Method, computer programme and system for analysing a sample comprising identifying or sorting cells according to the FTIR spectrum each cell produces
11598719 · 2023-03-07
Assignee
Inventors
Cpc classification
International classification
G01N21/27
PHYSICS
Abstract
The invention relates to a method for improving the screening of histological samples, especially samples that may include cancerous or precancerous cells, or cells having other disease states. The method involves analysing a sample obtained from a subject and comprises the steps of providing the spectra produced by scanning the sample using FTIR spectroscopy and identifying or sorting the cells in the sample according to the spectrum each produces.
Claims
1. A method for analysing a sample obtained from a subject, comprising the steps of: a. providing a spectrum produced by spectrally scanning the sample using a single crystal element ATR infrared spectroscopy technique, wherein the spectrum is an average spectrum over a substantial part of the sample; and b. identifying one or more tissue types in the sample according to the spectrum each such tissue type produces; wherein the identifying comprises a combination of steps of: (i) assigning the spectrum to either: a first classification; or a squamous classification, wherein the first classification includes but does not distinguish between non-dysplastic, high-grade-dysplastic and cancerous tissue types; (ii) assigning a spectrum from the first classification to either: an epithelial classification; or a lamina propria classification; (iii a) assigning an epithelial spectrum to either: a non-dysplastic classification; or a high-grade-dysplastic or cancerous classification; (iii b) assigning a lamina propria spectrum to either: a non-dysplastic classification; or a high-grade-dysplastic or cancerous classification; and c. determining the presence or absence of a disease state depending on whether high-grade-dysplastic or cancerous tissue types are identified in the sample.
2. A method according to claim 1, wherein the identifying further comprises combining classification results from two spectra, one from each side of the sample.
3. A method according to claim 1, wherein the sample is epithelial tissue from the esophagus.
4. A method according to claim 3, wherein the sample comprises two or more constituent tissue types, including a predominant tissue type and at least one minority tissue type.
5. A method according to claim 1, wherein the method includes the step of obtaining the sample.
6. A method according to claim 1, wherein the method includes the step of obtaining the spectrum.
7. A method according to claim 1, wherein assigning the spectrum under step (i) to either: the first classification; or the squamous classification is on the basis of 1385-1235 and 1192-1130 cm.sup.−1 spectral regions and/or the spectral region around 1153 cm.sup.−1.
8. A method according to claim 1, wherein assigning a spectrum from the first classification under step (ii) to either: the epithelial classification; or the lamina propria classification is on the basis of the spectral region around 1570 cm.sup.−1 and/or a 1610-1465 cm.sup.−1 spectral region.
9. A method according to claim 1, wherein assigning an epithelial spectrum under step (iii a) to either: the non-dysplastic classification; or the high-grade-dysplastic or cancerous classification is on the basis of a 1200-900 cm.sup.−1 spectral region, preferably a 1100-900 cm.sup.−1 spectral region, further preferably the spectral regions around 1082, 1043, and 974 cm.sup.−1.
10. A method according to claim 1, wherein assigning a lamina propria spectrum under step (iii b) to either: the non-dysplastic classification; or the high-grade-dysplastic or cancerous classification is on the basis of 1290-1210 and 1130-870 cm.sup.−1 spectral regions, preferably the spectral regions around 1221 and 1047 cm.sup.−1.
11. A method according to claim 1, wherein the method also comprises the step of shifting or calibrating the spectrum to take into account the hydration of the sample.
12. A computer programme comprising code means to carry out the method of claim 1.
13. A computer readable medium carrying a computer programme according to claim 12.
14. A system comprising a computer enabled to run the computer programme according to claim 12.
15. A system according to claim 14, further comprising a spectrometer.
16. A system according to claim 14, further comprising a library of spectra from known cells.
17. A method for diagnosing a disease state in a subject, comprising analysing a sample from the subject using the method of claim 1, wherein the presence of non-healthy cells or cells having a particular disease state is indicative of the subject having the disease state.
18. A method according to claim 1, wherein step b) is performed by comparing the spectrum to a reference data set.
19. A method according to claim 18 wherein the reference data set are spectra of samples with a known specified disease and/or disease state produced using a spectroscopy technique with microscopic resolution.
Description
BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS
(1) The invention will now be described by way of example only, with reference to the drawings.
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(17)
(18)
(19)
(20)
(21)
(22)
(23)
(24)
(25)
(26)
(27)
DESCRIPTION OF SPECIFIC EMBODIMENTS OF THE INVENTION
(28) Methods
(29) FTIR Spectroscopic Imaging
(30) FTIR spectroscopic images were measured using a Bruker IFS 66 spectrometer coupled with a Hyperion IRscope II microscope with a 15×0.4 NA objective, and a 128×128 pixel mercury-cadmium-telluride focal plane array (FPA) detector. The use of an array detector allows spectra to be obtained simultaneously from different spatial regions of the sample and is significantly faster than mapping the same area with a single element Detector.sup.34,38. An NDBE/LGD and a SQ sample were microtomed to an 8 μm thickness and mounted on 2 mm thick calcium fluoride windows and deparaffinised by a standardized xylene protocol [reference].
(31) Spectroscopic images were recorded using the FPA with a 96×96 pixel window, giving a field of view of 256×256 μm.sup.2. Using in-house developed macros, spectral images were acquired from different parts of the tissue and combined in Matlab to produce large data sets. Each sample was mapped to cover an area containing regions of epithelial and lamina propria cell types as well as different levels of dysplasia, which were previously graded by histopathologists. The FTIR images were binned in 4×4 matrices to increase signal to noise. Any pixels that contained an amide II integral of less than 0.5 between 1571-1490 cm.sup.−1 were excluded from further analyses. All other spectra were normalised to the height of the amide II peak and trough between 1555 and 1475 cm.sup.−1. The images were subsequently binned in 4×4 matrices to improve signal to noise ratios. Contributions of water vapor were removed by the subtraction of a pre-recorded water vapor reference spectra. After processing, differences between different cell types were revealed.
(32) ATR-FTIR Spectra Recording
(33) A Bruker Optics IFS 66/s FTIR spectrometer that records in the region of 6000-800 cm.sup.−1 was used to record spectra, however only the 4000-900 cm.sup.−1 region was collected, however, only the 2200-900 cm.sup.−1 region was analysed. The machine has a liquid nitrogen cooled MCT-A detector with an Attenuated Total Reflection (ATR) 3-reflection silicon prism with ZnSe optics. The spectra were recorded using the Bruker OPUS 6.5 software.
