See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/7644848

Diagnosing breast cancer by using Raman spectroscopy

Article in Proceedings of the National Academy of Sciences · September 2005

DOI: 10.1073/pnas.0501390102 · Source: PubMed

CITATIONS READS
553 305

6 authors, including:

Abigail S Haka Karen Elizabeth Shafer-Peltier

Weill Cornell Medical College MRIGlobal
81 PUBLICATIONS 3,874 CITATIONS 39 PUBLICATIONS 3,853 CITATIONS

SEE PROFILE SEE PROFILE

Maryann Fitzmaurice Joseph Crowe

Case Western Reserve University Cleveland Clinic
90 PUBLICATIONS 7,309 CITATIONS 82 PUBLICATIONS 4,547 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Complex fluids for CCUS and subsurface applications View project

Breast Cancer View project

All content following this page was uploaded by Maryann Fitzmaurice on 19 May 2014.

The user has requested enhancement of the downloaded file.

Diagnosing breast cancer by using
Raman spectroscopy
Abigail S. Haka*, Karen E. Shafer-Peltier†, Maryann Fitzmaurice‡, Joseph Crowe§, Ramachandra R. Dasari*,
and Michael S. Feld*¶
*G. R. Harrison Spectroscopy Laboratory, Massachusetts Institute of Technology, Cambridge, MA 02139; †Dade Behring, Newark, DE 19714; ‡University
Hospitals of Cleveland, Case Western Reserve University, Cleveland, OH 44106; and §The Cleveland Clinic Foundation, Cleveland, OH 44195

Edited by Inder M. Verma, The Salk Institute for Biological Studies, La Jolla, CA, and approved July 22, 2005 (received for review February 22, 2005)

We employ Raman spectroscopy to diagnose benign and malignant causing small lesions to go undetected. Furthermore, to obtain
lesions in human breast tissue based on chemical composition. In the sensitivity and specificity required for lesion detection
this study, 130 Raman spectra are acquired from ex vivo samples of in vivo, exogenous agents are often used to improve contrast (9).
human breast tissue (normal, fibrocystic change, fibroadenoma, Optical spectroscopic techniques are also under investigation
and infiltrating carcinoma) from 58 patients. Data are fit by using for breast cancer diagnosis. Unlike DOT, these techniques
a linear combination model in which nine basis spectra represent sample the tissue locally (⬇1 mm3 volume). Light delivery and
the morphologic and chemical features of breast tissue. The re- collection can be accomplished by using optical fibers that can
sulting fit coefficients provide insight into the chemical兾morpho- be incorporated into a biopsy needle. As opposed to biopsy, a
logical makeup of the tissue and are used to develop diagnostic spectroscopic needle measurement has the advantage of pro-
algorithms. The fit coefficients for fat and collagen are the key viding immediate diagnosis. Thus, spectroscopy has the potential
parameters in the resulting diagnostic algorithm, which classifies to reduce both the likelihood of a nondiagnostic needle biopsy
samples according to their specific pathological diagnoses, attain- (requiring repeat needle or surgical biopsy) and patient anxiety

