1. The document describes a study using spectroscopy to detect cervical dysplasia. Non-negative matrix factorization (NNMF) was used to decompose spectroscopy data into constituent source spectra and concentrations. 2. A machine learning model (Lasso regression) combined NNMF source concentrations to predict dysplasia levels. This improved prediction performance over individual reflectance or fluorescence data. 3. Two-dimensional disease maps were created locating cervical dysplasia tissue using the machine learning results. These maps correctly identified biopsy-confirmed normal and dysplastic tissue locations.

1. Identifying Constituent Spectra Sources in Multispectral Images To Quantify and Locate Cervical Neoplasia
Kevin C. Baker and Shabbir Bambot
Guided Therapeutics, Inc. 5835 Peachtree Corners East, Suite D Norcross, GA 30092
kbaker@guidedinc.com
5. Lasso regression machine learning modeling is used to identify the combination of different Figure 5. Lasso probability distribution functions and receiver operating characteristic curves.
INTRODUCTION source concentrations that best predict the amount of cervical dysplasia. (a) shows the PDF curves of the normal, CIN 1, CIN 2, and CIN 3 groups indicating higher values best
 Current methods to detect and manage cervical pre-cancers often miss the disease and generate  The Lasso is a shrinkage and selection method for linear regression that minimizes the sum of squared predict dysplasia tissue. (b) shows classification ROC curves to distinguish CIN 1 and CIN 2+ from normal
false positives. errors with a bound on the sum of the absolute values of the coefficients. tissue.
 This can lead to a delay in correctly diagnosing and over-treating cervical dysplasia.  The quantitative prediction labels used in the Lasso model are 1 for normal points, 2 for CIN 1, 3 for CIN
2, and 4 for CIN 3. Lasso linear regression combines NNMF source concentration information to improve cervical
 Recent clinical trials with two year follow up have shown that colposcopy (i.e. current triage standard of
care) can be inaccurate in determining the need for biopsies and locating appropriate biopsy sites.
 The point dataset is randomly subdivided into 60% training and 40% testing data to test for over fitting. dysplasia predication and classification performance.
 A receiver operating characteristic curve (ROC) analysis is done to determine the dysplasia tissue  There are 18 different features selected after performing the Lasso regression procedure on only the
classification performance. training data.
 Guided Therapeutics has developed a spectroscopic device to reduce the number of false positives  Testing dataset had a fitting error that is less than 0.1% different than the training dataset.
while retaining the high true positive predictive rate. 6. Two dimensional (2-D) disease maps are created to spatially locate and quantify cervical  ROC area under the curve (AUC) for CIN 1 and CIN 2+ classification is 0.75 and 0.86
Both reflectance and fluorescence spectroscopy are dysplasia tissue using the Lasso regression results obtained from the NNMF concentrations.  Detection specificity at 95% sensitivity (spec@95) for CIN 1 and CIN 2+ tissue is 30% and 58%.
Figure 1. Spectroscopy measurement device 
Sensitivity at 70% best spec model(sens@70) for CIN 1 and CIN 2+ tissue is 70% and 87%. analysis
utilized to detect morphological and biochemical  Each subject's 2-D false color disease map is created from the Lasso regression results and the 
Lasso specificity PDF Lasso best spec model ROC
abnormalities associated with cervical pre-cancer and corresponding measurement location. 3 1
cancer.  Finally, the 2-D disease maps are compared with cervical colposcopy images and biopsy results. Norm
CIN 1 0.9
 In a seven-center pivotal study, the device demonstrated
2.5 CIN 2
its potential as a cost effective and efficient modality for RESULTS CIN 3
0.8
the early detection of moderate and severe cervical
Sensitivity (True postive rate)
CIN 2+
0.7
intraepithelial neoplasia (CIN) disease in women at risk for Figure 3. Comparison between derived and actual source reflectance and fluorescence spectra. 2
cervical disease, while at the same time potentially (a) shows the reflection NNMF source signal that best predicts the dysplasia by itself. (b) shows the best 0.6
fluorescence source signal that predicts cervical dysplasia by itself. (c) shows a hemoglobin (Hb) and
Probability
reducing the number of colposcopies and biopsies from
normal and benign cervices. oxy-hemoglobin (HbO2) absorption spectrum that was measured by Scott Prahl at the Oregon Medical Laser 1.5 0.5
Center. (d) shows an actual porphyrin fluorophore spectra signal (i.e. protoporphyrin) at the same excitation
0.4
wavelength as (b) that was measured by DaCosta at the University of Toronto.
1
0.3
NNMF sources appear to characterize physiological cervical tissue spectra.
