1. Introductory remarks

At present the gold standard for diagnosing and monitoring cancer relies on a subjective visual microscopic analysis of excised samples of the patients’ tissue (biopsies) by a trained histopathologist. In general, positive patient outcomes correlate strongly with how early significant tissue abnormalities can be detected, and this in turn largely depends on the degree to which the cytological and architectural changes in the biopsy samples can be graded into an accepted range of normal and malignant classes with expediency and accuracy. As excised, a biopsy sample is essentially colourless, and the current “gold standard” of histopathological analysis employs a tissue staining process to visualise the microscopic tissue structure. Most commonly, in the “H+E” protocol, the two dyes Haemotoxylin and Eosin are used, and they stain the nuclear and cytoplasmic components of the biopsy specimen blue and pink respectively.

The technique of Hyper-Spectral imaging (also known as generating a “spectroscopic image”, SI) involves combining a series of images of an object taken at a range of different electromagnetic (EM) wavelengths. In the mid infrared (mid-IR), corresponding to wavelengths between approximately λ~3μm and λ~15μm, the molecules in tissue absorb radiation by a well-understood mechanism involving the creation of vibronic excitations that are localised to specific chemical bonds. Each chemical moiety absorbs with a well characterised strength and wavelength.In contrast to visible and near-IR hyperspectral images (which work with a mix of electronic and vibrational overtone absorption) the mid-IR is absorbed in a linear way, so that a transmission-mode image of a thin tissue section taken at the appropriate mid-IR wavelength can be processed to generate a quantitative 2D distribution map of the corresponding chemical moiety.The technique has been applied to a diverse range of biological and pathological problems, and there is a large literature encompassing quantitative analysis and classification of normal and tumour tissue as well as more diverse applications in drug discovery, forensic science, proteomics and homeland security [1].

Typically a SI is acquired by measuring, often with a Fourier Transform Interferometer linked to an IR microscope, the sample’s mid-IR absorption spectrum. Data is averaged over a point, or sometimes, in the case of systems with multi-channel detectors, a small patch containing a number of points [1] and the SI is built up by mechanically scanning the point/patch across the sample. With this approach, increasing the image’s spatial resolution by a factor q means increasing the number of spectra taken by ~q², whilst also shrinking the analysed area in a way that decreases the IR signal by ~q⁻², so that maintaining the signal-to-noise (SNR) in the image means increasing the scan time by a factor ~q⁴. The result is that achieving the sort of spatial resolution required for visual histopathology analysis usually requires data acquisition periods that take from tens of minutes to several hours. The resulting data sets are very large and sophisticated statistical data processing methods have been developed that reduce the data to images taken at only a few specific wavelengths.

These wavelengths are chosen according to statistical criteria that measure the degree to which absorption data at that wavelength correlates with differences in the pathological state of the tissue [1], that is they are chosen to maximise the ability of the SI technique to be able to discriminate spectrally between different tissue types. The drawback here is that the actual physical mechanism that results in a particular wavelength emerging from the statistical analysis is unknown. This means that, although the reduced data can be used to generate false colour images, even though the information in these images originated from quantitative chemical differences across the sample, there is no way for an expert clinician to interpret the contrast in the image in terms of their own biochemical and medical knowledge; they can work only with the morphology in the picture.

Here we adopt a new approach. Rather than collecting large amounts of SI data, then discarding the majority of it in a “hypothesis free” way, we start by devising a stripped-down imaging protocol that tries to mimic the accepted H + E protocol but uses IR spectroscopic methods to replace the wet chemistry. We start from a knowledge of the origins of features in the IR tissue spectra, we restrict our data to only a couple of chemical moieties (analogous to the two dyes used in the H + E protocol), and we develop a technology that aims to augment the accepted H + E process by yielding additional quantitative chemical information about chemical differences between different regions of the biopsy samples.

The long term aim is a bench-top instrument which is sufficiently economical, and quick and simple to use, that it can find a place in the clinical environment. Similar to the existing chemical staining protocols, this new “digistaining” method is tightly specified and, compared to current lab-based SI analysis, somewhat inflexible. However, we see this simplicity as a strength. It is quick and yields results that are potentially more useful and understandable to clinically trained personnel.

We believe that it will allow a wide range of comparative IR-based trials to take place, where the clinicians are able to make professional judgements that are, for the first time, based on both the morphological and the spectral information, in a way that is difficult with the often bewilderingly large data sets that are generated by the more technically complex laboratory SI systems.