(34) All measurements were recorded at 4 cm.sup.−1 resolution, giving a peak accuracy of approximately ±1 cm.sup.−1. 1000 background interferograms of the clean prism surface were averaged (taken after carefully cleaning the prism with water and 100% ethanol) and, after correct placing of the biopsy sample, 500 interferograms were averaged to produce single biopsy absorbance spectra.
(35) Data Processing
(36) All data was converted from Bruker OPUS 6.5 file format to ASCII file format from within the software. The data was then preprocessed and analyzed using in house scripts developed in MATLAB_R2012b and/or PLS toolbox v 7.0.3.
(37) Prior to analysis, four data pre-processing steps were applied to each spectra; in this order: spectral water subtraction using a reference water spectrum, spectral water vapor subtraction using a pre-recorded reference spectrum, normalization to the height of the amide II band, and second derivative calculation.
(38) Histopathology
(39) The gastrointestinal department at UCLH had two associated histologists. To ensure a correct diagnosis, both histologists independently verified any sample with a dysplastic diagnosis. The samples were stored in a 4% formalin solution, placed in embedding cassettes, and dehydrated by placement in the following solutions of ethanol for 2 hours: 70%, 80%, 95% and 100% with each solution being refreshed after one hour. Biopsies were then placed in xylene for three hours, changing the solution every hour. The biopsies were then placed in paraffin wax (˜57° C.) for 1.5 hours, and repeated before embedding into a paraffin block. Blocks were then sliced with a microtome into 4 μm sections in a 40-45° C. water bath, mounted on a glass slide and oven dried. The sample was then rehydrated in xylene for 5 minutes; the solution was changed and then repeated 3 times. The sample was then rinsed in 100% ethanol for three minutes, repeated, and followed by 3 minutes in 95% ethanol after which the sample was rinsed with distilled water and stained for inspection.
(40) The samples were then categorized as either healthy SQ epithelial cells, or one of the three classes of BE; NDBE, HGD or EAC. In the case of the intercepted-matched dataset, an additional class of NDBE-IM was included. An expert pathologist at UCLH diagnosed each biopsy, and if a biopsy was diagnosed with either HGD or EAC, an additional histologist independently verified the diagnosis.
(41) Biopsies classified as LGD were excluded from the training data of the model.
(42) Statistical Analyses
(43) Partial Least Square Discriminant Analysis
(44) Partial least square discriminant analysis (PLSDA) was applied for dimensionality reduction to maximize the covariance between explanatory, correlated variables (wavenumbers) and categorical variables (disease stage). Since the model was built using a multi-step process, the subset of variables (wavenumbers) changed depending upon the cell type. For the SQ separation the 1385-1235 and 1192-1130 cm.sup.−1 region was used, for the epithelium diagnostic model the 1200-900 cm.sup.−1 was used and for the lamina propria 1290-1210 and 1130-870 cm.sup.−1 region was used. These regions were then reduced into a lower dimensional space of uncorrelated variables, referred to as latent variables, the number of latent variables used are indicated within the results section. To calculate fast and accurate the PLS model, we follow the approach of De Jong, Sijmen.sup.39.
(45) Logistic Regression
(46) In order to discriminate disease stages we assigned a probability to each stage based on the scores generated from the PLSDA using logistic regression analysis. The logistic regression model is given by
(47)
where Y.sub.i describes the binary responses (disease stages), β.sub.0 is the intercept, β is the vector of coefficients and X* is the matrix of latent variables. From this equation we can calculate the probabilities for each disease stage and a classification rule must then be applied in order to identify a threshold between the two groups.
(48) Applying a Misclassification Cost
(49) To optimize the classification performance we applied misclassification costs to the decision problem. Given the data, X*, there were two possible decisions: No-treat, which corresponds to grouping an unknown biopsy spectrum as no-treat (SQ/NDBE) and treat, which corresponds to grouping an unknown biopsy spectrum as treat (HGD/EAC). No losses were applied to a correctly classified biopsy spectrum. If the decision was no-treat, but the true group was treat, then there is a cost of λ.sub.treat, which was fixed at 1. Similarly, the decision misclassification of treat biopsy as no-treat was assigned a cost λ.sub.no-treat, which was varied, refer to results section for λ.sub.no-treat. See below:
(50)
(51) The conditional risks (expected losses) are r(no-treat|X*)=λ.sub.treatp(treat|X*) and r(treat|X*)=λ.sub.no-treat p(no-treat|X*). The decision is no-treat, if r(no-treat|X*)<r(treat|X*), or equivalent λ.sub.no-treat p(no-treat|X*)>λ.sub.treatp(treat|X*), otherwise the decision is treat.
(52) Optimize the performance of the evaluation measurements by including an additional ‘inconclusive’ prediction class, which corresponds to the samples that, lie very close to the threshold (t). If G=|p(no-treat|X*)λ.sub.no-treat−p(treat|X*)λtreat|<t, then the sample is characterized as ‘inconclusive’, where t in practice is selected the q %-quantile of these differences in absolute values. On the other hand, the decision either threat or no-treat is done with 100%−q % confidence, if G>t.
(53) Cross Validation
(54) In order to most closely represent the clinical environment, the training dataset and the test dataset never contain the same patients, regardless of the number of biopsies taken from each patient. The process is as follows, where N is the number of patients: i) the biopsies are randomly split into N, not necessarily of equal size, sub-samples, such that biopsies that referred to the same patients belong to the same sub-sample, ii) use N−1 sub-samples of patients as training data and one sub-sample, out of the N patient, as test data and iii) repeat first two steps N times, one for each patient. The method is not so computationally expensive since the number of patients is much lower than the number of biopsies.
(55) Results
(56) FTIR Imaging to Generate a Library of Cell and Disease States Spectral Characteristics
(57) The single element ATR-FTIR method, though rapid, simple and with high signal/noise, has several inherent limitations. One of these being the large field of view (several mm.sup.2), which results in the averaging of spectra of all cell types and disease stages across a sample. This means that spectral differences between diseased and healthy cells are difficult to resolve. In particular, signals from a small number of diseased cells may be averaged out in a sample that is predominantly healthy. To overcome this limitation, FTIR imaging was used to generate a library of cell and disease stage characteristics. FTIR microspectroscopic images of 8 μm thick tissue sections containing known disease stages: SQ, NDBE with a LGD region, HGD and EAC were recorded. Characteristic features of cell types and disease stages could then be selected from these images and used to better identify spectral signatures of specific cell types in ATR-FTIR spectra.
(58) Cell Type Spectral Characteristics of a SQ Biopsy Section
(59) A SQ sample is expected to contain two tissue types: surface epithelium (EP) and underlying lamina propria (LP).