 Traditional supervised and unsupervised methods of analyzing fluorescence and reflectance spectra  The behavior of the reflectance spectra is opposite of the absorption spectrum, and you can see the 0.5
0.2
data have difficulty recovering physiological information. characteristic HbO2 absorption peaks as valleys in the reflectance NNMF source signal. CIN 1 auc: 0.75 spec@95: 0.3 sens@70: 0.7
0.1
 Supervised un-mixing methods require knowledge about the reflectance/fluorescence spectral patterns of  Porphyrin fluorescence spectrum has two peaks near 640nm and 700nm, and these two peaks appear to be CIN 2+ auc: 0.86 spec@95: 0.58 sens@70: 0.87
the constituent materials. also present in the fluorescence NNMF source signal. 0 0
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
(a) (b)
Relectance source spectra model: 4 source: 1 Fluorescence @460nm source spectra model: 4 source: 3 Predication values (a.u.) 1-Specificity (False postive rate)
 Reflectance and fluorescence spectral data are non-negative by nature, and spectral sources don’t always 100 60
appear to be statistically independent (e.g. collagen and elastin fluorescence).
90
 However, principal component analysis assumes that the underlying sources are uncorrelated and
50
Gaussian, and the resulting sources/concentrations can be negative. 80 Figure 6. Comparison between colposcopy image and biopsy with 2-D spectroscopy disease maps
 Also independent component analysis assumes that the sources are statistically independent, and the obtained from the Lasso regression results.
70
sources/concentrations can be negative. 40 Biopsy spatial locations are shown in the colposcopy images (left images) by an "x" indicator. (a)
60 Colposcopy image with biopsy diagnosed as normal. (b) 2-D spectroscopy disease map showing entire
Intensity (a.u.)
Intensity (a.u.)
cervix as normal. (c) Colposcopy image with biopsy diagnosed as CIN 3 and normal. (d) 2-D spectroscopy
 Therefore, non-negative matrix factorization (NNMF) is chosen to analyze the spectral data, since 50 30
disease map indicating where cervical dysplasia tissue exists.
it only assumes that the source and mixture data are non-negative. 40
20 The corresponding 2-D disease maps Figure 6 (b) and (d) correctly identify the biopsied normal and
30
dysplasia cervical tissue, and could have correctly been used to determine where to take a biopsy.
METHODS 20
10
1. Clinical colposcopist expert and biopsy results are used to create a point-by-point diagnosis. 10
 The diagnosis categories include: normal (e.g. normal squamous, normal columnar), CIN 1 (Grade 1), CIN
0 0
2 (Grade 2), CIN 3 (Grade 3), or CIN 2+ (either CIN 2 or CIN3). (a) 400 450 500 550 600 650 700 (b)500 520 540 560 580 600 620 640 660 680 700
Wavelength (nm) Wavelength (nm)
2. Measurement points and subjects are excluded if they were taken over non-cervical tissue,
5 4
x 10 x 10 Protoporphyrin fluorescence spectra @ 460nm excitation Normal
6 16
had excessive blood or mucus, had a specular reflection artifact (CCD saturation), or had an HbO2
unknown/uncertain disease classification. Hb 14
5
3. Spectral measurements are calibrated to account for any instrument and subject differences.
Molar extinction coefficient (cm- 1)
12
Fluorescence intensity (a.u.)
 Calibration sources are used to calibrate the system’s spectra response and perform wavelength calibration. 4
10
 Subject differences are calibrated by dividing the spectra with the mean intensity over all wavelengths.
(a) (b)
3 8
4. A NNMF approach is used to blindly decompose reflectance and fluorescence spectra into a
set of constituent source spectra curves and concentrations. 6
 NNMF approximates a non-negative matrix with the product of two other non-negative matrices: X≈A*S 2
where X is the measured spectral data matrix, S is an unknown spectral source matrix, and A is an 4
Normal
unknown mixing matrix. 1
 Matrices S and A are chosen to minimize the root-mean-squared residual (RMSR) between X and A*S. 2
 NNMF algorithm is an iterative approach and does not reach a unique solution.
0 0
(c) 400 450 500 550 600 650 700 (d)500 520 540 560 580 600 620 640 660 680 700
Wavelength (nm) Wavelength (nm)
Figure 2. Non-negative matrix factorization approach to construct constituent source spectra. CIN 3
(a) shows a fluorescence source signal obtained using a single NNMF source model. (b) shows the
calculated NNMF source signals obtained using a 3 source model. (c) shows how the RMSR decreases until (c) (d)
it becomes stable as the number of factors (sources) included in the model increases.
Figure 4. Probability distribution function (PDF) estimates for normal and dysplastic tissue. CONCLUSIONS
With a 3 source model, the source spectral signals appear to represent physiological fluorophores  The PDF curves represent the following different cervical tissue: normal (blue), CIN 1 (green), CIN 2
 Calculated sources from the NNMF approach appear to be represent physiological sources (e.g.
often present in cervical source spectra model: 1 source: 1 NADH, FAD/collagen,@340nm all points source curves for factor #: 3
Fluorescence @340nm epithelium tissue (e.g. Fluorescence porphyrin). (red), and CIN 3 (cyan). (a) shows the PDF estimates using the reflectance source concentrations that
correspond to the Hb source signal shown in Figure 3 (a). (b) shows the PDF curves using concentrations hemoglobin and porphyrin) and their concentration changes are in agreement with cervical
350 250
obtained from the fluorescence NNMF signal that appears to be a porphyrin spectrum shown in Figure 3 dysplasia changes.
data1 (b).