The key components of this imaging system as well as the samples used for the study are as follows:-

2. Mid-IR imaging system

2.1 Optical setup

The illumination source (Fig. 1 ) is a standard IR Globar (LotOriel Model 6363) IR emitter. Although the low specific brightness of these sources is the principal reason for the long data acquisition times in scanning SI systems, they are better suited for our purposes because ouroptics use a much larger fraction of their optical output than does a scanning system. The central Globar segment is a roughly 6mm x 20mm area of sintered silicon carbide, electrically heated to 1050 K. Its black body emissivity is 70% ± 10% over the 1<λ<25µm wavelength range, giving an irradiance that varies between 5x10⁻⁶ – 2x10⁻⁵ W m⁻² nm⁻¹ .

 

Fig. 1 Schematics of the optical setup. A Barium Fluoride condenser (f = 37.5mm) lens was placed between the Globar element and sample as shown to provide a sample illumination as uniform as possible. The f~50mm imaging lens was placed on a translational stage to allow focussing adjustments (needed to accommodate wavelength dispersion in the imaging lens) alongside the electronically operated shutter and narrow bandpass filters. The IR camera was placed in the image plane of the lens and interfaced with Labview to record data

Download Full Size | PDF

The optical arrangement satisfies several criteria for low noise imaging of the biopsy samples. Firstly the sample is illuminated as uniformly as possible by placing the condenser lens between the Globar source and sample in a 4f arrangement. This also serves to minimise radiative heating of the sample (Fig. 1). Secondly, because the imaging takes place over a wide range of optical wavelengths, allowances must be made for the optical dispersion in the lens material that results in chromatic aberrations which vary as the wavelengths are changed. By modelling the optics in Zemax software, it was possible to choose an appropriate imaging lens (Fig. 1).

2.2Mid-IR bandpass optical filters

Four bespoke optical narrow bandpass filters were commissioned which had transmission peaks (Table 1 ) designed to enable spectral imaging in wavelength ranges corresponding to key chemical components in the tissue (Fig. 2 ). Their full-width-at-half-maximum (FWHM) linewidths matched the typical peak widths in the absorption spectra.

Table 1. Characteristics of the Bespoke IR Bandpass Filters Used

View Table

 

Fig. 2 (Lower trace) Area-averaged absorbance spectrum of a 2.5mm x 2.5mm section of an unstained oesophageal biopsy section (sample 5601) recorded using a Jasco FTIR microscope at a spectral resolution of 4cm⁻¹. (Upper trace), measured transmission spectra of the 4 bespoke IR bandpass filters that were used in this study to obtain signal and baseline images at wavelengths corresponding to absorption by the Amide I and Phosphodiester chemical moieties. Variations in the areas of the transmission peaks were carefully measured and the respective Spectroscopic Image (SI) data arrays were scaled accordingly.

Download Full Size | PDF

For reference purposes a Jasco FTIR spectrometer was used to record an absorbance spectrum of sample 5601 (unstained) with a sampling area of 2.5mm x 2.5mm (Fig. 2).

To mimic the action of Haematoxylin, which stains the acidic cell nuclei blue, IR bandpass filters were designed to image and map the concentration of the (PO₂^-) moiety in the phosphodiester group that is present in cell DNA. A filter centred at λ = 8.13µm (1230cm⁻¹) was used for the absorption signal corresponding to this moiety, and a second, centred at λ = 8.50µm (1176cm⁻¹) was used to generate a baseline (Fig. 2) that was subtracted to correct for scattering and absorption in this region that was not specific to PO₂^-, and to correct for spatial non-uniformities in the sample illumination. Signals were corrected for the absolute filter transmission values, as separately recorded by the same Jasco FTIR spectrometer (Fig. 2(a)).

To mimic the Eosin, which stains cell cytoplasmic components, we used the “Amide I” vibronic band whose absorbance is a measure of the concentration of cytoplasmic proteins in human tissue and whose spectral prominence allows for good SNR imaging. The central maximum of the filter transmission peak was chosen to be λ = 6.00µm (1667cm⁻¹) with the baseline filter set at λ = 6.23µm (1605cm⁻¹).

