Clinical Spectroscopy 3 (2021) 100008

Contents lists available at ScienceDirect

Clinical Spectroscopy
journal homepage: www.elsevier.com/locate/clispe

Breast cancer detection using infrared spectral pathology from H&E stained
tissue on glass slides
Jiayi Tang a,b, Daniela Kurfürstová c, Peter Gardner a,b, *
a
Manchester Institute of Biotechnology, University of Manchester, 131 Princess Street, Manchester, M1 7DN, UK
b
Department of Chemical Engineering and Analytical Science, School of Engineering, University of Manchester, Oxford Road, M13 9PL, UK
c
Department of Clinical and Molecular Pathology, Institute of Molecular and Translational Medicine, Faculty of Medicine and Dentistry, Palacky University, Olomouc, Czech
Republic

A R T I C L E I N F O A B S T R A C T

Keywords: Infrared spectral pathology has gained significant attention in the last few years, since it has been demonstrated to be
Infrared able to readily identify cancerous tissue in biopsy samples. The Infrared technique, however, normally requires
Hyperspectral imaging tissue sections to be mounted on infrared transparent slides. Unfortunately, these slides are both expensive and
Pathology
particularly frangible. In addition, mounting samples on specialist slides is an additional step in the sample
Breast cancer
Glass slides
preparation workflow, which ideally should be avoided. Applying infrared imaging directly to the H&E stained
H&E tissue on the glass slides that are normally used by pathologists, could help the infrared imaging technique be
Cancer screening incorporated into current cancer diagnosis work flow and lower the total cost of detection. The disadvantage of using
glass slides is that the spectral range available is restricted to just the high wavenumber region (2500–3600 cm 1). In
this work a study has been conducted on 120 breast tissues biopsy cores from different patients, to demonstrate that
with the limited spectral information, breast cancer can be identified from the H&E glass slides. A four-class
histological Adboost classification model has been constructed. Optimisation of the classification threshold was
carried out to reduce the number of false negatives. Using a threshold of 0.1 the cancerous cores could be detected
with an accuracy of 95.8 %. This was incorporated into a simple traffic light system that could be used as a
prescreening tool. This work, demonstrating the use of infrared spectral pathology on standard pathology samples
slide, thus goes some way to overcome one of the barriers to successful translation of the infrared technique into the
clinic.

1. Introduction Infrared spectroscopy and more specifically infrared hyperspectral

imaging, has received particular attention recently, as it can interrogate
There are several steps in the diagnosis of breast cancer which may tissue samples based on the label-free chemical information inherent in
include a physical examination, a mammogram, ultrasound or MRI the sample [3–10]. Research has shown that infrared spectroscopy can be
imaging. A definitive diagnosis, however, normally needs a biopsy which used to distinguish cancerous and normal samples in a wide range of
involves the removal of a small amount of breast tissue for close different tissue types including, prostate [11–19], lung [20–23], colon
examination under a microscope by a trained pathologist. The pathologists [24] and breast [25–30].
will make decisions considering the deviations in the cell structures and/or Currently infrared imaging normally requires the tissue sections to be
the variations in the distribution of the cells across samples based on their mounted on infrared transparent slides most commonly calcium fluoride
years of training and often extensive personal experience. For most cases (CaF2) or barium fluoride (BaF2). Unfortunately, these are both expensive
this decision making process is considered as the gold standard for cancer and particularly frangible. The high price and nature of substrates
diagnosis [1]. The final judgment, however, is still somewhat subjective, therefore increases the difficulty of adopting infrared imaging into
and can potentially lead to considerable variability in diagnosis [2]. To current clinical practice. At the current costs of a 75 mm x 25 mm x1 mm
alleviate the problem and to help improve the reliability of diagnosis, the CaF2 slide, (50x the cost of a standard glass pathology slide) this is likely
development of automated cancer detection tools, which can augment the to be prohibitively expensive for adoption in the National Health Service
pathologists’ judgement, is curial. Such automated tools can be used to in the UK. The cost of CaF2 slides might drop dramatically if there was
provide a second opinion for patients or as a pre-screening tool for wide-scale adoption but whether this reduction in price would be
pathologists or even post-screening for quality control. sufficient is questionable. The use of cheaper infrared reflecting