(60) Cell Type Spectral Characteristics of a NDBE/LGD Biopsy Section
(61)
(62) Cell Type Spectral Characteristics of a HGD Biopsy Section
(63)
(64) Cell Type Spectral Characteristics of a Large EAC Tissue Section
(65)
(66) Comparisons of Disease Stage Spectral Characteristics in the Epithelial Tissue
(67) Although the NDBE, LGD and HGD+ CEP and the SQ EP are all epithelial cell types, they have different cell structures. Therefore, the spectral differences between CEP and SQ EP are expected to be larger than the spectral differences between disease stages of the CEP.
(68)
(69) Comparisons of Disease Stage Spectral Characteristics in the Lamina Propria
(70) The SQ LP and the LP present in BE samples (NDBE/LGD/HGD+) are expected to contain the same tissue components and therefore have similar spectra.
(71)
(72) Single Element ATR-FTIR Spectroscopy of Fresh Biopsies In total, 790 biopsy spectra of 414 biopsies from 131 patients were measured using single element ATR-FTIR spectroscopy. Where possible, at least one spectrum was recorded of each side of the biopsies. In cases of small biopsies, only one spectrum was recorded; conversely if the biopsy was large, multiple spectra of each side were taken. The spectra were corrected for water and water vapor contributions, normalized to the height of the amide II band and converted into their second derivative forms before further analyses. Of these 790 spectra, 80 were removed as outliers, leaving a total of 710 biopsy spectra from 379 biopsies and 122 patients (Table 1). A spectrum was determined as an outlier if it deviated from the mean plus or minus the standard deviation in over 75% of the 1800-850 cm.sup.−1 spectral region, after processing.
(73) Grouping Biopsy Spectra by Predominant Cell Type
(74) A further limitation to single element ATR-FTIR spectroscopy is the limited depth of penetration (several microns) of the evanescent wave.sup.37, which means that only the surface layers of cells are analyzed. Biopsy samples tend to be roughly disc-shaped, with one face derived from the exposed surface (either SQ EP or CEP) and the other from the underlying tissue (LP). To help overcome the depth of penetration limitation, and because these two surfaces contain different cell types, spectra from both sides of the biopsy were routinely recorded and categorized according to their predominant cell type, before being analyzed by their disease stage.
(75) The following predominant cell types were assumed to be present: EP or LP for SQ, NDBE and HGD; and EAC only for the fully cancerous samples. LGD samples have been purposefully removed from the training data of the model described here. This is because the inter-observer agreement of LGD diagnosis between histologists is low, where κ-values are reported as low as 0.27.sup.8, and there is also debate over whether these patients should be treated with ablative therapies or not.sup.4. For these reasons LGD patients were excluded from the main part of the analyses. However, the ability of the model to predict these patients will be discussed.
(76) In order to optimize the performance of the classification model, the spectra were sorted by the pipeline as illustrated in
(77) (i) Separation of NDBE/HGD/EAC From SQ Biopsies
(78) To effectively separate the SQ (SQ EP and SQ LP) tissue from all other disease stages (NDBE/HGD/EAC), a PLSDA with leave-one-patient-out cross validation was applied to the 1385-1235 and 1192-1130 cm.sup.−1 spectral regions of 710 individual spectra. This spectral region was selected based upon the differences between the SQ EP and SQ LP comparisons in
(79) The specificity of the SQ detection model was low at 64%. Before continuing to the next step, information from the FTIR image study was used to help improve the performance of the model. SQ EP was found to have a unique band at 1153 cm.sup.−1. The average integral of this group was −5.6483×10.sup.−5±0.15987×10.sup.−5 and the average integral of all other tissue types/disease stages was −0.0015±6.4367×10.sup.−4. Therefore, if the integral of this component was less than or equal to −8.5633×10.sup.−4 it was classified as SQ EP. If a biopsy had a spectrum present in the NDBE/HGD/EAC and the SQ group, the SQ spectrum was checked for the presence of the unique SQ EP peak. If the peak was present the previously misclassified NDBE/HGD/EAC spectrum, was then re-classified as SQ. This additional check resulted in the correct re-classification of 28 of the 60 incorrectly classified SQ, improving the specificity of NDBE/HGD/EAC versus SQ model to 81% (135/160) without misclassifying any of the NDBE/HGD/EAC biopsies.
(80) (ii) Separation of Epithelium From Lamina Propria in NDBE/HGD/FAC Spectra
(81) The NDBE/HGD/EAC spectra were then analyzed in terms of whether they represented predominantly EP or LP cell types. FTIR imaging revealed that spectra from the EP could be distinguished from the LP of NDBE and HGD by the presence of a second derivative peak at 1570 cm.sup.−1, and a shift of the amide II band. Based on this, k-means clustering analysis where k=2 (two groups), was performed on the 1610-1465 cm.sup.−1 region of NDBE and HGD single element ATR-FTIR measurements. The 1610-1465 cm.sup.−1 spectral differences between the EP and LP of NDBE and HGD samples were used to build a leave-one-patient out PLSDA model that would predict sidedness.
(82) (iii a) Disease Stage Separation in Epithelial Spectra
(83) After the separation of the NDBE/HGD/EAC spectra into EP and LP, these categories were further separated into NBDE or HGD/EAC disease stages.
(84) (iii b) Disease Stage Separation in Lamina Propria Spectra
(85)
(86) Combining Classification Results from Each Side of the Biopsy
(87) As stated, where possible each biopsy had a total of two spectra recorded, one from each side of the biopsy. Of the 340 total biopsy ATR-FTIR measurements of BE EP and BE LP (not including EAC), 158 of them had one EP and one LP spectrum, 96 had two EP readings, 40 had two LP readings, 23 had only a single EP spectrum and 23 had only a single LP spectrum. Of the multiple spectra recorded, results from the various models agreed on average 87% of the time. After averaging the prediction scores, the sensitivity of the SQ/NDBE versus HGD/EAC was 90% and the specificity was 71% where the HGD/EAC misclassification cost was 3.
(88) Optimizing the Model for Clinical Application
(89) The sensitivity and specificity of the model described above can be further optimized to meet clinical needs in order to be used as a dysplastic BE biopsy screening device to aid the clinician's decision making process. A screening device requires a minimum sensitivity of 95%, where specificity can be sacrificed as long as a there is a clear clinical benefit. To increase the sensitivity of this model an ‘inconclusive’ classification result was included. This step was used to improve the certainty of the two classes by reducing the number of false negatives and false positives in a single step. An inconclusive result was given if any of the following statements were true. First, the classification predictions from either side of the biopsy disagreed. Secondly if both of the cost adjusted p-values were above, or alternatively both spectra below a threshold of 0.8; meaning that if the model was certain that both spectra should fall into opposing classification groups, or alternatively if the model was uncertain that both spectra should fall into opposing classification groups, then the biopsy should be inconclusive. In the event that one spectra was above the threshold and the other was below the threshold, the biopsy would take the classification of the spectra above the given threshold. The inclusion of these rules resulted in an overall HGD/EAC sensitivity of 97%, a specificity of 83% and an inconclusive rate of 18%.