300 data2
 A machine learning algorithm (i.e. Lasso linear regression) improved the dysplasia prediction
200 data3 NNMF source concentrations behave similar as organic compound concentrations in the cervix. performance by combining source concentrations from both the reflectance and fluorescence spectra
 Figure 4 (a) results agree with the widely accepted knowledge that as dysplasia increases, the amount of
data.
250 angiogenesis increases and thus the HbO2 absorption (reflectance) increases (decreases).
 Higher porphyrin concentrations are well known to be present in pre-cancererous tissue.
150  Other prevalent cervical tissue fluorophore concentrations (not shown here) behave as expected as the level  2-D false color spectroscopy disease maps demonstrate the ability to quantify and spatially locate
Intensity (a.u.)
inensity (a.u.)
200 of dysplasia increases: NADH increases, FAD decreases, and collagen decreases. model: 4 source: 3
Relectance PDFs model: 4 source: 1 Fluorescence @460nm PDFs dysplasia cervical tissue.
600 180
Norm
150 CIN 1
160  These results offer the potential to reduce the number of false positive cases while maintaining a
100 500 high enough detection rate necessary for primary screening, and may assist in identifying biopsy
CIN 2
140
CIN 3 locations.
100 400 120
Probability (a.u.)
Probability (a.u.)
50 Norm
100
CIN 1
REFERENCES
50 300
CIN 2 Jeronimo, J., Shiffman, M., “Colposcopy at a crossroads,” Am. J. Obstet. Gynecol. 195, 349-353 (2006).
80
CIN 3
Twiggs, L., B., Chakhtoura, N., A., Werner, C., L., Griffith, W., F., Flowers, L, C., Lashgari, M., Ferris, D., G., Winter, M., L., Sternfeld,
0 0 200 60 D., R., Burnett, A., F., Wilkinson, E., J., and Raab, S., S., “Multimodal Spectroscopy As A Triage Test For Women At Risk For Cervical
400 450 500 550 600 650 700 400 450 500 550 600 650 700
(b) Neoplasia: Results Of A 1,607 Subject Pivotal Trial,” J. Lower Gen. Tract. Dis. 14(3), 264-269 (2010).
(a) Wavelength Fluorescence @340nm all points
(nm) wavelength (nm) 40
NNMF factor # residual: min
0.14 100 Chang, K., S., Mirabal, N., Y., Atkinson, N., E., Cox D., Malpica, A., Follen, M., and Richards-Kortum, R., “Combined reflectance and
20 fluorescence spectroscopy for in vivo detection of cervical pre-cancer,” J. Biomed. Eng. 10(2), 024031 (2005).
0.12
0 0 Hastie, T., Tibshirani, R., and Friedman J., [The Elements of Statistical Learning: Data Mining, Inference, and Predication], Springer, New
4 5 6 7 8 9 0.005 0.01 0.015 0.02 0.025 York NY, (2009).
Root mean residual error
(a) Source concentration (a.u.) -3 (b) Source concentration (a.u.)
x 10
Root mean residual error
0.1 Berry, M., W., Browne, M., Langville, A., N., Pauca, V., P., and Plemmons, R., J., “Algorithms and application for approximate
nonnegative matrix factorization,” Computation Statistics and Data Analysis, 52(1), 155-173 (2007).
0.08
DaCosta, R., S., Andersson, H., and Wilson, B., C., “Molecular fluorescence excitation-emission matrices relevant to tissue spectroscopy,”
Photochem Photobiol. 78(4), 384-392 (2003).
0.06
ACKNOWLEDGEMENTS http://omlc.ogi.edu/spectra/hemoglobin/
The project described was supported in part by a grant award from the National Cancer Institute. The content is
Chang, S., K., Marin, N., Follen, M., and Richards-Kortum, R., “Model based analysis of clinical fluorescence spectroscopy for in vivo
0.04 solely the responsibility of the authors and does not necessarily represent the official views of the National detection of cervical intraepithelial dysplasia,” J. Biomed. Eng. 11(2), 024008 (2006).
Cancer Institute or the National Institutes of Health.
0.02 Wu, Y., Xi, P., and Qu, J., Y., “Depth-resolved fluorescence spectroscopy of normal and dysplastic cervical tissue,” Optics Express 13(2),
1 1.5 2 2.5 3 3.5 4 4.5 5
(c) ©2011 Guided Therapeutics, Inc. 382-388 (2005).
Factor #