* Corresponding author at: Manchester Institute of Biotechnology, University of Manchester, 131 Princess Street, Manchester, M1 7DN, UK.
E-mail address: [email protected] (P. Gardner).

http://doi.org/10.1016/j.clispe.2021.100008
Received 10 October 2020; Received in revised form 23 March 2021; Accepted 24 March 2021
2666-0547/© 2021 Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
J. Tang et al. Clinical Spectroscopy 3 (2021) 100008

substrates is controversial due to the so-called infrared electric field The microscope utilises 15 Cassegrain optics (NA = 0.62) with a
standing wave effect which introduces non-Beer Lambert behaviour [31]. resultant field of view of 704  704 mm and a corresponding pixel size of
In an attempt to address issue of substrates, Bassan et al. proposed, and 5.5 mm.
subsequently tested, applying infrared imaging directly to glass slides
substrate [32]. The down side of this approach, however, is that glass is 2.3. Experiment Procedure
not transparent over most of the diagnostically useful infrared range,
meaning approximately two thirds of the spectrum, at the longer Infrared hyperspectral images were collected using transmission
wavelengths, is lost. In a follow up study Pilling et al. showed that not only mode. The spectral resolution used was 5 cm 1 and for the intererograms
could you use glass slides but you could also use haematoxylin and eosin a Happ-Genzel apodisation was used. The background and sample spectra
(H&E) stained sections on glass i.e. samples exactly as used by the were obtained using 128 and 96 coadded scans respectively. Background
pathologists, thus eliminating the need for a non-standard substrate. In spectra were collected from a clean area of the glass slide some distance
their study Pilling et al. used 182 prostate tissue cores obtained from 100 away from the coverslip with the infrared light passing through the glass
different patients, on 18 separate H&E slides [13]. Utilising a Random slide only. The H&E stained slide was loaded in the purge box
Forest classification model, they were able to demonstrate that they could (surrounding the microscope sample stage) overnight to minimse the
rapidly classify four classes of histology of an independent test set with a effect of water vapour and spectra were only acquired once the relative
high degree of accuracy (>90 %) even though only the high wavenumber humidity of the sample compartmet was below 1%.
region of the spectrum was available. In this current paper we
demonstrate that similar methodology can be used on breast tissue, thus 2.4. Data Analysis
strengthening the feasibility of cancer diagnosis using H&E stained tissue
on glass slides and infrared hyperspectral imaging. All data were pre-processed with MATLAB 2018a. Infrared spectra for
each biopsy core were extracted from the mosaic as a 256  256 = 65,536
2. Methodology spectral array with each spectrum containing 364 data points.
False greyscale chemical images obtained for each of the breast tissue
2.1. Sample preparation core were rendered using the peak height of the amide A band at 3298 cm
1
. The IR images were compared to the H&E stained sections, and both
The sample used for this project is a breast TMA H&E slide, BR20832 epithelium and stroma areas were identified and annotated on the IR
(prepared by US BioMax.Inc). The breast tumour tissue microarray images, based on WHO Classification of Tumours in the Breast [33]. The
contains 192 cases of breast carcinoma, 13 normal associated tissue software used for annotation were CaseViewer 2.3.0.99276 and GIMP
(NATs) and 3 normal breast tissue cores. Each core is from one patient, so 2.8.14.
in total 208 patients are represented. The paraffin embed slide was To simplify the cancer prediction problem, this research was
mounted onto a glass slide and H&E stained after de-waxing. The conducted using only cores from grade II breast cancer patients.
dewaxing protocol involved warming to 60  C for 20 min, before Differences caused by grades of breast cancer are not discussed. The
immersion in xylene for 10 min, followed by immersion in 100 %, 95 % dataset was separated into two groups, a training set and an independent
and 70 % ethanol each for 5 min and then dried at room temperature. A test set. The training set contained 17 cancerous and 7 NAT cores while
cover glass slide was applied to the TMA after H&E staining. the independent test set contained 4 cancerous and 2 NAT cores in total 6
cores. Four classes (cancerous epithelium, NAT epithelium, cancerous
2.2. Instrumentation stroma and NAT stroma) were under-sampled before model training to
eliminate bias caused by unbalanced datasets. The same number of pixels
The infrared hyperspectral images were acquired using an Agilent 670 was randomly selected from each core. The number of pixels annotated
FTIR spectrometer connected to an Agilent 620 infrared imaging from each class were 5952, 3135, 3570 and 31,050 for cancerous
microscope incorporating a liquid nitrogen cooled 128  128 Focal epithelium and stroma and NAT epithelium and stroma respectively. To
Plane Array (FPA) detector. Resolutions Pro 5.2 software was used to build the model, the maximum number of pixels available to achieve a
adjust and control spectrometer and microscope during measurement. balanced data set i.e. 3135 were randomly selected used for each class.