(90) To test how the diagnostic model would perform for LGD patients, 27 spectra of 14 biopsies from 10 LGD patients were tested. After combining the predictions from both sides of the spectra as described above, 7 biopsies were classified as SQ/NDBE (where 2 were SQ); 2 were HGD/EAC and 5 were inconclusive. The inconclusive rate for LGD biopsies was 36%, higher than the inconclusive rate of 18% for the other classes.
(91) Discussion
(92) Here, we describe a technique that enables single element ATR-FTIR spectroscopy to be used as a real-time point-of-care screening device for HGD/EAC biopsies. Like other vibrational spectroscopic methods, ATR-FTIR spectroscopy can provide a clinical tissue diagnosis based on its biochemical profile. Where other methods try to offer an in vivo diagnostic that has a high initial cost and a requirement for a specialist operator, we suggest a more simplistic approach that can be operated by a nurse. One of the major benefits of single element ATR-FTIR spectroscopy is that is not limited to a single sample type, and no damage are caused to the samples. Therefore, it can be used for analysis of solids or liquids and the same sample can be sent on for classical diagnosis if needed, thus making it a versatile tool and applicable to many different clinical settings.
(93) Biochemical changes between SQ, NDBE and HGD/EAC tissue can be seen in the 1200-900 cm.sup.−1 spectral region, particularly the 1082, 1043, and 974 cm.sup.−1 bands of the EP and the 1290-1210 and 1130-870 cm.sup.−1 regions in the LP. These biochemical changes were modeled via PLSDA using a leave-one-patient-out cross validation built with 710 ATR-FTIR spectra of 379 fresh biopsies from 122 patients. There were three possible outcomes: SQ/NDBE, or no treatment required; HGD/EAC, where immediate treatment would be required; or inconclusive where the certainty of the result is not high enough to classify it. Each result had an associated certainty level (p-value), which could be displayed to the clinician in order to aid the clinical decision making process. The model has an overall accuracy of 90%, with a sensitivity of 97%, a specificity (not including inconclusive results) of 83% and an inconclusive rate of 18%. When the model was tested on LGD patients, 50% of the 14 biopsies from 10 patients were classified as SQ/NDBE, 14% were classified as HGD/EAC and 36% were inconclusive. Ideally all LGD biopsies would either fall into the inconclusive or HGD/EAC. However, the model was trained using histology results for which the inter-observer agreement was less than 50% for LGD diagnosis. Therefore it was expected that the model would not classify LGD biopsies consistently into a single group. It is however encouraging that 36% of the results were inconclusive as LGD is a dysplastic stage between the NDBE and HGD/EAC classes. With more samples it would be possible to include an additional LGD group into the model.
(94) If a spectrometer were to be installed with the model that we have presented here, a reduction of histological bulk by at least 50% could be achieved by only sending those biopsies with an uncertain prognosis for further analysis. This is based on the fact that over 90% of all Barrett's surveillance biopsies sent to histopathology are healthy. A reduction of histological bulk of this size would make considerable cost savings to the healthcare provider. Furthermore, there is potential for this model to provide benefit to those patients predicted to be HGD/EAC. When a patient is predicted to be HGD/EAC, the model is 83% certain that this is true. If the clinician assessed this p-value, and agreed with this prediction based on what they see at endoscopy and the patient's clinical history, there is potential for the patient to be treated immediately.
(95) The study was conducted using a liquid nitrogen cooled single element ATR-FTIR spectrometer, which is not appropriate for use in the clinic. However, there are portable, room temperature devices available that claim to produce the same data quality as lab grade equipment in less than 10 seconds. In order to translate this device into the clinic, a larger study would need to be conducted on one of these smaller bench-top single element ATR-FTIR spectrometers.
(96) Application of FTIR Imaging and ATR-FTIR Spectroscopy to Lung Cancer Diagnosis
(97) FTIR imaging and ATR-FTIR spectroscopy as described above were also applied to lung cancer. FTIR spectroscopic imaging in transmission mode was used to characterise cell and disease progression of lung squamous cell carcinoma (SCC) in a single, deparaffinised, 8 μm thick biopsy section that contained histologically-defined areas of disease progression. Disease stages that were present in this biopsy included healthy, mild/moderate/severe dysplasia and SCC in situ. The use of such a sample was to eliminate inter-sample and inter-patient spectral differences that might occur that are unrelated to carcinogenesis and due to it being rare for a single sample to display a complete disease transition from healthy to carcinoma in situ. The present study describes for the first time FTIR spectroscopic imaging of such a sample. The cell type information gained from the FTIR image was used to develop an algorithm to sort a small dataset of fresh lung biopsies from 21 patients. Their disease stage differences were then assessed.
(98) Use of FTIR Imaging to Characterize Spectral Features of a Bronchial Biopsy
(99) Cell Type Differences
(100) The largest spectral differences appear to be found between the cell types within a tissue. Therefore, in this study, the cell types were first separated before assessing disease stage differences within the cell types.
(101)
(102) Disease Type Differences
(103) The disease stages in the EP and LP were analysed separately to eliminate misinterpreting spectral differences arising from cell type as differences in disease stages.
(104) To increase the SNR, the image was first binned in 4×4 matrices, where each pixel had an approximate size of 10.8×10.8 m. However, this was still insufficient to accurately de-convolute the small signal differences that could be used to distinguish disease stages. To increase the SNR, spectra from large areas of the image were selected and averaged.
(105) Analysis of the Epithelium
(106) A total of fifteen areas of the image were manually selected along the EP (
(107) According to the histological analysis of the H&E stained section (
(108) To assess whether choosing a spectral region, as opposed to a single peak, helped resolve disease stages, a three component PCA was performed on the 1100-1030 cm.sup.−1 spectral region (
(109) Analysis of the Lamina Propria
(110) Since SCC originates in the surface EP, the major spectral differences are expected to arise in this layer of cells. To investigate whether the progression of dysplasia also affected spectral properties of the underlying tissue, fifteen LP areas were manually selected from the mapped FTIR image (
(111) The disease stage of the LP was defined according to the histopathology of the adjacent area of EP: areas 1-3, healthy; areas 4-6, mild dysplasia; areas 7-9, moderate dysplasia and areas 10-15, severe dysplasia/carcinoma in situ.
