2.1 Preparation and Staining of Mohs Surgical Skin Excisions
Skin excisions from Mohs surgery were obtained from our collaborating Mohs surgeon in the Dermatology Service at Memorial Sloan-Kettering Cancer Center, following approval from our Institutional Review Board. During surgery, each excision is processed by a Mohs technician and 3 to 5 frozen histology sections are prepared, each section being ~5 to 6 mm thin. After the Mohs surgeon has examined the frozen sections, the remainder of the excision is discarded. These discarded excisions were collected. Often, triangular-shaped excisions of normal skin are made adjacent to the crater-shaped wound that remains after removal of the tumor. Such excisions are necessary for cosmetically efficient suturing and closing of the wound. These excisions are called “dog-ears” because of their shape. Fresh normal dog-ear skin excisions were also collected soon after they were discarded.
The collected Mohs excisions were frozen in the embedding medium that is used to prepare histology sections. Each excision was thawed, rinsed in isotonic saline solution for 30 s, and placed in a solution of acridine orange (Fluka, 01660) in saline for various times and concentrations. Concentrations of 0.3 mM and 1.0 mM were tested for staining skin excisions by immersing for 5 s, 20 s, 1 min, 3 min, and 5 min. Five excisions were tested for each concentration and time, resulting in a total of 50. After staining, excess dye was rinsed from the dermis by dipping in isotonic saline solution for 2 to 3 s. The excisions were mounted in a custom-designed tissue fixture and imaged with a confocal microscope.
Fifty Mohs surgical excisions of BCCs were imaged in the process of optimizing the staining procedure. The images were visually evaluated, and the minimum time was determined for brightening of nuclei and detectability of BCCs. The evaluation was limited to only the exposed surface of the excisions since the Mohs surgeon mainly needs to determine the lateral spread of tumor on the excised surface. Concentration of 1 mM and immersion time of 20 s was chosen as the optimum for rapid staining and strong fluorescence. After this preliminary feasibility test, for further instrumentation development, excisions were stained with 1 mM acridine orange for immersion time of 20 s. Subsequently, 50 more excisions were similarly stained and imaged for the clinical comparison-to-histology work.
2.2 Diffusion of Contrast Agent into Tissue
Only the excised tissue surface is examined in Mohs histology, to determine the lateral extent of tumor. Thus, for imaging, staining with acridine orange was necessary only superficially through a few cell layers. A cell layer in skin is about ~10 mm thin, and the Mohs surgeon usually examines 3 to 5 frozen histology sections, each being ~5 to 6-mm thin. Hence, we consider imaging to a maximum depth of ~30 mm, which corresponds to ~3 cell layers. (This depth is also approximately the maximum to which real-time confocal imaging is possible in dermal tissue with very low milliwatt power blue 488-nm illumination.)
Rapid staining depends on the diffusion of the fluorescent dye into the excised tissue. The uptake kinetics of any particular dye depends on molecular weight and tissue conditions such as pressure in interstitial spaces.11
The average time t
(s) for diffusion across a distance x
(cm) may be determined from the diffusion coefficient D
The diffusion coefficient varies with molecular weight Mr
in tumor and normal tissue12
by the following power law:
Experiments by Nugent and Jain12
determined the coefficients to be a
=−2.96 for normal tissue, and a
=2.51 × 10−2
=−1.14 for tumor (VX2 carcinoma, rabbit). For acridine orange, the diffusion coefficient is then calculated to be 15.2 × 10−3
/s) in normal tissue and 24.5 × 10−3
/s) in tumor, assuming similar order-of-magnitude parameters for normal skin and BCC tumors. Using x
m in Eq. (1)
, the average diffusion time is 0.6 (ms) for tumor and 0.37 (s) for normal tissue.