Fig. 1. Examples of A) images of 16 H&E stained breast biopsy cores B) annotation on Infrared based chemical images of corresponding cores with red representing
cancerous epithelium, purple cancerous stroma, green non-cancerous epithelium, orange noncancerous stroma.

2
J. Tang et al. Clinical Spectroscopy 3 (2021) 100008

For the independent test dataset, 624, 2908, 1198 and 20,154 pixels almost certainly due to the fact that the cancerous tissue contains more
were used for cancerous epithelium and stroma, and NAT epithelium and holes in the tissue and therefore contains more mounting medium.
stroma respectively. Later the constructed model was applied to 120 However, what is also noticeable is that the Amide A band has structure to
breast tissue cores. it. The peak position is different in the epithelium and the stroma being
Spectra were quality tested based on the height (absorbance) of the 3300 cm-1 in the former and 3320- 3330 cm-1 in the latter. There are also
amide A band. Spectra that have absorbance between 0.1 and 1.2 were subtle changes in peak shape.
retained. The spectral range 3040 to 3600 cm 1 was used to avoid the Although difficult to interpret by eye, these differences can be used for
possible saturation of the CH stretch region, caused by the mounting classification using multivariate analysis. A classification model was
material of the slides. Each spectrum was converted to first derivative established with a training dataset having 12,540 spectra which consist of
with a Savitzky Golay with 19 point smoothing [13]. four classes of spectra in equal number. An independent test data set was
The AdaBoost classification method was applied to construct the constructed as well where, for a single test the highest classification
model [34]. It is the first practical boosting algorithm, and remains one of accuracy is 99.5 % for NAT stroma spectra and the lowest accuracy is 75.2
the most widely used and studied, with applications in numerous fields % for the NAT epithelium spectra. To further validate the independent
[35]. AdaBoost works by obtaining weighted majority votes of weak classification results, all data processing procedures from annotated pixel
hypothesis. Each hypothesis is assigned weight and conducted by every extraction to model construction were repeated ten times and the
weak learner [35]. In this study, an AdaBoost classification model was classification accuracies of each class was averaged and shown in the
constructed using MatLab 2018a built-in Statistical and Machine Table 1. The classification accuracies obtained each time is consistent.
Learning Toolbox. The AdaBoost M2 algorithm was used for this four- The highest classification accuracy is for NAT stroma, 99.6 %, and the
class problem. 500 iterations were applied with learning rate (to train an lowest is on NAT epithelium, which is 73.1 %. The low accuracy
ensemble using shrinkage) equalled to 0.1. associated with the NAT epithelium may be related to the fact that the
normal associated tissue may still contain information from a benign
3. Results and discussion condition and these benign conditions are not identical. For example, M1,
M3 and M4 are NAT from benign, adenosis, M9 is NAT from benign,
Spectra from annotated pixels from each training core were extracted fibrocystic mastopatia and M13 is from benign adenomyoepithelioma.
for analysis and an example is shown in Fig. 1. A typical ROC curve for a single run of the model is shown in Fig. 3. To
The average spectra of each class in the training set are shown in Fig. 2, quantify the model performance in AUC value, 0.99, 0.99, 0.99 and 0.99
from 2500 cm 1 to 3600 cm 1.The absorption bands between 2800 cm 1 are obtained, which indicates good performance as they all very close to 1.
and 3000 cm-1 are CH stretching modes that are predominantly from the The model was then applied to all the grade II cores measured from
mounting medium used to mount the tissue on the glass slide. There will Br20832 TMA. False colour images of 120 cores was generated and shown
also be a contribution to the CH stretching bands from the lipid material in Fig. 4, where red represents predicted cancerous epithelium, purple
the tissue (mainly cell membranes) but this is largely obscured. The shows the predicted stroma in the cancerous cores, green is the predicted
absorption bands between 3000 cm-1 and 3600 cm-1 are predominantly NAT epithelium regions and yellow shows the stroma areas in the
the amide A at 3300 cm-1 and the amide B at 3060 cm-1 due to the N H predicted NAT cores.
stretch and C–H stretch in proteins. Importantly there is also a To further determine whether a core is cancerous or not, the traffic
contribution to the amide A from the first overtone of the Amide I light colour system was applied to all 120 cores. In the final step of core
vibration (fundamental at 1650 cm-1) which means that information assignment, after obtaining classification results from the model, the
from this diagnostically useful band, that is obscured by the glass number of spectra predicted as cancerous and NAT were compared.
absorption, is captured at the higher wavenumber albeit at very much If the number of cancerous spectra, saying either predicted cancerous
reduced intensity. What is immediately obvious, however, is that there is epithelium or stroma spectra, is larger than the number of NAT spectra
a noticeable difference in the cancerous spectra compared with the non- (predicted NAT epithelium and stroma respectively), the core is assigned
cancerous spectra. The ratio of the CH stretching bands to the protein as a cancerous core. In this case, the comparison follows the rule of
bands is very different, being much higher in the cancerous spectra. This is majority wins. However, the real clinical conditions, the simple majority-