BIOPHYSICS
ing 94% sensitivity and 96% specificity for distinguishing cancer- (by eliminating the currently unavoidable wait for pathology
ous tissues from normal and benign tissues. The excellent results diagnosis). Furthermore, with the development of minimally
demonstrate that Raman spectroscopy has the potential to be invasive breast cancer therapies, such as radiofrequency abla-
applied in vivo to accurately classify breast lesions, thereby reduc- tion, there is the potential for diagnosis and treatment to be
ing the number of excisional breast biopsies that are performed. performed in a single procedure (10).
Optical techniques that have been applied to breast tissue
spectral diagnosis 兩 optical vibrational disease include fluorescence, reflectance, and Raman spectroscopies.
Fluorescence spectroscopy has been applied to ex vivo breast
n the United States, ⬇216,000 new cases of breast cancer are tissue, and trends correlating with disease have been observed
I diagnosed each year, and 40,000 women die from the disease
(1). Mammography, the most common technique for detecting
(11–13). Fluorescence produces relatively large signals. How-
ever, the small number of endogenous fluorophores in breast
nonpalpable, highly curable breast cancer, employs x-rays to tissue and their broad spectral lineshapes are limiting factors.
quantitatively probe density changes in breast tissue. Because There are preliminary studies using diffuse reflectance spec-
these density changes are not uniquely correlated with breast troscopy to diagnose breast lesions (14) by monitoring changes
cancer, mammography serves as a screening technique rather in absorption and scattering. Combining fluorescence and dif-
than a diagnostic tool. Thus, a lesion found through either fuse reflectance spectroscopies has shown promise in the breast,
clinical breast examination or mammography is always biopsied. as well as high sensitivity and specificity for cancer detection in
Because of current limitations, 70–90% of mammographically several other organ systems (15).
detected lesions are found to be benign upon biopsy (2). Breast Raman spectroscopy can provide detailed chemical informa-
biopsy is most often performed by surgical excision that removes tion about a tissue sample and thus insight into the chemical
the entire lesion or by core needle biopsy that removes 5–12 changes that accompany breast disease. In contrast to fluores-
cores of tissue, typically 1 mm in diameter and several centime- cence, there are a large number of Raman active molecules in
ters long, to ensure proper sampling (3). The complete diagnos- breast tissue, and their spectral signatures are sharp and well
tic process, from start to finish, may take months and may delineated. The ability to measure several different chemicals is
include multiple biopsies. of particular importance in studying breast cancer because of the
A desire to reduce the number of biopsies performed on heterogeneity of the disease. Although Raman spectra provide
benign tissue and the patient trauma, time delay, and high high information content, the signals are orders of magnitude
medical costs involved has motivated researchers to explore a weaker than fluorescence. However, with careful system design,
variety of minimally invasive optical imaging and spectroscopy collection of clinical data in relevant times with safe laser powers
techniques to improve breast cancer diagnosis, especially the can be accomplished. For these reasons, we have investigated
ability to distinguish benign lesions from malignant ones. These Raman spectroscopy as a clinical tool for the diagnosis of a
techniques employ visible or near-infrared light, have the po- variety of breast pathologies.
tential to provide chemical information, and are less invasive Raman spectroscopy is an inelastic scattering process in which
than current diagnostic procedures. photons incident on a sample transfer energy to or from mo-
Diffuse optical tomography (DOT) studies the propagation of lecular vibrational modes (16). It is a coherent two-photon
amplitude-modulated pulses of light through the breast. It is
noninvasive and can detect lesions deep within the tissue (4–8). This paper was submitted directly (Track II) to the PNAS office.
An array of sources and detectors in a measurement cup enables
Abbreviations: DEH, ductal epithelial hyperplasia; ROC, receiver operating characteristic;
3D images to be constructed. By using light of different wave- N兾C, nuclear-to-cytoplasm.
lengths, information about scattering and absorption can be ¶To whom correspondence should be addressed at: G. R. Harrison Spectroscopy Laboratory,
extracted to measure oxy- and deoxy-hemoglobin concentrations Massachusetts Institute of Technology, 166 Albany Street NW-14, Cambridge, MA 02139.
and the presence of lipids. However, DOT can detect only a E-mail: [email protected].
limited number of chemicals and is generally of low resolution, © 2005 by The National Academy of Sciences of the USA

www.pnas.org兾cgi兾doi兾10.1073兾pnas.0501390102 PNAS 兩 August 30, 2005 兩 vol. 102 兩 no. 35 兩 12371–12376