(112) Spectral differences were evident in their second derivative troughs at 1334 and 1279 cm.sup.−1 and peaks at 1215 and 1066 cm.sup.−1. The intensities of these bands decreased as the disease progressed. The 1279 cm.sup.−1 band and the 1080-1050 cm.sup.−1 spectral region also exhibited some shifts as the disease progressed (
(113) Spectral regions that contained the largest differences between disease stages in the LP are those from 1350-1196 cm.sup.−1 and 1097-1041 cm.sup.−1 spectral regions.
(114) Single Element ATR-FTIR Spectroscopy of Fresh Bronchial Biopsies
(115) Assessing the Sidedness of the ATR-FTIR Bronchial Biopsies
(116) Based on the above analysis of the 8 μm thick deparaffinised bronchial biopsy FTIR image, two cell types were expected in a typical sized biopsy (1-2 mm in diameter and up to 1 mm in thickness). Biopsies are roughly disc shaped where EP cells are expected on the surface, and LP on the underlying side.
(117) The orientation of the biopsy on the ATR prism was not known, and, if the biopsy was twisted or folded, it was possible that the biopsy was oriented in such a way that both surface and underlying layers were in contact with the prism. Therefore, the sidedness of the biopsy could not be determined using ATR-FTIR spectroscopy alone.
(118) To separate the spectra of fresh biopsies recorded with ATR-FTIR spectroscopy into groups based on predominant cell types (i.e. surface EP or underlying LP), the 1614-1465 cm.sup.−1 spectral region was used in a HCA. This region of the spectrum was found to contain differences arising predominantly from cell type differences. Three main clusters were produced from the HCA, the average second derivative spectrum from these clusters can be seen in
(119) Table 4 shows the distribution of spectra across the three possible classes in
(120) Assessing the ATR-FTIR Differences in Disease Stages of the Lung
(121) To make accurate distinctions between disease states. The three predominant EP and LP groups were first created from the HCA (see above), before disease staging of the different groups were assessed.
(122) Disease Stage Comparisons of Spectra From the Epithelium
(123)
(124) Disease Stage Comparisons of Spectra From the Lamina Propria
(125) Since bronchial dysplasia develops in the EP, minimal change in the LP tissue was expected. However, the FTIR imaging study of the single biopsy suggested that the LP might also display features characteristic of the disease stage of the adjacent EP.
(126) Normal LP shows a second derivative lipid trough at 1743 cm.sup.−1, where as LGD, HGD and cancer have a shifted trough at 1738 cm.sup.−1, where the trough in cancer samples was much larger than the other disease stages. There appears to be a transition from normal to cancer with a decrease in the 1360 cm.sup.−1 second derivative peak and a total of 2 cm.sup.−1 shift to a lower frequency. Changes in the spectral region between 1163-1171 cm.sup.−1 were also observed. Normal LP had a single broad peak at 1163 cm.sup.−1, where as LGD and HGD have a combination of bands at 1163 and 1171 cm.sup.−1, and cancer has a much larger 1171 cm.sup.−1 trough (
(127) Disease Stage Comparisons of Spectra From Mixed Cell Types
(128) The mixed cell type group exhibited too much variation within each disease group to pick up any significant spectral differences that could be used to distinguish between them. The mixed cell type is likely present due to poor orientation of the biopsy on the prism, for example, if the biopsy was twisted, the resulting spectrum containing both the surface EP and the underlying EP. During the studies with both BE and lung cancer, one of the main drawbacks of using ATR-FTIR spectroscopy for clinical diagnosis of fresh biopsies was its orientation on the prism. The fresh lung biopsies were recorded before the sidedness issue was known. Therefore, if future collections of data were recorded, a protocol would be in place to help ensure that the biopsy was not twisted or folded. This would help prevent multiple cell types being in-contact with the prism.
(129) Results
(130) As with the main BE study, FTIR imaging was used to generate a library of typical cell type spectra and disease type spectra. A single 8 μm thick deparaffinised lung biopsy that contained a complete gradation of diseases from healthy to SCC in situ was used for the analysis. The cell types were significantly different and were easily separated by the integration of 1591, 1334, 1215 and 1275 cm.sup.−1 bands, or by HCA of the 1614-1465 cm.sup.−1 region, which was used to separate the spectra into two distinct cell type groups. There is little information present in the literature regarding spectral cell type differentiation of lung tissue. Bird et al. present an example where they used FTIR imaging to separate the different cell types in a single sample. This was done by using a 10 class HCA, which was assigned to cell/tissue types such as the LP with fibroblasts, LP with abundant lymphocytes, blood vessels, macrophages and mucinous glands. However, they did not specify the spectral features that arise from these cell types and so cannot be compared with the data here. The cell type differences found here do, however, bear similarity to the cell type differences found between the EP and LP in BE. This similarity was in the second derivative 1633 cm.sup.−1 spectral region where there was an amide I shoulder present in the BE CEP and lung EP and absent in the BE LP and lung LP. Although, the shoulder in the lung EP was not as prominent as the BE FTIR image data. The other prominent feature that could be used to determine sidedness in the 1614-1465 cm.sup.−1 spectral region of BE FTIR image data, was the second derivative peak at 1570 cm.sup.−1. The differences in lung EP and LP in this region of the FTIR imaging data were not as prominent. Nevertheless, the differences in the 1570 cm.sup.−1 bands in ATR-FTIR spectra of fresh lung EP and LP biopsies were more similar to those found between the EP and LP ATR-FTIR spectra of fresh BE biopsies. The reason that this differences was less apparent in the FTIR image data was unknown. However, it is possible that it arose because of effects of dehydration on the imaging sample spectra, which can affect the IR spectra. However, due to the consistency of spectra across the image, it is most likely that the differences in the 1614-1465 cm.sup.−1m arose from real differences in protein types. The LP is a structural tissue comprised of a network of fibrous tissue that contains more collagen and blood vessels than the epithelium. Healthy EP is a thin layer of epithelial cells and will contain more cells than the LP. The spectral differences between the EP and LP in the 1560-1190 cm.sup.−1 spectral region have not been assigned to any individual components due to the complex overlapping components likely to contribute to this region. However, these differences are most likely to be related to differences of the fibrous connective tissue in LP compared to the layer of EP cells which will contain more metabolites are carbohydrates.