2.3 Confocal Microscope and Tissue Fixture
shows the experimental setup. The confocal microscope was described in detail in our earlier reports on reflectance mosaicing.1,3,4
Briefly, the original reflectance microscope (VivaScope 2000, Lucid, Inc., Rochester, New York) was modified by incorporating an argon-ion laser for fluorescence excitation at 488 nm and a corresponding fluorescence detection channel. The illumination of tissue, is with low-level power of 0.3 to 1.0 (mW). The fluorescence detection optics consists of a dichroic beamsplitter (Chroma, 510DC-SPRX), an excitation rejection filter (Omega Optical, 510EFLP) to block extraneous reflected light, and an avalanche photodiode (Perkin-Elmer, Quebec, Canada, C30659-900-R8A). Detection in the fluorescence channel was mainly of the emission from the acridine orange–stained nuclei, with almost none from the cytoplasm. Autofluorescence from the dermis is a few orders of magnitude weaker than the fluorescence from acridine orange and hence was not detected for the low illumination power and tightly confocal (i.e., small pinhole) detection conditions.
Fig. 1 Optical design of the confocal microscope with the following components: argon-ion laser, 7× beam expander (B/E), 488-nm-excitation selection filter (F1), spinning polygon (P), galvanometric scan mirror (G), relay telescopes (T), and objective (more ...)
A custom-designed water immersion objective lens (StableView, Lucid, Inc.) was used for imaging through a 1-mm-thick glass slide. Instead of thin coverslips that are conventionally used with objective lenses, a thick glass slide is necessary for mounting and stabilizing unconventional tissue specimens such as surgical excisions. The objective lens features 30× magnification to provide a 430-μ
m field of view. With a numerical aperture (NA) of 0.9, the calculated axial section thickness is: Δz
m and the lateral resolution is: Δx
=0.46 λ/NA=0.25 μ
m, which is adequate for imaging nuclear morphology. As previously explained,4
water was often substituted with a water-based gel as an immersion medium.
Surgical excisions are often thick, large, and of unusual shape. Furthermore, the tissue is fresh, living, hydrated, and mechanically compliant and hence not easy to mount in a microscope. A custom tissue fixture was engineered for Mohs surgical excisions to be mounted and gently compressed onto a microscope slide. Design details of the original tissue fixture are in an earlier report.4
The operation requires a threaded piston to be tightened for gently compressing and embedding in a gel disk, so as to stabilize the excision. However, the rotational motion of the piston caused the gel and subsequently the edges of the tissue to twist and distort. Such distortion is now prevented with a design modification that includes placing a thin polycarbonate disk and needle-roller bearings (Part No. 5909K31, McMaster-Carr, Dayton, New Jersey) between the piston and the gel disk. The fixture allows repeatable and accurate control of the flattening, tip, tilt, sag, and stability of the tissue surface to be imaged. The functionality of the tissue fixture mimics the operation of a cryostat, which is the standard equipment for preparing frozen histology sections for Mohs surgery. Imaging in reflectance was used to guide z
distance and tip and tilt adjustments such that the tissue surface was oriented exactly parallel to the focal plane of the objective lens. The process involved translating the mounted excision laterally and adjusting four thumbscrews until the focus moved along the reflective water/tissue interface. This alignment enabled acquisition of images contiguously over large areas of 10 to 20 mm. (Engineering drawings, manufacturing methods, and operation details are available to researchers who may be interested.)
After the excision is mounted in the tissue fixture and properly positioned and oriented, confocal images were acquired. Images were acquired of the surface of the excision. Because of the thawing, staining, and rinsing process, small distortions in the imaged surface were expected due to the compliance of the tissue. Thus, the mosaics were expected to show a close but not an exactly one-to-one correspondence to the frozen sections that were prepared by the Mohs technician during surgery.
2.4 Acquisition of Images
The small field of view of the 30× objective lens (430 μm) and the large region of interest (~10 to 20 mm of excised tissue) necessitated the construction of mosaics. A mosaic to display a large area is created by stitching together a two-dimensional (2-D) matrix of confocal images. Mosaics allow observation of large fields of view without sacrificing resolution. The matrix of images was acquired with a continuous step-and-capture routine while translating the tissue fixture with stepper motors–driven linear XY stages (Hayden, Inc., Stamford, Connecticut). An overlap of 10% was included in the translation step distance to correct for a field curvature–induced artifact in the image, as is further explained later. The overlap also prevents loss of detail at the edges between images. The amount of overlap determines the number of images in the matrix that must be acquired, which then determines the total time of acquisition.