process in which a molecule simultaneously absorbs an incident in 5-␮m-thick sections, and stained with hematoxylin兾eosin. The
photon and emits a Raman photon, accompanied by its transi- histological slides were examined by an experienced breast
tion from one energy level to another, giving rise to a frequency pathologist who was blinded to the outcome of the Raman
(i.e., energy) shift of the emitted photon. Because the energy spectroscopy analysis. A total of 130 spectra from 58 patients
levels are unique for every molecule, Raman spectra are chem- were examined by using Raman spectroscopy: 49 normal tissues
ical-specific. Individual bands in the Raman spectrum are char- from 25 patients (20 female, 2 male, and 3 unknown; 8 white, 12
acteristic of specific molecular motions. Raman spectroscopy is black, and 5 unknown) with a mean age of 34.3 years (range of
particularly amenable to in vivo measurements, because the 13–75 years); 50 from benign lesions, 16 with fibrocystic change
powers and excitation wavelengths that are used do not affect the from 16 patients (10 female, 1 male, and 5 unknown; 5 white, 3
tissue and the penetration depth is relatively large (17). black, and 8 unknown) with a mean age of 40.3 years (range of
Early studies based on small sample sets observed spectral 13–75 years), 3 from lesions diagnosed as ductal epithelial
trends, indicating the promise of Raman spectroscopy for breast hyperplasia (DEH) from 2 white female patients with ages of 11
cancer diagnosis (18–21). However, this initial work, which and 49 years, and 31 from fibroadenomas from 6 patients (4
relied on peak height ratios for data analysis, was unable to female and 2 unknown; 2 black and 4 unknown) with a mean age
differentiate benign lesions from malignant lesions. Given the of 20.8 years (range of 13–40 years); and 31 from malignant
intended application, in which there is an a priori expectation of lesions all diagnosed as infiltrating carcinoma (10 ductal, 2
nonnormal tissue, accurate distinction of benign and malignant lobular, 1 ductal and lobular, and 1 mammary, not otherwise
lesions is crucial. In view of the wealth of information available specified) from 16 patients (8 female and 8 unknown; 5 white, 2
from Raman spectroscopy and the biochemical complexity of black, and 9 unknown) with a mean age of 57.6 years (range of
breast lesions, a method of analysis that utilizes the entire Raman 46–77 years). Patient information was not available for a sample
spectrum, rather than peak height ratios, is necessary to distin- that was diagnosed as fat necrosis. Because multiple spectra were
guish between benign and malignant lesions. collected from each patient, some tissue samples are included in
Our own initial research, based on principal component both the normal and diseased categories, depending on the
analysis of tissue Raman spectra, showed that benign and pathology underlying the exact region of data collection.
malignant tumors could be differentiated (22). Although prin- The mean ages and age ranges for the patients in this study
cipal component analysis utilizes the entire Raman spectrum, it reflect the natural age incidence of each lesion (27). The peak
affords little insight into the chemical changes responsible for age of incidence for stromal fibroplasia, the predominant man-
disease diagnosis. To provide information about the chemical ifestation of fibrocystic change encountered in this study, is in the
basis for diagnosis and understand the relationship between a fourth and fifth decades (30s and 40s). That of fibroadenoma is
tissue’s Raman spectrum and its disease state, we developed a much earlier [the third decade (20s)], and fibroadenomas ac-
spectroscopic model of breast tissue (23). This model fits mac- count for the majority of breast lesions requiring biopsy in that
roscopic tissue spectra with a linear combination of basis spectra age range. The peak age of incidence for infiltrating carcinoma
derived from Raman microscopy of various breast tissue mor- is the sixth decade (50s).
phological structures. These basis spectra represent the epithe-
lial cell cytoplasm, cell nucleus, fat, ␤-carotene, collagen, cal- Raman Spectroscopic Measurements. Data were acquired by using
cium hydroxyapatite, calcium oxalate dihydrate, cholesterol-like a Raman system described in refs. 28 and 29. The excitation spot
lipid deposits, and water. This modeling approach is based on the is ⬇100 ␮m in diameter, and light diffusion in the tissue results
assumptions that the Raman spectrum of a mixture is a linear in a sampled volume of ⬇1 mm3. Raman spectra were acquired
combination of the spectra of its components and that signal with a 10- to 30-s integration time, depending on signal intensity,
intensity and chemical concentration are linearly related (24). and a spectral resolution of 8 cm⫺1. The average laser excitation
The resulting fit coefficients yield the contribution of each basis power varied between 100 and 150 mW. The fluences used in this
spectrum to the macroscopic tissue spectrum, thereby elucidat- study are safe for clinical investigations (30). No tissue damage
ing the chemical兾morphological makeup of the lesion. These was observed, either grossly or upon histological review.
same morphological changes are routinely used by pathologists
to diagnose disease. However, unlike conventional pathology, Data Processing. Data processing was preformed as described in
which is subject to interobserver variation, Raman spectroscopy refs. 23 and 28. Model fitting was performed by using a linear
assesses these changes in an objective, reproducible manner combination of basis spectra with a nonnegativity constraint.
without the need for tissue removal (25, 26). The contribution of each basis spectrum, obtained from the
Here, our Raman spectroscopic model is used to characterize model described above, to the breast tissue specimens was
the chemical兾morphological composition of a range of ex vivo acquired by normalizing the fit coefficients (excluding water,
breast tissue specimens and pathologies and to predict the breast because it is applied exogenously) such that they sum to one. To
tissue disease state. The data demonstrate the efficacy of the determine the error in our fit coefficients, we used a ␹2 analysis
model in both diagnosing breast disease and understanding the (31). ␹2 analysis is a well known method for calculating the
chemical兾morphological variations associated with disease pro- goodness of a fit as well as the error associated with model fitting.
gression. The results indicate that the technique has the potential The error bars (one SD), shown in Fig. 3, are generated from this
to be applied in vivo to accurately classify breast lesions, thereby analysis. The Raman spectra in each diagnostic group have
reducing the number of excisional breast biopsies that are different signal-to-noise ratios; thus, mean errors are reported
performed. for each pathology. Fitting errors for the two diagnostic model
components, fat and collagen, are 0.010 and 0.006 for normals,
Materials and Methods 0.040 and 0.022 for fibrocystic change, 0.012 and 0.006 for
Tissue Preparation. Breast tissue was obtained from patients who fibroadenoma, and 0.034 and 0.016 for infiltrating carcinoma,
were undergoing surgical breast biopsy reduction mammoplas- respectively. Errors are slightly larger for fat than for collagen
ties and prophylactic mastectomies. Upon removal, the samples because the Raman spectrum of fat has more similarity to other
were snap-frozen in liquid nitrogen for storage and then pas- model components than that of collagen.
sively thawed at room temperature and kept moist with PBS. Logistic regression, a discriminate analysis technique, was
After spectral acquisition, specimens were marked with India ink used to correlate the normalized fit coefficients with the diag-
to indicate the region sampled, fixed in formalin, routinely nostic categories (32) for all combinations of the eight morpho-
processed, paraffin-embedded, cut through the marked locations logical components in the model. A likelihood ratio test was used