1
Fig. 2. Average spectra of four classes of training spectra of the classification model (2500-3700 cm ), where blue stands for cancerous epithelium, red stands for
cancerous stroma, yellow stands for NAT epithelium and purple stands for NAT stroma.

3
J. Tang et al. Clinical Spectroscopy 3 (2021) 100008

Table 1
Average of classification accuracies among four classes from the annotated areas repeated for ten times.
Average of ten classification process from annotated areas Predicted / %

True / % Cancerous Cancerous NAT NAT

Epithelium Stroma Epithelium Stroma
Cancerous 88.6 7.1 4.3 0
Epithelium
Cancerous 4.6 94.3 0 1.1
Stroma
NAT 21.0 3.1 73.1 2.8
Epithelium
NAT 0 0.1 0.3 99.6
Stroma

95.0 %. The threshold was reduced to 0.2 which meant that if the
predicted cancerous spectra was larger than 20 % of the total number of
predicted epithelium spectra, or the number of the predicted cancerous
stroma spectra was larger than 20 % the total number of predicted stroma
spectra. The classification accuracy in cores is 95.8 %, where 115 of the
120 cores were classified correctly, (Fig. 5 C). Finally when the threshold
was reduced to 0.1, (Fig. 5 D), where 115 of the 120 cores were classified
correctly, he classification accuracy is 95.8 %.
The summarised classification accuracies were shown in the Table 2. It
can be seen that with the decrease of threshold, from 0.5 to 0.2, the
classification accuracy increases. When the threshold is reduced from 0.2
to 0.1, the classification accuracy stays the same. However, if we pay
attention the individual mis-classified cores, the number of mis-classified
cancerous cores is reduced, while the number of mis-classified NAT cores
is increased. In term of cancer diagnosis, a false positive is generally better
than a false negative although both must be minimised. In the present
study therefore, the optimum threshold for this model is 0.1.
In order to investigate why cores in yellow boxes (shown in Fig. 5) are
miss-classified, the cores have been re-examined with regards to the
original Biomax classification. It was found that most of the miss-
classified cores (for example: A5, B11, F9 and G11) are grade III breast
cancer cores instead of grade II as labelled in the slide information. It is
Fig. 3. A typical ROC of the independent test of four-class model classification on likely that the overall tumours from those patients are grade II tumours,
the annotated areas of cores. but the thin-sliced sample cores that were supplied for this study are
substantially grade III. In addition, given this mismatch between the
win rule may not be appropriate, since if a pathologist observed even a Biomax classification and our pathology classification of the tissue it is
small amount of cancerous cells, the sample will still be classified as a possible that the initial training may not have been as robust as it might
cancerous core (only applies when the number of cells in the sample is have been. With that information, when the threshold is reduced, there is
larger than six [36]). With consideration of miss-classification rate, what a chance those miss-classified spectra in the first place will lead to miss-
is the best comparison threshold value for core assignment? Four typical label of cores in the traffic light system.