Of the 3 to 5 frozen histology sections that are prepared during surgery, the Mohs surgeon usually examines the first to determine the lateral spread of the tumors on the excised surface. Occasionally, if the quality of the first section is poor or if the determination of tumor margins is not very clear, the Mohs surgeon will examine the remaining sections. Additional sections are sometimes prepared if the Mohs surgeon needs to further examine deeper layers of tissue. For this study, however, images were acquired and mosaics created only of the exposed surface of the discarded excisions. This exposed surface corresponds to the last Mohs frozen section. (Subsequent comparison of the mosaics to histology was therefore limited to the last frozen section.)
Confocal mosaics can be quickly acquired. For acquisition, a continuous step-and-capture routine requires about 5 min for 36 × 36 images. Transferring and archiving images followed by processing to create a mosaic requires another 4 min on another PC. Thus, total time to create a mosaic is up to 9 min at present.
2.5 Calibrations for Illumination Artifacts and Vignetting
At the edges of images, dark bands due to field curvature–induced artifact and illumination falloff due to vignetting were noticeably seen. These were corrected for in the image-stitching algorithm, based on calibration measurements in the confocal microscope.
The Petzval field curvature in the microscope was calculated to be ~3.8 μm and measured to be ~5 μm. When focused at the surface of the excision, field curvature results in tissue being seen in the center but surrounded with an annular ring at the periphery in the image. The ring is due to the overlying glass window and results in an illumination artifact. The artifact appears as bright bands in reflectance but dark bands in fluorescence. By focusing slightly deeper than 5 μm beneath the tissue surface, the artifact was often minimized. As a result of deeper focusing, small mismatches to the frozen section of the surface and small losses in correlation to frozen histology were anticipated. The dark bands were largely eliminated by cropping the 10% overlap between images in the image-stitching algorithm.
To characterize vignetting, the illumination falloff across the field of view was measured with a standard fluorescent target. A drop of acridine orange was compressed between a microscope slide and coverslip, and an axial stack of images was acquired. The images were averaged in depth to determine the vignetting in both x and y directions (). The vignetting was corrected for with an inverted-brightness polynomial fit in the image-stitching algorithm.
Fig. 2 Calibration measurements showing: (a) illumination falloff due to vignetting in the x and y directions across the field of view, and (b) 3 × 3 mosaic of a reflective grating target indicating angular alignment and lateral registration to within (more ...)
2.6 Calibrations for Angular Misalignments and Lateral Registration
The translation of the linear XY stages must be parallel to the x and y directions of the optical raster scan in the confocal microscope. Mosaics of a reflective grating test target were used to calibrate for angular misalignments. shows a mosaic of a reflective grating target (Ronchi ruling with 200 lpi, Edmund Industrial Optics) that demonstrates angular alignment and lateral registration in both x and y directions. The mosaic was created by cropping the 10% overlap between images and stitching 3 × 3 images. The grating lines appear continuous to within 5 pixels between images. However, as explained here, full-size mosaics are scaled down by 8 to 10× such that the lateral mismatch is within a subpixel and not noticeable in the final display.
2.7 Image-Stitching Algorithm
Mosaics were created with MATLAB software (version 7.4, MathWorks, Natick, Massachusetts). The algorithm implements the following steps: cropping of the overlap between images, merging of images into a single composite mosaic, and correction for the residual dark bands between images. The amount of cropping was 10%, as predefined by the stepping distance of the XY-translation stages during the image acquisition, and further precisely adjusted by measurements of overlap using image analysis software (IPLab Spectrum, version 3.6.5, BD Biosciences, Inc., Rockville, Maryland). Based on our experience, the overlap between images remained repeatedly consistent across large mosaics with minimal errors. After cropping, the images were concatenated into a single composite mosaic. The cropping removes the dark bands due to field curvature. The illumination falloff due to vignetting was then corrected for with inverted-brightness correction polynomial fits in both x
directions. The polynomials were empirically designed to flatten the illumination falloff across images. The design of the polynomials is specific to our fluorescence mosaics but is based on a destripe filter that was originally authored by Marc Lehman and is available as open-source software called GNU Image Manipulation Program (GIMP). Lehman's executable and source code can be downloaded at www.GIMP.org
The pixel gray scale or brightness is defined as I(x, y), where x and y are column and row positions, respectively, of individual pixels. Horizontally across the entire mosaic, the mean brightness profile Ī(x) is determined by averaging pixel values in columns as a function of x-pixel location. Similarly, vertically across the entire mosaic, the mean brightness profile Ī(y) is determined by averaging pixel values in each row as a function of y-pixel location. These are:
where N is the total number of pixels (or rows) in each column,
where M is the total number of pixels (or columns) in each row.