(131) The differences between the disease stages within the EP and LP were largely in the 1350-1000 cm.sup.−1 spectral region. The EP demonstrated differences between healthy and dysplastic/carcinoma in situ in band integrals at 1163, 1095, 1074 and the 1036 cm.sup.−1. However, no significant differences could be found between the intermediate dysplastic stages when comparing intensities of integrals of these components. However, some distinction between disease stages could be further resolved with a PCA of the 1100-1030 cm.sup.−1 spectral region. However, to determine the significance of these spectral changes, more samples would be required in the study. Comparisons between the changes in the spectrum reported here and model compounds suggest that the changes in the 1074 and 1036 cm.sup.−1 might be attributed to changes in glycogen related compounds. A recent biochemical study on a lung cancerous cell line versus a healthy cell line by Chaudhri et al. suggests that there is a decrease in metabolites, including glucose, from the healthy cells to cancerous cells, supporting this finding.
(132) The spectral differences between the disease stages were more prominent in the LP. The main differences arose at 1334, 1279, 1215 and 1066 cm.sup.−1 bands. The differences between the disease stages was further resolved with a PCA using the 1350-1196 and the 1097-1041 cm.sup.−1 spectral regions. Particularly interesting was a band shape change around, 1066 cm.sup.−1, indicative of the introduction, and/or change, of one or more biochemical components. It is known that cancer development triggers the inflammatory response. It is possible that such additional component(s) were caused by leukocytes and other cells/proteins recruited to the area in response to the inflammatory process. It was possible that changes in the 1066 cm.sup.−1 region was due to an increase in the amount of DNA relative to protein from the additional cells from the inflammatory response. However proportional changes in other DNA bands were not seen.
(133) Bird et al. describe the changes between SCC and healthy tissue at 1235, 1090, 1065 and 965 cm.sup.−1, which they attributed to changes in DNA. These are in part similar to the 1095 and 1066 cm.sup.−1 band changes seen in the present study. However, it is difficult to make direct comparisons as it was not clear whether the EP and LP had been separated in the Bird et al. study. Another FTIR imaging study of SCC and healthy lung tissue by Yano et al. reported discrimination based on the height of the 1045 cm.sup.−1 band, when normalised to the amide II band. They attributed this change to collagen based on their previous work with pulverised fresh biopsies in FTIR transmission mode. Whilst a band at 1045 cm.sup.−1 was not found here, a band at 1036 cm.sup.−1 that might be related to the same component was found. However a more detailed analysis would be required to confirm this. As well as the aforementioned possible change in glycogen, DNA/RNA are well known to contributors in the 1100-900 cm.sup.−1 spectral region. Since the EP contains more cells and therefore nuclei, it was likely that some of the changes in this 1100-900 cm.sup.−1 region reported here were attributed to changes in DNA.
(134) The disease stage differences between the ATR-FTIR measurements of fresh biopsies, from the predominantly EP group, was seen in the 1273 and 1738 cm.sup.−1 bands. However, these bands only showed a difference between healthy tissue and cancerous tissue. The 1273 cm.sup.−1 this peak was larger in healthy tissue compared to the other diseased EP signatures, however, this band was not supported by FTIR imaging. The trough at 1738 cm.sup.−1 in cancerous tissue was more prominent than all other stages of disease. This part of the spectrum was tentatively assigned to lipid. However, this band was not observed in the FTIR imaging data, which could be due to the fact that the imaging sample was deparaffinised and this process may well have washed away lipid components. To assess whether there were any lipid changes, a more detailed analyses is required.
(135) In conclusion, spectral differences between the EP and LP of the 8 m thick lung biopsy section could be seen in the 1614-1465 cm.sup.−1 region of second derivative spectra; differences which were also observed between CEP and LP of BE biopsies. Spectral signatures showing disease progression in the EP of the tissue section from SQ to carcinoma in situ were seen in 1350-1000 cm.sup.−1 region of both the EP and LP. EP features at 1163, 1095, 1074 and the 1036 cm.sup.−1 were integrated and showed a clear distinction between SQ EP and dysplastic/carcinoma in situ tissue. However, the SNR was not high enough to distinguish between the dysplastic disease stages with integration alone. PCA of the 1100-1030 cm.sup.−1 spectral region showed that further separation of the dysplastic stages could be achieved. Integrals of the LP features at 1334, 1279, 1215 and 1066 cm.sup.−1 showed a clear progression from SQ to carcinoma in situ. This progression could be further resolved using PCA of the 1350-1196 and the 1097-1041 cm.sup.−1 spectral regions.
(136) The fresh lung biopsy dataset recorded with ATR-FTIR spectroscopy had low numbers of samples in each disease class. The biopsy spectra could be separated into their predominant cell types, which were the EP, LP and a mixed class. The differences between the cell types could be seen at the 1614-1465 cm.sup.−1 spectral region, consistent with the cell type differences in the FTIR imaged data. The spectra that contained predominantly EP spectra showed a difference between SQ and carcinoma in the 1273 and 1738 cm.sup.−1 bands of the second derivative spectra. However, these bands could not be used to distinguish between the LGD and HGD disease classes.
(137) Effects of Acetic Acid
(138) Two experiments were carried out:
(139) 1. To test the possible effects acetic acid has on tissue in a controlled environment two concentrations of acetic acid were sprayed onto porcine oesophagus. Small sections of the oesophagus were then cut and measured using IR spectroscopy.
(140) 2. The effects of acetic acid, throat spray, 1/100,000 adrenaline, NAC and throat spray on human tissue were analysed by comparing IR measurements of human biopsy tissue from patients with and without the use of the drugs.
(141) The spectral changes associated with acetic acid on human and porcine tissue were then compared.
(142) (i) Porcine Samples
(143) Method
(144) Oesophaguses from two different pigs were used, and the experiment was conducted approximately 3 hours after the pigs were slaughtered. The oesophaguses were transported from the abattoir to the lab on ice and were then dissected and washed with distilled water to remove any remaining food in the gullet.
(145) Table 5 shows number of samples and the number of measurements recorded in each condition. All samples were handled with tweezers and between each measurement the sample was lifted and this prism cleaned with distilled water and allowed to dry.
(146) After thoroughly washing the oesophaguses with distilled water, two samples were cut from each and measured. Part of the oesophagus was then washed with 2.5% acetic acid; the tissue was cut and immediately measured. This was repeated for the 5% acetic acid on a new area of the oesophagus.
(147) Measurement Parameters
(148) A Perkin Elmer Spectrum 2 fitted with a single bounce diamond ATR prism and a DTGS detector was used. Spectra were recorded in absorbance mode between 4000 and 400 cm.sup.−1 at a 1 cm.sup.−1 resolution with 10 co-added and averaged scans. The resolution of the spectra was then subsequently reduced to 4 cm.sup.−1 to improve signal to noise and to equal that of the spectrometer used in the human acetic acid experiments.