12372 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0501390102 Haka et al.

 BIOPHYSICS
Fig. 1. Model fits. Normalized Raman spectra (solid lines), model fits (dotted lines), residuals (shown below), fit coefficients, and images from hematoxylin兾
eosin (H&E)-stained sections used to make the histopathologic diagnosis for normal breast tissue (A), fibrocystic change (B), fibroadenoma (C), and infiltrating
carcinoma (D). The India ink used to record the region of spectral examination is seen as a black line on the tissue surface in the H&E images. (Scale bars, 100 ␮m.)

to determine which fit coefficients were significant for diagnosis demonstrates that the model accounts for the majority of the
and what probability thresholds, based on these fit coefficients, spectroscopic features observed and describes the data well. The
correctly classified the most samples. To cross-validate our fit coefficients, also displayed in Fig. 1, represent the amount that
algorithm, we used a leave-one-out cross-validation analysis, each model basis spectrum must be weighted to recreate the
which enabled recycling of the data to produce a more robust tissue spectrum, thereby providing insight into the chemical兾
diagnostic algorithm. Using maximum likelihood estimation, we morphological makeup of the tissue. Fit coefficients are a
determined the probability that a breast sample is normal, function of both the concentration of a particular model com-
fibrocystic change, fibroadenoma, or invasive carcinoma. Re- ponent and its Raman scattering cross section (which indicates
ceiver operating characteristic (ROC) curves were generated by the strength of the signal at unit concentration).
changing the probability threshold for assigning a classification. Four of the 130 spectra in our data set were excluded from
The ROC curves in Fig. 4 display unique values of sensitivity and diagnostic algorithm development because of an insufficient
specificity. number of samples: three from benign samples diagnosed as
DEH and one from a benign sample diagnosed as fat necrosis.
Results and Discussion The sample diagnosed as fat necrosis exhibited an intense
Model Fits. To understand the relationship between a tissue fluorescent background, most likely due to lipids oxidized during
sample’s Raman spectrum and its disease state, we examined the tissue necrosis that, with a larger data set, may prove to be
contribution of each model basis spectrum to spectra acquired diagnostic of fat necrosis.
from a variety of pathologies. Model fits to Raman spectra
acquired from normal, benign, and malignant samples of breast Spectral Fit Coefficients and Tissue Morphology. The fit coefficients,
tissue are shown in Fig. 1, with corresponding images from given by the model and normalized to sum to one, represent
hematoxylin兾eosin-stained sections used to make the his- contributions of chemicals and morphological features to the
topathologic diagnosis. The difference between the measured macroscopic tissue spectrum. Fig. 2 shows histograms of the
spectrum and the model fit, the residual, is shown below each average fit coefficients for normal breast tissue, fibrocystic
spectrum. The lack of significant structure in the residuals change, fibroadenoma, and infiltrating carcinoma. In Fig. 2A,