Averaging across columns and rows of the entire mosaics provides a globally smoothed low-frequency estimate of the spatial high-frequency variations in fluorescence from the tissue. The mean brightness profiles in Eqs. (3)
represent both the fluorescence signal from the central regions of the images and the illumination falloff at the edges. Polynomial fits for brightness in the x
direction and the y
direction are further modeled in terms of a rolling average of the mean brightness profiles:
In Eqs. (5)
, the rolling average–based polynomial fits are locally smoothed versions of the mean brightness profiles. The rolling average polynomial fits, too, represent both the fluorescence signals from the central regions and the illumination falloff near the edges of the images. As will be evident from further equations later, the purpose of these polynomial fits is to spatially isolate the regions of low-frequency illumination falloff near the edges of the images from the relatively high-frequency fluorescence signals in the central regions. Isolating the two regions subsequently allows corrections to be applied mainly to the illumination falloff, but none to the fluorescence signals. The parameter W
is the window width of a rolling average filter. The width of the filter W
was empirically tested in the range of 24 to 100 pixels. On the basis of visual examination, a small width of 36 pixels was found to provide the optimum match of the rolling average polynomial fit to the mean brightness profile. Smaller windows produced too little smoothing of the differences between the spatially high-frequency fluorescence signals in the central regions of the images and the relatively low-frequency illumination falloff near the edges, and therefore provided inadequate isolation of the two regions. This resulted in (unwanted but) minimal corrections to the fluorescence signal but too little correction for illumination falloff. Larger windows resulted in too much smoothing and, again, inadequate isolation, which resulted in too much correction of illumination falloff and also unwanted large corrections to the fluorescence signal.
Subtracting the mean brightness profiles from rolling average polynomial fits substantially removes the fluorescence signals from the central regions and mainly represents the illumination falloff near the edges. This subtraction results in an inverted-brightness correction polynomial that is determined to be:
The correction polynomials C(x) and C(y) are minimal in the central regions of images of the mosaic, which are relatively uniformly illuminated. Visual examination showed that, with the rolling average filter width of 36 pixels, the correction polynomials were close to zero. However, they were appropriately large near the edges that display illumination falloff. Thus, the correction polynomials are applied mainly in the regions of illumination falloff but not in the central regions of images.
The inverted-brightness correction polynomials are added back to the original mosaic to flatten the illumination falloff across the images. The illumination falloff at the edges between images are corrected in the y direction (row by row) and in the x direction (column by column) to produce a new brightness or pixel grayscale distribution defined by:
In Eqs. (9)
is the width of rolling average filter (as before), and D
is a scaling factor or “gain” for image brightness. Since the corrections are based on a rolling average approximation of the actual pixel brightness distribution, a scaling factor is necessary to adjust the brightness in the regions of fluorescence signal and vignetting to appropriate levels. The scaling factor D
was empirically tested in the range of 8 to 128. On the basis of visual examination, a factor of 32 was determined to provide the optimum brightness for the mosaics. Smaller scaling factors result in dark-appearing mosaics, whereas larger scaling factors result in too much brightness and saturation. To every row (or column) of pixel gray scales I
), the inverted-brightness polynomial fits are applied proportionally to the locally averaged fluorescence signal, as described in Eqs. (9)
. This approach minimizes the unnecessary corrections in relatively uniformly illuminated central regions of images and limits corrections mainly to the illumination falloff near the edges.