thresholds, 0.5, 0.3, 0.2, 0.1 were applied to in the final assignment
process to answer the previous question. In addition, to reduce false- 4. Conclusion
negative rate for cancer diagnosis, the threshold for NAT diagnosis was
raised to 0.9. In this paper we have demonstrated that, with careful pre-processing
A threshold of 0.5 was first applied, which represented the ratio process and data selecting method, FTIR hyperspectral images of H&E
between predicted cancerous epithelium/stroma spectra and the total stained breast tissue on glass can be used for cancer prediction. Using the
number of predicted epithelium/stroma spectra. Both the presence of AdaBoost algorithm classification accuracies for grade II breast cancer
predicted epithelium and stroma spectra were used for traffic light and NAT breast tissues of 94 and 99 % were obtained for the stoma
decision making process. A core would be coloured red, considered tissue. The accuracies for epithelial tissue were slightly less at 88 and 73 %
cancerous, if the presence of either predicted cancerous epithelium or respectively.
stroma spectra was higher than the threshold. A core would be coloured Importantly this study supports the work of Pilling et al. who
green if the ratio between number of predicted NAT epithelium/stroma performed a similar study on prostate cancer tissue and demonstrates the
spectra (must be larger than the predicted cancerous epithelium/ stroma wider applicability of this methodology. This has implications for possible
spectra) and the total number of predicted epithelium spectra is higher introduction of the IR hyperspectral imaging based diagnostics. Being
than 0.9. Any other cases would be coloured in amber, which means a able to measure IR spectra on samples that are currently used minimises
further observation should be conducted to draw a conclusion. The false disruption to the current sample preparation workflow. There are also no
colour image of these cores is shown in Fig. 5 A. It can be seen that 108 of additional sample preparation costs or any additional cost of substrates.
the 120 cores were classified correctly. The classification accuracy in Being able to work on glass slides removes significant barrier to clinical
cores is 90.0 %. When the threshold was reduced to 0.3, 114 of the 120 adoption but at present it has only been demonstrated as a screening tool.
cores were classified correctly (Fig. 5 B). The classification accuracy is In order to obtain new information in future studies, need to include more

4
J. Tang et al. Clinical Spectroscopy 3 (2021) 100008

Fig. 4. False colour images of 120 grade II breast cancer cores, where red indicates cancerous epithelium, purple indicates cancerous stroma, green indicates NAT
epithelium and yellow shows NAT stroma.

Fig. 5. False colour images of 120 grade II breast cancer cores using traffic light system with a threshold of A) 0.5, B) 0.3, C) 0.2, D) 0.1, where red represents cancerous,
green shows normal cores and amber represents cores that could not be classified as either cancerous or normal and therefore should be looked at by a pathologist. The
yellow boxes around a core indicate cores which were wrongly classified judged against the original Biomax classification.