Haka et al. PNAS 兩 August 30, 2005 兩 vol. 102 兩 no. 35 兩 12373
 or with a strong enough Raman scattering cross section to
appreciably contribute to the macroscopic Raman spectrum.
The fit coefficients of the breast lesions indicate a markedly
different chemical兾morphological composition than that of nor-
mal breast tissue. First, the amount of collagen increases in all
abnormal breast tissues. This finding is consistent with known
breast pathology, because lesion formation is often accompanied
by fibrosis, a scarring process characterized by an increased
stromal component, and thus by both proliferation of fibroblasts
and accumulation of collagen. The relative increase in collagen
is most pronounced in fibrocystic change, a benign condition that
can manifest itself as fibrosis, adenosis (an increase in the
number of ducts), or cyst formation (dilation of ducts with fluid).
Each of these changes can occur with or without the others. In
the fibrocystic lesions examined in the present study, increases in
the fit coefficients of the collagen, epithelial cell cytoplasm, and
cholesterol-like basis spectra replace the large contribution from
fat in normal breast tissue.
Fibroadenoma is a benign tumor of a different lineage than all
other lesions in this study (34). It is most closely related to
phylloides tumors, the malignant counterpart of which is not
carcinoma but cystosarcoma phylloides, in which the stroma,
rather than the epithelium, is malignant. Lesions diagnosed as
fibroadenoma show an increased contribution from collagen due
to fibroblast proliferation and accumulation of collagen that
results in expansion of the stroma (35). They also show an
increased contribution from both the cell nucleus and epithelial
cell cytoplasm basis spectra as a consequence of the number of
fibroblasts and epithelial cells present.
Similar to fibroadenoma, lesions diagnosed as infiltrating
carcinoma show an increased contribution from collagen, in this
case due to fibroblast proliferation in response to stromal
Fig. 2. Histograms displaying the average composition of samples diagnosed
as normal (white), fibrocystic change (gray), fibroadenoma (striped), and
invasion by the malignant epithelial cells (35). The fit coefficients
infiltrating carcinoma (black). The one-SD confidence intervals are shown for of such lesions also display an increase in the amount of epithelial
each model component. (A) Data are grouped according to pathological cell cytoplasm and cell nucleus. Both infiltrating carcinomas and
diagnosis: fat (a), collagen (b), cell nucleus (c), epithelial cell cytoplasm (d), fibroadenomas exhibit large increases in the number of cells
calcium oxalate (e), calcium hydroxyapatite ( f), cholesterol-like (g), and ␤- relative to other lesions. Enlargement of cell nuclei is a hallmark
carotene (h). (B) Data are clustered by model component, and each histogram of cancer, and thus a higher nuclear-to-cytoplasm (N兾C) ratio is
cluster is normalized to the largest of the four values. a diagnostic criterion routinely used by pathologists (36, 37). In
our studies, the spectroscopic parameter characterizing the N兾C
ratio is obtained by dividing the fit coefficient of the cell nucleus
data are grouped according to pathologic diagnosis; in Fig. 2B, basis spectrum by that of the epithelial cell cytoplasm basis
data are clustered by model component. In Fig. 2B, each spectrum. Fibrocystic change and fibroadenoma have mean N兾C
histogram cluster is normalized to the largest of the four values parameters of 0.01 and 0.04, respectively, whereas infiltrating
to highlight the relative changes between pathologies. The carcinoma has a much higher mean N兾C parameter of 0.08.
one-SD confidence intervals are also displayed in Fig. 2 to Although clear trends are seen in the mean values, there is
illustrate the variation in fit coefficients of a particular model significant variability of the N兾C parameter within pathologies.
component between samples with the same pathology. Further differences between lesions diagnosed as fibroade-
Examination of the fit coefficients for each pathology provides noma and infiltrating carcinoma exist in the amount of fat
insight into the chemical changes associated with breast disease. present. Samples diagnosed as fibroadenoma have less fat than
The fit coefficients of normal breast tissue indicate that it is those diagnosed as infiltrating carcinoma, as can be seen in Fig.
primarily composed of fat. Normal breast tissue contains both 2, because fibroadenoma is an expansile lesion that grows by
glandular and adipose tissues (33). Glandular tissue consists of pushing the fatty breast tissue aside. Infiltrating carcinoma, in
ducts lined by epithelial cells and a supportive collagenous contrast, infiltrates in between the fat cells, so some adipocytes
extracellullar matrix. Adipose tissue is primarily composed of are retained within the carcinoma.
adipocytes (cells containing large amounts of cytoplasmic fat), Other breast pathologies are much less common and were not
represented in the study specimens. Three samples were diag-
although small quantities of extracellular matrix are present.
nosed as DEH, a proliferation of the ductal epithelium that
Overall, ducts represent only a small volume of the tissue, and
confers an increased risk for breast cancer (38). The mean N兾C
thus our model accurately characterizes normal breast tissue as parameter of these three lesions is 0.05, intermediate between
predominately composed of fat with small contributions from that of normal breast and infiltrating carcinoma, indicating the
collagen. Contributions from fat are particularly prominent potential for detecting precancerous changes in the breast by
because adipose tissue has a large Raman scattering cross section using Raman spectroscopy.
relative to most other model components. The model does not The histograms in Fig. 2 exhibit relatively little contribution
show a contribution from epithelial cells to normal breast tissue. from the two types of calcifications found in breast tissue,
However, the lack of an epithelial cell contribution does not calcium hydroxyapatite and calcium oxalate dihydrate. Because
mean that the normal samples do not contain epithelial cells; of their diagnostic importance, microcalcifications in fresh breast
rather, it means that they are not present in sufficient quantities tissue are not typically made available for scientific research, and

12374 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0501390102 Haka et al.