Based on our experience, this algorithm works effectively to correct the well-defined loss of illumination at the edges of images due to vignetting. The algorithm was applied to the 50 mosaics that were acquired to achieve repeatable results. The advantage of this algorithm is that the correction polynomials may be determined in a “blinded” manner to any given mosaic without requiring a priori
knowledge of vignetting in the microscope. The mean brightness profiles and rolling average polynomial fits [Eqs. (3)
] produce an estimate of the illumination falloff due to vignetting. Any additional instrumental errors in illumination are also estimated and corrected for. (An alternative is to directly fit polynomials based on instrument calibrations such as shown in . This may require frequent alignment and measurements of the microscope—a task that is easily performed in laboratory settings but not always in a clinical setting.) The main requirement is that values for W
must be initially determined in two to three mosaics. We have found that the values may be consistently used afterwards. The disadvantage of this algorithm, however, is unwarranted decrease or increase of contrast in tissue morphologic features, especially at the edges of BCC tumors, due to the unnecessary corrections produced by the smoothing effect of the rolling average model. However, the corrections are close to zero in the fluorescence signal such that visual gain or loss of contrast appeared to be minimal and did not appear to affect subsequent analysis of mosaics and comparison to histology. The final processed mosaic is saved in TIFF format. (The algorithm and MATLAB software, including implementation details, are available to researchers who may be interested.)
2.8 Display of Large-Area Mosaics
Each image consists of 640 × 480 pixels, is 8-bit gray scale, and requires about 1/4 MB of memory. Thus, a full mosaic of up to 36 × 36 images consists of 23,040 × 17,280 pixels and requires 325 MB of memory. The mosaics are scaled down using bilinear interpolation to make the displayed lateral resolution and pixelation equivalent to that of a 2×-magnified view of histology. The scaling down of the optical resolution and pixelation to match that of a 2× view was explained in detail in our earlier report on reflectance mosaicing.4
The final mosaic is displayed to the Mohs surgeon with lateral resolution of ~4 mm, consists of ~2500 × 2500 pixels, and requires less than 4 MB.
Mohs excisions are sometimes oval or elongated in shape and may be as long as 30 mm. For such excisions, larger mosaics displaying up to 30 × 10 mm were created by acquiring multiple adjacent 10 × 10 mm mosaics and joining them in Adobe Photoshop.
We observe mosaics on a 30-in. flat-screen monitor (Dell 222-7175 with a GeForce 8800 GTS video card) of 2500 × 1600 pixels. A full-size mosaic is equivalent to a 2×-magnified view. With digital zooming, smaller portions, called submosaics, are also observed. The display of submosaics is equivalent to higher magnifications of 4× and 10×.
2.9 Comparison of Mosaics to Frozen Histology
Fifty mosaics were compared to the corresponding Mohs frozen histology sections. The frozen sections were those that were prepared during surgery for the Mohs surgeon. These sections were prepared with standard hematoxylin-and-eosin (H&E) stains. The imaged surface of the frozen-thawed discarded excision corresponds to the final section that was prepared. Therefore, mosaics were compared to the last Mohs frozen section.
Nuclear and morphological features were evaluated in the mosaics and compared to the histology. The Mohs surgeon evaluated those features that are routinely examined in histology and are necessary to detect BCC tumors versus normal skin. The features for the presence of BCC tumors are nuclear pleomorphism (atypical shapes and sizes), increased overall nuclear density, palisading (“picket-fence” type arrangement of nuclei around the inner periphery of tumor), clefting (dark-appearing “moats” around the outer periphery of tumors that are filled with optically clear mucin), and the presence of inflammatory infiltrates. The features for normal skin are epidermal margin (epidermis along half the periphery of the excision), hair follicles, sebaceous glands, and eccrine glands.
The Mohs surgeon uses an objective lens with 2× magnification to quickly examine large areas of histology. Objective lenses with 4× and 10× magnifications are used, when necessary, for closer inspection of nuclear detail. Entire mosaics and submosaics were evaluated at equivalent magnifications. The submosaics usually consisted of a quarter of the entire mosaic, since the Mohs surgeon usually records the presence or absence of BCC tumor in quadrants.
2.10 Effects of Acridine Orange Staining on Subsequent Frozen Histology
Experiments were performed to investigate the effects of exposure for 20 s to 1 mM acridine orange on subsequent histology. Possible effects include tissue decay, degradation of RNA and DNA due to autolysis, room-temperature digestion of tissue due to proteolytic enzymes, swelling of cytoplasm, and separation of the epidermis from the dermis. These effects may lead to prevention of H&E staining and a “washed out” appearance. Ten excisions were reprocessed for frozen H&E–stained histology sections after acridine orange staining and confocal imaging. The frozen sections were compared to the corresponding Mohs sections. The evaluation and comparison, performed under “blinded” conditions by the Mohs surgeon, analyzed possible tissue disruptions and distortions as well as the chromaticity of staining.