(149) Analysis
(150)
(151) There appears to be a band in the distilled water-washed tissue at 1399 cm.sup.−1 that was shifted to 1412 cm.sup.−1 in the tissue samples washed with 2.5% and 5% acetic acid. This band is not consistent with the reference spectra and is most probably due to a conformational change. The main component of this band is the amide III bond in proteins, which supports a protein conformational change hypothesis.
(152) It is possible to correct the spectra measured from tissue where acetic acid has been produced, this can be done by a simple subtraction of a spectrum of acetic acid from the sample spectrum, followed by a correctional shift of the amide III bands. Therefore the evidence presented here supports the use of our algorithm with or without the use of acetic acid.
(153) (ii) Human Tissue
(154) Method
(155) During a routine endoscopy, samples were intercepted from patients consented onto the BOOST study at UCL. The samples were transported on ice and in a moist sealed environment to the lab where they were measured. Some patients did not have any topical drugs sprayed on the surface of their oesophagus during the procedure. These patients were named as ‘no drugs’ in this analysis. Some patients had one of the following drugs sprayed onto their oesophagus: 2.5% acetic acid, 1:100,000 adrenaline, NAC or throat spray. Table 6 shows the breakdown of patient, sample and spectra numbers.
(156) Measurement Parameters
(157) Spectra were recorded in absorbance mode between 4000 and 400 cm.sup.−1 using a Bruker Optics IFS 66/s FTIR spectrometer. A liquid nitrogen cooled MCT-A detector, KBr beamsplitter and a carbon globar was used. The aperture was set to 1.5 mm and a scanner velocity of 40 kHz was used. The spectrometer was purged with dried air. All measurements were recorded at 4 cm.sup.−1 resolution, giving a peak accuracy of approximately ±1 cm.sup.−1. Spectra were recorded in ATR mode with a SensIR 3-reflection silicon prism with ZnSe optics, 1000 background interferograms of the clean prism surface were averaged (taken after carefully cleaning the prism with water and 100% ethanol) and, after orienting the sample onto the prism, 500 interferograms were averaged to produce a single sample absorbance spectrum. All ATR-FTIR spectra were recorded using Bruker OPUS 6.5 software.
(158) Analysis of the Effects of Acetic Acid
(159)
(160) The effects of acetic acid of porcine tissue in this experiment were much greater than the effects seen in human samples. This was to be expected as the porcine samples were prepared and analysed in a controlled environment in which the oesophagus was held horizontal, allowing the acetic acid to sit on the tissue for an extended amount of time. In reality, when acetic acid is applied in vivo to a human, it quickly runs off in to the stomach and is further washed away by saliva.
(161) Analysis of the Effects of 1:100,000 Adrenaline
(162) The effects can be seen
(163) Analysis of the Effects of NAC
(164) The effects can be seen in
(165) Analysis of the Effects of Throat Spray
(166) The effects can be seen in
(167) Tables
(168) TABLE-US-00001 TABLE 1 Total number of patients, biopsies, and ATR-FTIR spectra recorded at each disease stage according to their histological diagnosis, after the removal of outliers. Patients Biopsies Spectra SQ 70 87 167 NDBE 75 222 412 HGD 10 31 58 EAC 21 39 73 TOTAL 122 379 710
(169) TABLE-US-00002 TABLE 2 Confusion matrix for the prediction of SQ spectra versus NDBE/HGD/EAC spectra PLSDA with a leave-one-patient out cross validation applied to the 1385-1235 and 1192-1130 cm.sup.−1 regions. Actual class SQ NDBE/HGD/EAC Sen Spe Predicted SQ 107 7 0.64 0.99 class NDBE/HGD/EAC 60 536 0.99 0.64
(170) TABLE-US-00003 TABLE 3 Confusion matrix for the prediction of SQ/NDBE or HGD/EAC when including an inconclusive group on a per biopsy basis. Actual class SQ/NDBE HGD/EAC Sen Spe Predicted SQ/NDBE 223 2 0.83 0.97 class HGD/EAC 46 60 0.97 0.83 Inconclusive 60 7 Inconclusive Rate: 0.18
(171) TABLE-US-00004 TABLE 4 Distribution of the sidedness of normal bronchiole biopsies. A) Distribution of all the spectra across the possible IR cell types B) Distribution of pairs of spectra from the same biopsy: whether they had the same cell type, different cell types, or whether the biopsy had only one spectrum. Table A Spectra Epithelium 41 Lamina propria 16 Mixed 24 TOTAL 81 Table B Biopsies Epithelium only 9 Epithelium and lamina propria 6 Epithelium and mixed 14 Lamina propria only 2 Mixed only 2 Biopsies with single spectra 9 TOTAL 42
(172) TABLE-US-00005 TABLE 5 Pig data recorded (number of samples and the number of measurements recorded in each condition). Number of Number of Number of Condition: Washed with Pigs samples Spectra Distilled water only 2 4 7 (4 from the epithelium and 3 from the underlying tissue) Distilled water followed 2 4 8 (4 from the by 2.5% acetic acid epithelium and 4 from the underlying tissue) Distilled water followed 2 4 11 (5 from the by 5% acetic acid epithelium and 6 from the underlying tissue)
(173) TABLE-US-00006 TABLE 6 Human tissue data (shows the breakdown of patient, sample and spectra numbers). Number Number of Number of Drug of patients biopsy samples biopsy spectra No drugs 117 367 698 2.5% Acetic acid 3 10 24 1:100,000 adrenaline 4 8 16 NAC 2 5 10 Throat spray 7 13 25
REFERENCES