 Fig. 4. ROC curves illustrating the ability of Raman spectroscopy to separate
Fig. 3. Scatter plot displaying the fat and collagen content for all patholo- lesions diagnosed as infiltrating carcinoma from benign and normal breast
gies encountered in this study. Error bars (one SD) are indicated. Several of the tissues. The ROC curve of two indistinguishable populations, represented by
error bars are smaller than the symbols used to denote each data point. the dashed line, is included for comparison.
Normal, gray stars; fibrocystic change, diamonds; fibroadenoma, triangles;
infiltrating carcinoma, squares; DEH, white stars.
compares the pathologic diagnosis with that of the Raman
diagnostic algorithm for our data set. The algorithm yields a

BIOPHYSICS
thus our study did not include lesions containing microcalcifi- sensitivity of 94% (29兾31), specificity of 96% (91兾95), and
cations. However, in a parallel study, we have shown that the overall accuracy of 86% (108兾126) for detecting infiltrating
Raman spectra of microcalcifications contain significant diag- carcinoma. Also shown in Fig. 3 are the fat and collagen fit
nostic information (39). As this technique moves into a clinical coefficients for the three samples diagnosed as DEH. In future
setting, microcalcifications will be encountered more frequently applications, this diagnostic algorithm could be used in a pro-
and in greater abundance, which should provide additional spective manner, and the fit coefficients of collagen and fat
diagnostic information and increase diagnostic accuracy. simply plotted to determine where they fall in the diagnostic
plane.
Diagnostic Algorithm. The fit coefficients not only provide insight Fig. 4 displays a ROC curve that illustrates the ability of
into the composition of the tissue but also are used to develop Raman spectroscopy to separate lesions diagnosed as infiltrating
diagnostic algorithms. A diagnostic algorithm was developed carcinoma from benign and normal breast tissues. A ROC curve
that examines all of the data simultaneously and is, to our illustrates the tradeoff between sensitivity and specificity by
knowledge, the first spectral-based algorithm to separate breast plotting the true-positive rate against the false-positive rate for
tissues according to specific pathological diagnoses. the different possible probability thresholds of a diagnostic test.
The fit coefficients corresponding to the fat and collagen basis The closer the curve comes to the 45-degree diagonal, shown as
spectra were the key diagnostic parameters in differentiating a dashed line, the less accurate the diagnostic test. The ROC
pathologies. Samples were initially divided into two groups based curve clearly illustrates the ability of Raman spectroscopy to
on their collagen [FC(Coll)] and fat [FC(Fat)] content. One accurately diagnose breast cancer and demonstrates how the
group contained infiltrating carcinomas and fibroadenomas; the diagnostic scheme can be adjusted to obtain the desired degree
second contained normal tissue and fibrocystic lesions. The of sensitivity at the cost of specificity.
equation for the decision line, drawn by logistic regression, is
FC(Coll) ⫽ ⫺0.85FC(Fat) ⫹ 0.60. The resultant two clusters Effect of Age on the Diagnostic Algorithm. The female breast
were then further subdivided by using logistic regression. Again, undergoes substantial biochemical alterations at menopause. Spe-
the diagnostic power of all eight model components was as- cifically, a large amount of collagen is replaced by fat, resulting in
sessed. Fibrocystic change and normal tissue were separated breast tissue that is less dense on mammography. Because our
based on their fat and collagen contents, with the decision line diagnostic algorithm is based on fat and collagen content as assessed
given by FC(Coll) ⫽ ⫺0.06FC(Fat) ⫹ 0.10. Fibroadenoma and by Raman spectroscopy, we investigated trends in these two pa-
infiltrating carcinoma were separated based solely on their fat rameters as a function of patient age. It is notable that although the
content. The decision line for this separation is FC(Fat) ⫽ 0.02. normal samples span a wide age range, our technique characterizes
Fig. 3 displays a scatter plot of FC(Coll) and FC(Fat) for all them as predominately composed of fat because the Raman
pathologies encountered in this study, as well as the decision scattering cross section is much higher for fat than for collagen.
lines that separate samples according to diagnoses. Table 1 Thus, we observe significant contributions from collagen only when

Table 1. Comparison of the pathologic diagnosis with that of the Raman diagnostic algorithm
Pathology diagnosis

Normal Fibrocystic change Fibroadenoma Infiltrating carcinoma

Raman diagnosis (49 spectra) (31 spectra) (15 spectra) (31 spectra)

Normal 45 1 0 0
Fibrocystic change 4 22 0 0
Fibroadenoma 0 7 12 2
Infiltrating carcinoma 0 1 3 29