2.3 The Mid-IR camera

The IR camera is an uncooled silicon microbolometer array (CEDIP IR Systems) featuring a 13.4mm x 10.1mm, focal plane array (FPA), comprising 384 x 288 elements each (35μm)² manufactured by ULIS IR. Each pixel signal is read out by its own complementary metal-oxide silicon (CMOS) read-out integrated circuit (ROIC), and the array has a thermoelectric temperature stabiliser. The FPA is housed in an evacuated enclosure to avoid water condensation whose window adds its own transmission characteristics to the system but the IR filter (usually supplied to limit the FPA spectral response to the λ>7µm atmospheric window, for thermal imaging applications) was removed to increase the usable spectral region to λ~ 2–14µm.

The absorption of IR radiation by a pixel element in the FPA changes its electrical resistance, generating a voltage that is first measured (over a selectable time period set here at 16msec) by the analogue part of the ROIC, before being digitised and sent to the controlling computer via a USB interface, at a capture rate of 50Hz and a 16msec exposure time. Images were recorded at a pixel bit depth of 14 corresponding to 16,384 absolute digital greyscale values per pixel.

2.4 FPA non-uniformity correction

Although using a FPA allows for much higher data rates than conventional single-channel scanning imagers, it introduces the problem of pixel-to-pixel sensitivity variation. Each detector element is characterised by its own zero offset and sensitivity and these vary with experimental parameters such as wavelength and exposure time. In order to obtain numerically reliable imaging data, these effects needed to be measured and corrected for.

For the FPA non-uniformity correction (FPANUC) the FPA was uniformly illuminated with a given peak wavelength and the raw digitised pixel data were captured as the intensity was slowly changed by varying the Globar temperature. Each pixel response curve was parameterised with a gain and offset parameter which could be subsequently used to adjust the raw digital values from each pixel so as to match the average for the whole array.

A small (0.5%), randomly distributed fraction of the pixels had responses which were too non-linear to be adequately described with this simple two parameter fitting process. These were labelled as “bad pixels” by the FPANUC algorithm, on the basis of manually set threshold criteria, and their intensity values were replaced by an average of those of the eight nearest neighbour pixels.

2.5 Spatial resolution

The ultimate resolution is set by the physics of diffraction in the optical system; applying Rayleigh’s criterion (Eq. (1))

A USAF 1951 resolution test chart (Fig. 3(a) ) was employed to assess the spatial resolution of the imaging system. The test pattern consists of a series of slits of known dimensions deposited onto a polyethylene sheet. Imaging it at λ~6.0µm gave the results of Fig. 3(b). Subjectively, we estimate the smallest resolvable group of slits is group 5, set 2, corresponding to 36 lines per mm, implying that the effective physiological resolution of the system is closer to r~14µm. This is roughly 1.5 times the Rayleigh Limit, partly due to optical aberrations in the non-ideal spherical imaging lens but mostly due to the ~11.1 µm effective pixel resolution at the magnification (3.1) used in this study.

 

Fig. 3 A USAF 1951 Test Resolution Pattern imaged at a wavelength of λ~6.00µm with the use of the appropriate narrow bandpass filter. By determining the smallest line pair elements resolvable on screen (Group 5; Set 2- 36.0 line pairs/mm) the useful spatial resolution of the imaging system was estimated at r~14µm.

Download Full Size | PDF

2.6 Single wavelength transmission mode imaging

All image acquisition, processing and shutter control was carried out through a National instruments Labview software interface developed in-house. 3 types of image were combined to generate the chemical maps, namely

Raw image (R_ij): The IR radiation transmitted through the biopsy sample is collected by the imaging lens and focused onto the FPA of the IR camera, successively at each of the 4 filter wavelengths.

Background Image (B_ij): To calculate the fraction of IR radiation absorbed by the sample, an image is recorded with the biopsy sample removed but all other optical elements unchanged.

Noise Offset (O_ij): This image was taken to correct both for the effects of ambient thermal radiation reaching the FPA, and for the pixel to pixel voltage offset variations. This image is acquired with the biopsy sample in place and the electronic shutter closed.

The raw and background images were each corrected for offset before being ratioed, to give an absolute value for the fraction of IR light absorbed by the sample (Eqs. (2)–(4)). With the Beer Lambert Law (Eq. (5)) it is then possible to calculate the absorbance map (A_ij)of the sample.

The pixel array(A_ij) provides quantitative information about the biological composition of the tissue being analysed and could readily be computer rendered into false coloured images. Typically each (A_ij)array was composed from a total of 200 image frames, corresponding to a total acquisition time of only 4 seconds.

