41-54.
`Rosenfeld, A. & Kak, A.C. (1982) Digital Picture Proresing. Academic Press, New York.
`Silage, D.A. (1982) Microcomputer image cellular morphometry. Etlgi~~eenng 111 Heulrh Cure ( I E E E ~ E M B S I ,
`4, 152-154.
`Silage, D.A. & Gil, J. (1982) Digital image analyzer for the morphometric reconstruction of biological tissue.
`Medlcal Computer Sciences ( I E E E I C S ) , 1, 456-458.
`Silage, D.A. & Gil, J. (1984) The use of a touch-sensitive screen in interactive morphometrv. J. hl~irosi..
`134, 315-321.
`Street, C.H. & Mize, R.R. (1983) A simple microcomputer-based three-dimensional serial section
`reconstruction system. J. Neurosc~. Methods, 7, 359-375.
`Walmsley, J.G. (1983) Vascular smooth muscle orientation in straight portions of human cerebral arteries. 3.
`M~crosc. 131, 361-375.
`Ware, R.W. & LoPresti, V. (1957) Three-dimensional reconstruction from serial sections. Inr. Hezl. (,ystol.
`40, 325440.