Haka et al. PNAS 兩 August 30, 2005 兩 vol. 102 兩 no. 35 兩 12375
it is present in large quantities (for instance, in the setting of dense Implications. We have demonstrated the ability of Raman spec-
stromal fibrosis). To rigorously confirm that our diagnostic algo- troscopy to diagnose benign and malignant breast lesions with
rithm is not influenced by patient age, we examined the correlation high sensitivity and specificity ex vivo in a laboratory setting.
between the spectroscopic parameter characterizing the fat-to- These diagnoses are based on chemical兾morphological changes
collagen ratio, obtained by dividing FC(Fat) by FC(Coll), and age. that are known to accompany breast disease. These excellent
Fibroadenoma was excluded from this analysis because it is a ex vivo results are a necessary step toward data acquisition in
juvenile disease and has a much different age range than the other hospital settings, both ex vivo directly following biopsy and in vivo
pathologies in our study (27). We found a correlation coefficient of during breast surgery. These preclinical and clinical studies are
⫺0.140 for the spectroscopic parameter characterizing the fat-to- made possible by the recent development of a Raman optical
collagen ratio and age. We also examined the relationship between
fiber probe designed for medical applications (40). The probe is
the fat-to-collagen ratio and age within individual pathologies,
optimized to collect high signal-to-noise ratio data from tissue in
obtaining correlation coefficients of ⫺0.104, 0.137, and ⫺0.025 for
invasive carcinoma, fibrocystic change, and normals, respectively. clinically relevant times (1 s).
Thus, we do not observe age-dependent trends in the fat and
collagen contents of our data. Our results indicate that Raman This work was supported by National Institutes of Health Grants
spectroscopy is much less sensitive to breast density and meno- RR02594 and HL51265 and Pathology Associates of University Hospi-
pausal status than many other optical techniques. tals, Cleveland.