5
J. Tang et al. Clinical Spectroscopy 3 (2021) 100008

Cancer biopsies with Gleason grade and tumour stage, Eur. Urol. 50 (2006) 750–761,
doi:http://dx.doi.org/10.1016/j.eururo.2006.03.031.
Table 2 [12] M.J. Baker, E. Gazi, M.D. Brown, J.H. Shanks, N.W. Clarke, P. Gardner, Investigating
Classification accuracies obtained when applying with different thresholds. FTIR based histopathology for the diagnosis of prostate cancer, J. Biophotonics 2
(2009) 104–113, doi:http://dx.doi.org/10.1002/jbio.200810062.
Threshold Classification accuracy
[13] M.J. Pilling, A. Henderson, J.H. Shanks, M.D. Brown, N.W. Clarke, P. Gardner, Infrared
0.5 90.0 % spectral histopathology using haematoxylin and eosin (H&E) stained glass slides: a
0.3 95.0 % major step forward towards clinical translation, Analyst 142 (2017) 1258–1268, doi:
0.2 95.8 % http://dx.doi.org/10.1039/C6AN02224C.
0.1 95.8 % [14] M.J. Pilling, A. Henderson, B. Bird, M.D. Brown, N.W. Clarke, P. Gardner, High-
throughput quantum cascade laser (QCL) spectral histopathology: a practical
approach towards clinical translation, Faraday Discuss. (2016) , doi:http://dx.doi.
org/10.1039/c5fd00176e.
annotated cores including all three grades of breast cancer (same number [15] R. Bhargava, D.C. Fernandez, S.M. Hewitt, I.W. Levin, High throughput assessment of
of cores in each grade) will be used to establish a more robust breast cells and tissues: bayesian classification of spectral metrics from infrared vibrational
spectroscopic imaging data, Biochim. Biophys. Acta Biomembr. (2006) , doi:http://
cancer prediction model. dx.doi.org/10.1016/j.bbamem.2006.05.007.
[16] R. Bhargava, Towards a practical Fourier transform infrared chemical imaging
Declaration of Competing Interest protocol for cancer histopathology, Anal. Bioanal. Chem. 389 (2007) 1155–1169, doi:
http://dx.doi.org/10.1007/s00216-007-1511-9.
[17] J.T. Kwak, S.M. Hewitt, A.A. Kajdacsy-balla, S. Sinha, R. Bhargava, Automated
The authors declare that they have no known competing financial prostate tissue referencing for cancer detection and diagnosis, BMC Bioinformatics
interests or personal relationships that could have appeared to influence (2016) 1–12, doi:http://dx.doi.org/10.1186/s12859-016-1086-6.
[18] T.P. Wrobel, J.T. Kwak, A. Kadjacsy-balla, R. Bhargava, High-Definition Fourier
the work reported in this paper. Transform Infrared Spectroscopic Imaging of Prostate Tissue, vol. 9791(2016) , pp. 5–
11, doi:http://dx.doi.org/10.1117/12.2217341.
Dedication [19] P. Bassan, A. Sachdeva, J.H. Shanks, M.D. Brown, N.W. Clarke, P. Gardner, Automated
high-throughput assessment of prostate biopsy tissue using infrared spectroscopic
chemical imaging, in: M.N. Gurcan, A. Madabhushi (Eds.), International Society for
We would like to dedicate this paper our very dear friend and Optics and Photonics, 2014, doi:http://dx.doi.org/10.1117/12.2043290 p. 90410D.
collaborator Dr Vlastimil Mašek who sadly passed away on the 14th July [20] X. Mu, M. Kon, A. Ergin, S. Remiszewski, A. Akalin, C.M. Thompson, M. Diem,
2020 at the age of 43. He was a valued member of the International Statistical analysis of a lung cancer spectral histopathology (SHP) data set, Analyst.
140 (2015) 2449–2464, doi:http://dx.doi.org/10.1039/C4AN01832J.
Society for Clinical Spectroscopy (CLIRSPEC) and will be sadly missed by [21] P.D. Lewis, K.E. Lewis, R. Ghosal, S. Bayliss, A.J. Lloyd, J. Wills, R. Godfrey, P. Kloer, L.