3. The oesophageal cancer samples

Two sets of human oesophageal cancer biopsy sections were obtained from University of Oxford under a Local Research Ethics Committee permission Number 07-Q1604-17. The biopsies were processed by fixation in 10% neutral buffered formalin for 24 hours before being immersed in 70% ethanol for a further 24 hours then being embedded in paraffin wax at 60°C. Sections were cut with a microtome at a thickness 5µm. Two adjacent sections from each biopsy were placed separately, one on a 3mm x 25.4 mm diameter IR-transmitting Barium Fluoride (BaF₂) disc, the other on a standard glass microscope slide. These sections were then immersed in a series of 3 Xylene baths for 10 minutes each, totalling 30 minutes, to remove the paraffin wax.

One of the sections from each set (the one on the glass slide) was processed for H + E staining, while the other was left unstained. The sections were then covered and protected with, in the case of the unstained sections, BaF₂ cover discs (25.4mm x 0.5mm. These sections were labelled appropriately for traceability (biopsy samples 5601 and 5933 stained and unstained) and imaged using an optical microscope. The stained images (Figs. 4(a) , 6(b), and 6(d)) were then analysed by a trained histopatholgist for classification purposes.

 

Fig. 4 (a) Visible light image of a adjacent section taken from sample 5601, stained with a standard H+E protocol, (b) False colour chemical map from an adjacent, unstained section from the same biopsy sample, plotted to show the distribution of the concentration of the Amide moiety. This image is acquired by subtracting the image acquired through the λ = 6.23µm bandpass filter from that acquired at λ = 6.00µm.

Download Full Size | PDF

4. Results and discussion

Figure 4(b) shows the baseline-corrected absorbance signal, proportional to the Amide I concentration (C_ij) across sample 5601, rendered as a false colour image. On its own this signal reflects the density of biological material, with the highly proteinaceous squamous epithelium clearly absorbing much more strongly than the more hydrogel-rich areas such as the stroma, which has more complex polysaccharides, such as hyaluronic acid which is amide-free.

It was thought advisable to test the hypothesis that the baseline-corrected PO₂^- signal was a reliable measure of DNA concentration, as opposed to, say, elevated levels of ATP (adenosine triphosphate) and ADP (adenosine diphosphate) that may be present in the cytoplasmic components. To do this we tested the correlation between the PO₂^--related IR absorbance values and the density of nuclei that were visible in the H + E samples. The plot in Fig. 5 was generated by counting the cell nuclei, in the various rectangular (as delineated in Fig. 6 ) regions that were deemed, on the basis of the histological evidence, to be of a particular given tissue type (epithelial, stroma, adenocarcinoma etc.). The number of cell nuclei in the stained sample was normalised to the rectangle area and plotted against the mean phosphate absorbance of that same area in the corresponding unstained tissue section. The results (Fig. 5) show a strong linear correlation between nuclei density and total absorbance for a given area for both healthy and diseased tissue types, strongly supporting the assumption that the measured phosphate absorbance is a reliable measure of nuclear DNA concentration.

 

Fig. 5 Plot of the mean absorbance of selected rectangular areas of tissue, with widely varying nuclear density in samples 5601 and 5933 versus the areal density of nuclei found in that area in the corresponding H & E stained section. These areas are shown highlighted in Fig. 6. The strong linear correlation argues that the main contribution to the latter comes dominantly from PO₂^- moieties in the phosphodiester groups in DNA and less from free ADP/ATP. The vertical errors are smaller than the symbol sizes in all cases.

Download Full Size | PDF

 

Fig. 6 (a) and (c): false-colour chemical maps plotted to represent the ratio of the concentrations of the amide and phosphodiester moieties in the unstained sections of biopsy samples 5601 and 5933 respectively. (b) and (d):- standard H+E stained sections from adjacent sections of the same two biopsies, for comparison. Note the pronounced increase in the relative DNA concentration in the high grade dysplasia area (delineated by the dotted line) of the sample 5933 compared with the healthy epithelial tissue (delineated by the solid line). Tissue regions were delineated by eye.