(174) 1. Bhat, S. et al. Risk of malignant progression in Barrett's esophagus patients: results from a large population-based study. J. Natl. Cancer Inst. 103, 1049-1057 (2011). 2. Simard, E. P., Ward, E. M., Siegel, R. & Jemal, A. Cancers with increasing incidence trends in the United States: 1999 through 2008. CA Cancer J Clin 62, 118-128 (2012). 3. Reid, B. J., Li, X., Galipeau, P. C. & Vaughan, T. L. Barrett's oesophagus and oesophageal adenocarcinoma: time for a new synthesis. Nat Rev Cancer 10, 87-101 (2010). 4. Fitzgerald, R. C. et al. British Society of Gastroenterology guidelines on the diagnosis and management of Barrett's oesophagus. gut 63, 7-42 (2014). 5. Haidry, R. J. et al. Radiofrequency ablation for early oesophageal squamous neoplasia: outcomes form United Kingdom registry. World J Gastroenterol 19, 6011-6019 (2013). 6. Haidry, R. J. et al. Improvement over time in outcomes for patients undergoing endoscopic therapy for Barrett's oesophagus-related neoplasia: 6-year experience from the first 500 patients treated in the UK patient registry. gut (2014). doi:10.1136/gutjnl-2014-308501 7. Levine, D. S. et al. An endoscopic biopsy protocol can differentiate high-grade dysplasia from early adenocarcinoma in Barrett's esophagus. Gastroenterology 105, 40-50 (1993). 8. Kerkhof, M. et al. Grading of dysplasia in Barrett's oesophagus: substantial interobserver variation between general and gastrointestinal pathologists. Histopathology 50, 920-927 (2007). 9. Downs-Kelly, E. et al. Poor Interobserver Agreement in the Distinction of High-Grade Dysplasia and Adenocarcinoma in Pretreatment Barrett's Esophagus Biopsies. Am J Gastroenterol 103, 2333-2340 (2008). 10. Chisholm, J. A., Mayne, G. C., Hussey, D. J. & Watson, D. I. Molecular biomarkers and ablative therapies for Barrett's esophagus. Expert Review of Gastroenterology & Hepatology 6, 567-581 (2012). 11. Sharma, P. et al. Standard endoscopy with random biopsies versus narrow band imaging targeted biopsies in Barrett's oesophagus: a prospective, international, randomised controlled trial. gut 62, 15-21 (2012). 12. Curvers, W. et al. Mucosal morphology in Barrett's esophagus: interobserver agreement and role of narrow band imaging. Endoscopy 40, 799-805 (2008). 13. Curvers, W. L. et al. Endoscopic trimodal imaging versus standard video endoscopy for detection of early Barrett's neoplasia: a multicenter, randomized, crossover study in general practice. Gastrointestinal Endoscopy 73, 195-203 (2011). 14. Evans, J. A. & Nishioka, N. S. The Use of Optical Coherence Tomography in Screening and Surveillance of Barrett's Esophagus. Clinical Gastroenterology and Hepatology 3, S8-S11 (2005). 15. Lovat, L. & Bown, S. Elastic scattering spectroscopy for detection of dysplasia in Barrett's esophagus. Gastrointestinal Endoscopy Clinics of North America 14, 507-517 (2004). 16. Zhu, Y. et al. Elastic scattering spectroscopy for detection of cancer risk in Barrett's esophagus: experimental and clinical validation of error removal by orthogonal subtraction for increasing accuracy. J Biomed Opt 14, 044022-044022-7 (2009). 17. Sturm, M. B. et al. In Vivo Molecular Imaging of Barrett's Esophagus With Confocal Laser Endomicroscopy. Gastroenterology 145, 56-58 (2013). 18. Goetz, M. & Kiesslich, R. Confocal endomicroscopy: In vivo diagnosis of neoplastic lesions of the gastrointestinal tract. Anticancer Res 28, 353-360 (2008). 19. Estores, D. & Velanovich, V. Barrett esophagus: epidemiology, pathogenesis, diagnosis, and management. Current Problems in Surgery 50, 192-226 (2013). 20. Almond, L. M. & Barr, H. Advanced endoscopic imaging in Barrett's oesophagus. International Journal of Surgery 10, 236-241 (2012). 21. Diem, M. et al. Applications of Infrared and Raman Microspectroscopy of Cells and Tissue in Medical Diagnostics: Present Status and Future Promises. Spectroscopy: An International Journal 27, 463-496 (2012). 22. Kendall, C. et al. Vibrational spectroscopy: a clinical tool for cancer diagnostics. Analyst 134, 1029-1045 (2009). 23. Wang, T. D. et al. Detection of endogenous biomolecules in Barrett's esophagus by Fourier transform infrared spectroscopy. Proc Natl Acad Sci USA 104, 15864-15869 (2007). 24. Quaroni, L. & Casson, A. G. Characterization of Barrett esophagus and esophageal adenocarcinoma by Fourier-transform infrared microscopy. Analyst 134, 1240-1246 (2009). 25. Zhao, R., Quaroni, L. & Casson, A. G. Fourier transform infrared (FTIR) spectromicroscopic characterization of stem-like cell populations in human esophageal normal and adenocarcinoma cell lines. Analyst 135, 53-61 (2009). 26. Maziak, D. E. et al. Fourier-transform infrared spectroscopic study of characteristic molecular structure in cancer cells of esophagus: an exploratory study. Cancer Detect. Prev. 31, 244-253 (2007). 27. Wang, J.-S. et al. FT-IR spectroscopic analysis of normal and cancerous tissues of esophagus. World J Gastroenterol 9, 1897-1899 (2003). 28. Kendall, C. A. et al. Raman spectroscopy for the diagnosis of dysplasia in columnar and squamous epithelium. in 4161, 131-137 (SPIE, 2000). 29. Shetty, G., Kendall, C., Shepherd, N., Stone, N. & Barr, H. Raman spectroscopy: elucidation of biochemical changes in carcinogenesis of oesophagus. Br. J. Cancer 94, 1460-1464 (2006). 30. Li, S.-X. et al. Study of support vector machine and serum surface-enhanced Raman spectroscopy for noninvasive esophageal cancer detection. J Biomed Opt 18, 027008-027008 (2013). 31. Kendall, C. et al. Raman spectroscopy, a potential tool for the objective identification and classification of neoplasia in Barrett's oesophagus. J. Pathol. 200, 602-609 (2003). 32. Hutchings, J., Kendall, C., Shepherd, N., Barr, H. & Stone, N. Evaluation of linear discriminant analysis for automated Raman histological mapping of esophageal high-grade dysplasia. J Biomed Opt 15, 066015-066015 (2010). 33. Almond, L. M. et al. Endoscopic Raman spectroscopy enables objective diagnosis of dysplasia in Barrett's esophagus. Gastrointestinal Endoscopy 79, 37-45 (2014). 34. Kazarian, S. G. & Chan, K. L. A. ATR-FTIR spectroscopic imaging: recent advances and applications to biological systems. Analyst 138, 1940 (2013). 35. Almond, L. M. et al. Assessment of a custom-built Raman spectroscopic probe for diagnosis of early oesophageal neoplasia. J Biomed Opt 17, 81421-81421 (2012). 36. Kendall, C. et al. Evaluation of Raman probe for oesophageal cancer diagnostics. Analyst 135, 3038-3041 (2010). 37. Gajjar, K. et al. Diagnostic segregation of human brain tumours using Fourier-transform infrared and/or Raman spectroscopy coupled with discriminant analysis. Anal. Methods 5, 89 (2012). 38. Dorling, K. M. & Baker, M. J. Rapid FTIR chemical imaging: highlighting FPA detectors. Trends Biotechnol. 31, 437-438 (2013). 39. De Jong, S. SIMPLS: an alternative approach to partial least squares regression. Chemometrics and Intelligent Laboratory Systems 18, 251-263 (1993).