all who had the pleasure to know him. A.J. Mur, Evaluation of FTIR Spectroscopy as a diagnostic tool for lung cancer using
sputum, BMC Cancer (2010) , doi:http://dx.doi.org/10.1186/1471-2407-10-640.
[22] F. Großerueschkamp, A. Kallenbach-Thieltges, T. Behrens, T. Brüning, M. Altmayer, G.
Acknowledgements Stamatis, D. Theegarten, K. Gerwert, Marker-free automated histopathological
annotation of lung tumour subtypes by FTIR imaging, Analyst. 140 (2015) 2114–2120,
Special thanks to Dr Alex Henderson for all the data processing doi:http://dx.doi.org/10.1039/C4AN01978D.
[23] C. Krafft, V. Sergo, Biomedical applications of Raman and infrared spectroscopy to
support and constructive discussion on code writing. diagnose tissues, Spectroscopy. 20 (2006) 195–218, doi:http://dx.doi.org/10.1155/
This work was supported in part by the European Regional 2006/738186.
Development Fund - Project ENOCH (No. CZ.02.1.01/0.0/0.0/16_019/ [24] C. Kuepper, F. Großerueschkamp, A. Kallenbach-Thieltges, A. Mosig, A. Tannapfel, K.
Gerwert, Label-free classification of colon cancer grading using infrared spectral
0000868). histopathology, Faraday Discuss. 187 (2016) 105–118, doi:http://dx.doi.org/
The FTIR microscope was funded by a generous donation from the 10.1039/C5FD00157A.
Williamson Trust, UK. [25] H. Fabian, N.A.N. Thi, M. Eiden, P. Lasch, J. Schmitt, D. Naumann, Diagnosing benign
and malignant lesions in breast tissue sections by using IR-microspectroscopy,
Biochim. Biophys. Acta Biomembr. 1758 (2006) 874–882, doi:http://dx.doi.org/
References 10.1016/j.bbamem.2006.05.015.
[26] D.M. Mayerich, M. Walsh, A. Kadjacsy-Balla, S. Mittal, R. Bhargava, Breast
[1] L.B. Rorke, Pathologic diagnosis as the gold standard, Cancer 79 (1997) 665–667, doi: histopathology using random decision forests-based classification of infrared
http://dx.doi.org/10.1002/(SICI)1097-0142(19970215)79:4<665::AID- spectroscopic imaging data, in: M.N. Gurcan, A. Madabhushi (Eds.), Prog. Biomed. Opt.
CNCR1>3.0.CO;2-D. Imaging - Proc., SPIE, 2014, doi:http://dx.doi.org/10.1117/12.2043783 p. 904107.
[2] C. Demir, B. Yener, Automated Cancer Diagnosis Based on Histopathological Images: a [27] P. Bassan, M.J. Weida, J. Rowlette, P. Gardner, Large scale infrared imaging of tissue
Systematic Survey, Dept. Comput. Sci., Rensselaer Polytech. Inst., Troy, NY, USA, micro arrays (TMAs) using a tunable Quantum Cascade Laser (QCL) based microscope,
Tech. Rep, 2005 https://doi.org/10.1.1.61.1199. Analyst. 139 (2014) 3856–3859, doi:http://dx.doi.org/10.1039/C4AN00638K.
[3] M. Pilling, P. Gardner, Fundamental developments in infrared spectroscopic imaging [28] S. Mittal, T.P. Wrobel, L.S. Leslie, A. Kadjacsy-balla, A Four Class Model for Digital
for biomedical applications, Chem. Soc. Rev. 45 (2016) 1935–1957, doi:http://dx.doi. Breast Histopathology Using High- Definition Fourier Transform Infrared (FT-IR)
org/10.1039/C5CS00846H. Spectroscopic Imaging, vol. 9791(2016) , pp. 1–8, doi:http://dx.doi.org/10.1117/
[4] M.J. Baker, H.J. Byrne, J. Chalmers, P. Gardner, R. Goodacre, A. Henderson, S.G. 12.2217358.
Kazarian, F.L. Martin, J. Moger, N. Stone, J. Sulé-Suso, Clinical applications of infrared [29] M. Sattlecker, R. Baker, N. Stone, C. Bessant, Support vector machine ensembles for