1. American Cancer Society (2005) Breast Cancer Facts and Figures 2004. Avail- 21. Redd, D. C., Feng, Z. C., Yue, K. T. & Gansler, T. S. (1993) Appl. Spectrosc.
able at www.cancer.org兾docroot兾STT兾stt㛭0㛭2004.asp. Accessed Aug. 5, 2005. 47, 787–791.
2. Johnson, J., Dalton, R., Wester, S., Landercasper, J. & Lambert, P. (1999) Arch. 22. Manoharan, R., Shafer, K., Perelman, L., Wu, J., Chen, K., Deinum, G.,
Surg. (Chicago) 134, 712–715. Fitzmaurice, M., Myles, J., Crowe, J., Dasari, R. & Feld, M. (1998) Photochem.
3. Liberman, L., Dershaw, D. D., Rosen, P. P, Abramson, A. F., Deutch, B. M. Photobiol. 67, 15–22.
& Hann, L. E. (1994) Radiology 192, 793–795. 23. Shafer-Peltier, K. E., Haka, A. S., Fitzmaurice, M., Crowe, J., Dasari, R. R. &
4. Fantini, S., Walker, S. A., Franceschini, M. A., Kaschke, M., Schlag, P. M. & Feld, M. S. (2002) J. Raman Spectrosc. 33, 552–563.
Moesta, K. T. (1998) Appl. Opt. 37, 1982–1989. 24. Buschman, H. P., Deinum, G., Motz, J. T., Fitzmaurice, M., Kramer, J. R., van
5. Hebden, J., Veenstra, H., Dehghani, H., Hillman, E., Schweiger, M., Arridge,
der Laarse, A., Bruschke, A. V. & Feld, M. S. (2001) Cardiovasc. Pathol. 10,
S. & Delpy, D. (2001) Appl. Opt. 40, 3278–3287.
69–82.
6. Quaresima, V., Matcher, S. J. & Ferrari, M. (1998) Photochem. Photobiol. 67,
25. Elston, C. W., Sloane, J. P., Amendoeira, I., Apostolikas, N., Bellocq, J. P.,
4–14.
Bianchi, S., Boecker, W., Bussolati, G., Coleman, D., Connolly, C. E., et al.
7. Tromberg, B. J., Coquoz, O., Fishkin, J. B., Pham, T., Anderson, E. R., Butler,
J., Cahn, M., Gross, J. D., Venugopalan, V. & Pham, D. (1997) Philos. Trans. (2000) Eur. J. Cancer 36, 1769–1772.
R. Soc. London B 352, 661–668. 26. Tsuda, H., Akiyama, F., Kurosumi, M., Sakamoto, G. & Watanabe, T. (1999)
8. Shah, N., Cerussi, A., Eker, C., Espinoza, J., Butler, J., Fishkin, J., Hornung, Jpn. J. Clin. Oncol. 29, 413–420.
R. & Tromberg, B. (2001) Proc. Natl. Acad. Sci. USA 98, 4420–4425. 27. Rosen, P. P. & Oberman, H. A. (1992) Atlas of Tumor Pathology: Tumors of the
9. Ntziachristos, V., Yodh, A. G., Schnall, M. & Chance, B. (2000) Proc. Natl. Mammary Gland (Armed Forces Institute of Pathology, Washington, D.C.),
Acad. Sci. USA 97, 2767–2772. 3rd Series, pp. 101, 157.
10. Fornage, B. D., Sneige, N., Ross, M. I., Mirza, A. N., Kuerer, H. M., Edeiken, 28. Shafer-Peltier, K. E., Haka, A. S., Motz, J. T., Fitzmaurice, M., Dasari, R. R.
B. S., Ames, F. C., Newman, L. A., Babiera, G. V. & Singletary, S. E. (2004) & Feld, M. S. (2002). J. Cell. Biochem. 87, S125–S137.
Radiology 231, 215–224. 29. Brennan, J. F. B., Wang, Y., Dasari, R. R. & Feld, M. S. (1997) Appl. Spectrosc.
11. Gupta, P. K., Majumder, S. K. & Uppal, A. (1997) Lasers Surg. Med. 21, 51, 201–208.
417–422. 30. Motz, J. T. (2003) Ph.D. thesis (Massachusetts Institute of Technology,
12. Majumder, S. K., Gupta, P. K., Jain, B. & Uppal, A. (1998) Lasers Life Sci. 8, Cambridge, MA).
249–264. 31. Kenney, J. F. & Keeping, E. S. (1951) Mathematics of Statistics (Van Nostrand,
13. Yang, Y., Katz, A., Celmer, E. J., Zurawska-Szczepaniak, M. & Alfano, R. R. Princeton), Part 2, 2nd Ed.
(1997) Photochem. Photobiol. 66, 518–522. 32. Shama, S. (1996) Applied Multivariate Techniques (Wiley, New York).
14. Bigio, I. J., Bown, S. G., Briggs, G., Kelley, C., Lakhani, S., Pickard, D., Ripley, 33. Rosen, P. P. (1997) Breast Pathology (Lippincott–Raven, Philadelphia).
P. M, Rose, I. G. & Saunders, C. (2000) J. Biomed. Opt. 5, 221–228.
34. Houssami, N., Cheung, M. N. & Dixon, J. M. (2001) Med. J. Aust. 174,
15. Yang, Y., Katz, A., Celmer, E. J., Zurawska-Szczepaniak, M. & Alfano, R. R.
185–188.
(1996) Lasers Life Sci. 7, 115–127.
35. Cotran, R. S., Kumar, V. & Robbins, S. L. (1989) Robbins Pathologic Basis of
16. Raman, C. V. & Krishnan, K. S. (1928) Nature 121, 501–502.
17. Hanlon, E. B., Manoharan, R., Koo, T. W., Shafer, K. E., Motz, J. T., Disease (Saunders, Philadelphia), 4th Ed., pp. 1189, 1195.
Fitzmaurice, M., Kramer, J. R., Itzkan, I., Dasari, R. R. & Feld, M. S. (2000) 36. Elston, C. W. & Ellis, I. O. (1991) Histopathology 19, 403–410.
Phys. Med. Biol. 45, R1–R59. 37. Hoda, S. A. & Rosen, P. P. (2002) Am. J. Clin. Pathol. 118, 101–108.
18. Alfano, R. R., Liu, C. H., Sha, W. L., Zhu, H. R., Akins, D. L., Cleary, J., 38. Page, D. L. & Dupont, W. D. (1990) Cancer 66, 1326–1335.
Prudente, R. & Cellmer, E. (1991) Lasers Life Sci. 4, 23–28. 39. Haka, A. S., Shafer-Peltier, K. E., Fitzmaurice, M., Crowe, J., Dasari, R. R. &
19. Frank, C. J., McCreery, R. L. & Redd, D. C. B. (1995) Anal. Chem. 67, 777–783. Feld, M. S. (2002) Cancer Res. 62, 5375–5380.
20. Frank, C. J., Redd, D. C. B., Gansler, T. S. & McCreery, R. L. (1994) Anal. 40. Motz, J. T., Hunter, M., Galindo, L. H., Gardecki, J. A., Kramer, J. R., Dasari,
Chem. 66, 319–326. R. R. & Feld, M. S. (2003) Appl. Opt. 43, 542–554.

12376 兩 www.pnas.org兾cgi兾doi兾10.1073兾pnas.0501390102 Haka et al.

View publication stats