Download Full Size | PDF

It is widely accepted that the nuclear to cytoplasmic ratio is a useful parameter in use by histopathologists as a morphological marker for judging the malignancy of a tumour. The common observation is that malignant cells possess a larger proportional nuclear volume [2], and this feature can be exploited in our measurements. Furthermore, malignant tumors have an abnormal karyotype with multiple structural and numerical aberrations of chromosomes – so-called `aneuploidy'. Malignant cells frequently contain multiple centrosomes [which are required for proper chromosome segregation], which can lead to aberrant mitosis and errors in chromosomal segregation. The chromosome number in cancer cells is highly variable, with elevated levels of chromosomal loss and gain, so-called `chromosomal instability' (CIN), resulting in ongoing karyotypic changes [3]. Both factors lead to an H + E stained biopsy that appears bluer than a healthy portion of the same tissue (Fig. 4(a)). This in turn suggests that more nuclear material may be present in malignant tissue, which is supported by observation [4] and which can impact on patient survival [5].

On this basis, we mapped the baseline-corrected DNA absorption (N_ij) across sample 5601 by subtracting the λ~8.50µm absorbance map from the at λ~8.13µm one, and used it, with a similarly derived map of the cytoplasmic protein absorbance, to generate a false-coloured image that corresponded to the ratio, AP_ij, of Amide/Phosphodiester concentrations, i.e.

This ratioing procedure has the effect of increasing image noise (particularly in dense parts of the section where two small transmittance values are being ratioed) so the resulting AP_ij arrays (Fig. 6) needed further smoothing, with a median filter developed in-house, to facilitate visual differentiation, before being rendered into a false-colour image to maps of the cytoplasmic to DNA ratio for the two samples.

The morphology of the “digitally-stained” sections correlates well with their conventionally H + E stained counterparts (Fig. 6). In sample 5601, the portions of the squamous epithelium and the adenocarcinoma, that had a higher proportion of nuclear material (both of which appear bluer in the stained section, (Fig. 6(b)) were highlighted in the corresponding digitally stained section (Fig. 6(a)) as areas of low Amide/Phosphate ratio. Their AP_ij values were less than half those for neighbouring stroma tissue, the Phosphate differences presumably reflecting the structural differences in the cells of the two types of tissue.

From a diagnostic perspective however, the data from sample 5933 are rather more illuminating. The H+E stained section shows an area of epithelial tissue (Fig. 6(d)), part of which has entered a dysplastic phase. Here the ratioed image dramatically highlights the chemical changes accompanying dysplasia; the AP_ij values averaged over the two rectangular areas highlighted in Fig. 6(c) differ by a factor 2.07±0.06. We see this as a very encouraging result; in the H+E stained section the dysplastic tissue is only differentiable on the basis of its morphology, whereas the digitally-stained image reproduces this morphological information, and augments it with hard quantitative numerical information.

Furthermore, we believe that the high accuracy of the numerical data indicates a potential for developing a test for the onset of dysplasia based on the AP_ij values, that is significantly more sensitive and specific than the present morphology based methods that use the H + E sections on their own. Pertinent to our novel analytical method is a recent report that for diagnosis in Barrett's oesophagus there is quite a divergence of the opinions of pathologists [6] as to the grading of the disease. We believe our new approach may enable better inter-pathologist concordance and thus lead to improved clinical outcomes.

5. Conclusion

In conclusion, we have demonstrated, for the first time, that modern IR camera technology makes it possible to use narrow bandpass mid-IR spectral filters and simple thermal IR light sources in a way that allows spectral images to be captured quickly and simply enough to make spectroscopic imaging, or at least a stripped down version thereof, a practical proposition for the working histopathology lab.

Although the technique we propose is rather less versatile than a full SI implementation, this makes it favourable for routine use in the clinical environment, where it offers with considerable advantages over conventional FTIR Imagers particularly in terms of reliability, ease of use and cost.

Although future technology improvements will doubtless improve the sensitivity and spatial resolution, the latter will always be limited (to ~10μm) by the physics of diffraction at these longer wavelengths. This is, of course, rather more than in standard H + E images, but nonetheless we believe that, especially if the two techniques are used together, the quantitative chemical information provided by this technique will prove to be a powerful adjunct, both in advancing the sensitivity of standard diagnosis, and as a measure of e.g. disease dynamics.

Clinical studies are planned to quantify these potential improvements, and we believe that, subject to the development of appropriate protocols, and correlation with existing markers for malignancy in biological tissue, this imaging approach is capable of delivering significant improvements in the sensitivity and specificity of the diagnosis and monitoring of cancers.