and Raman spectroscopy: state of play and future challenges, Analyst (2018) , doi: breast cancer type prediction from mid-FTIR micro-calcification spectra, Chemometr.
http://dx.doi.org/10.1039/c7an01871a. Intell. Lab. Syst. (2011) , doi:http://dx.doi.org/10.1016/j.chemolab.2011.05.007.
[5] S. Tiwari, R. Bhargava, Extracting knowledge from chemical imaging data using [30] M. Verdonck, A. Denayer, B. Delvaux, S. Garaud, R. De Wind, C. Desmedt, C. Sotiriou,
computational algorithms for digital cancer diagnosis, Yale J. Biol. Med. 88 (2015) K. Willard-Gallo, E. Goormaghtigh, Characterization of human breast cancer tissues by
131–143http://www.ncbi.nlm.nih.gov/pubmed/26029012. infrared imaging, Analyst (2016) , doi:http://dx.doi.org/10.1039/c5an01512j.
[6] I.W. Levin, R. Bhargava, FOURIER TRANSFORM INFRARED VIBRATIONAL [31] P. Bassan, J. Lee, A. Sachdeva, J. Pissardini, K.M. Dorling, J.S. Fletcher, A. Henderson, P.
SPECTROSCOPIC IMAGING: integrating microscopy and molecular recognition, Gardner, The inherent problem of transflection-mode infrared spectroscopic microscopy
Annu. Rev. Phys. Chem. 56 (2005) 429–474, doi:http://dx.doi.org/10.1146/annurev. and the ramifications for biomedical single point and imaging applications, Analyst. 138
physchem.56.092503.141205. (2013) 144–157, doi:http://dx.doi.org/10.1039/C2AN36090J.
[7] L.S. Leslie, T.P. Wrobel, D. Mayerich, S. Bindra, R. Emmadi, R. Bhargava, High [32] P. Bassan, J. Mellor, J. Shapiro, K.J. Williams, M.P. Lisanti, P. Gardner, Transmission
definition infrared spectroscopic imaging for lymph node histopathology, PLoS One FT-IR chemical imaging on glass substrates: applications in infrared spectral
10 (2015) e0127238, doi:http://dx.doi.org/10.1371/journal.pone.0127238. histopathology, Anal. Chem. (2014) , doi:http://dx.doi.org/10.1021/ac403412n.
[8] D. Mayerich, M.J. Walsh, A. Kadjacsy-balla, P.S. Ray, S.M. Hewitt, R. Bhargava, Stain- [33] S.R. Lakhani, International Agency for Research on Cancer, World Health
less staining for computed histopathology, Technology (Singap World Sci) 3 (2015) Organization, WHO Classification of Tumours of the Breast, 4th edition, International
27–31, doi:http://dx.doi.org/10.1142/S2339547815200010. Agency for Research on Cancer, Lyon, France, 2012.
[9] K. Malek, B.R. Wood, K.R. Bambery, FTIR imaging of tissues: techniques and methods [34] Y. Freund, R.E. Schapire, A desicion-theoretic generalization of on-line learning and
of analysis, Challenges Adv. Comput. Chem. Phys. (2014) , doi:http://dx.doi.org/ an application to boosting, Comput. Learn. Theory Second Eur. Conf. (1995) 23–37,
10.1007/978-94-007-7832-0_15. doi:http://dx.doi.org/10.1007/3-540-59119-2_166.
[10] I.U. Rehman, R.S. Khan, S. Rehman, Role of artificial intelligence and vibrational [35] B. Schölkopf, Z. Luo, V. Vovk, Empirical inference: Festschrift in honor of Vladimir N.
spectroscopy in cancer diagnostics, Expert Rev. Mol. Diagn. (2020) , doi:http://dx.doi. Vapnik, Empir. Inference Festschrift Honor Vladimir N. Vapnik, (2013) , pp. 1–287,
org/10.1080/14737159.2020.1784008. doi:http://dx.doi.org/10.1007/978-3-642-41136-6.
[11] E. Gazi, M. Baker, J. Dwyer, N.P. Lockyer, P. Gardner, J.H. Shanks, R.S. Reeve, C.A. [36] J. Crocker, P.G. Murray (Eds.), Molecular Biology in Cellular Pathology, John Wiley &
Hart, N.W. Clarke, M.D. Brown, A correlation of FTIR spectra derived from prostate Sons, Ltd, Chichester, UK, 2003, doi:http://dx.doi.org/10.1002/0470867949.