High Correlation of Double Debye Model

Parameters in Skin Cancer Detection

Bao C. Q. Truong1, H. D. Tuan1 Member, IEEE, Anthony J. Fitzgerald2, Vincent P. Wallace2 , and H.T. Nguyen1
Senior Member, IEEE

Abstract—The double Debye model can be used to capture casian population [10]. Mohs’s micrographic surgery (MMS)
the dielectric response of human skin in terahertz regime due to is currently the treatment method providing the highest cure
high water content in the tissue. The increased water proportion rate. However, this technique is intensively time-consuming
is widely considered as a biomarker of carcinogenesis, which
gives rise of using this model in skin cancer detection. Therefore, and its cure efficiency depends on experiences of pathologist
the goal of this paper is to provide a specific analysis of the in determining the histological margin of cancerous regions.
double Debye parameters in terms of non-melanoma skin cancer Therefore, accurate delineation of NMSC is essential for not
classification. Pearson correlation is applied to investigate the only saving time and effort in terms of recurrent surgeries and
sensitivity of these parameters and their combinations to the biopsy but also simplifying the treatment modality [11]. Good
variation in tumor percentage of skin samples. The most sensitive
parameters are then assessed by using the receiver operating correlation between terahertz images and their histology sec-
characteristic (ROC) plot to confirm their potential of classifying tions suggested the application of the THz imaging technique
tumor from normal skin. Our positive outcomes support further in vivo to delineate tumor margins pre-operatively [12].
steps to clinical application of terahertz imaging in skin cancer
delineation.
Preliminary studies demonstrated that the dielectric proper-
ties of human skin in terahertz frequencies can be modeled by
I. I NTRODUCTION two Debye relaxation processes, known as the double Debye
Intensive studies of terahertz (THz) radiation (T-ray) have model [13]. The capability of the double Debye model to
continuously presented the potential of terahertz pulse imaging specify the pathology of tissue was initially investigated by
(TPI) and terahertz pulse spectroscopy (TPS) in medical ap- [14]. However, its outcomes were limited because the fitting
plications, especially cancer detection [1]–[3]. The THz wave, procedure to extract the model parameters from measured
which is non-ionizing, only utilises a low power level within optical properties was recognised mathematically difficult
safety guide [4]. As THz frequencies (0.2 − 10 THz) lie in [15]. To overcome this challenge, a further study by [16]
the excitation range of torsional and vibrational motions in proposed a global optimization-based approach to extract the
molecular systems, TPI and TPS are capable of providing double Debye model parameters which provide the optimal
spectroscopic information of biological tissue [5]. The sig- fit to measured complex permittivities of human skin, both
nificantly high sensitivity of waves in the terahertz regime to healthy and cancerous. Additionally, contrast values in these
water is a well-known feature contributing to advantages of parameters between the two types of skin tissue suggested
the imaging technique as biological tissues comprise a large the possibility for skin cancer classification. Thereupon, the
proportion of water [6]. aim of the present paper is to identify the potential appli-
Contrast images between healthy and abnormal tissues such cants for skin cancer classification from the double Debye
as skin cancer and breast cancer have been previously studied. parameters via statistical analysis of pathological correlation
A THz imaging system based on reflection was especially and the receiver operating characteristic (ROC) plots of these
used to non-invasively detect cancerous regions of human skin parameters. The global optimization-based approach proposed
while ex vivo images of this skin cancer also proved the ability by [16] is employed to extract the parameters of the double
of TPI to distinguish between basal cell carcinoma (BCC) and Debye model corresponding to a variety of skin samples. To
normal tissue [2], [7]. THz imaging might be also beneficial the authors’s best knowledge, this paper is the first study in
to breast-conserving surgery as [8] investigated the possibility providing a specific statistical analysis of these parameters. Its
of using THz pulses to correctly map breast tumor margins. outcomes support further steps towards applying the double
Recent advances in THz technology triggered applications in Debye model for THz imaging in order to assist the recent
other cancers occurring at less accessible areas such as cervix tumor-removal surgeries.
and colon [9].
Non-melanoma skin cancer (NMSC), especially basal cell
carcinoma, is the most popular cancer occurring within Cau- The paper has the following structure. Section II present the
double Debye model, experiment data and applied parameter
1 Centre for Health Technologies, University of Technology Syd-
extraction procedure. Section III describe the statistical anal-
ney, Ultimo 2007, Australia; Email: [email protected], ysis of correlation and classification possibility. Discussion
[email protected], [email protected].
2 School of Physics, University of Wertern Australia, Crawley 6009, Aus- on the results are demonstrated in Section IV. Finally, we
tralia; Email: [email protected]. summarise our contributions in Section V.
978-1-4244-7929-0/14/$26.00 ©2014 IEEE 718

Authorized licensed use limited to: Zhejiang University. Downloaded on February 27,2024 at 05:54:29 UTC from IEEE Xplore. Restrictions apply.
 Table I
T HE DOUBLE D EBYE PARAMETERS

Sample BCC percentage (%) εs ε2 ε∞ τ1 (ps) τ2 (ps)

Normal 0 27.13 ± 10.00 4.64 ± 0.18 2.89 ± 0.19 4.30 ± 2.27 0.11 ± 0.02
BCC < 30 44.79 ± 32.37 4.83 ± 0.26 3.06 ± 0.13 6.56 ± 5.34 0.12 ± 0.02
BCC 30 − 50 53.91 ± 34.69 4.77 ± 0.02 2.92 ± 0.06 8.25 ± 5.67 0.12 ± 0.01
BCC > 50 73.11 ± 55.17 5.04 ± 0.15 2.97 ± 0.11 9.71 ± 7.66 0.12 ± 0.02

II. PARAMETER E XTRACTION from (2). The global optimization-based approach by [16] is
A. The double Debye model employed to find the optimal solution. According to recent
studies, τ 1 is reasonably constrained in the range 0.1 − 20
References [11], [13], [16] confirm the possibility of using
picoseconds for all samples while the constraint of τ 1 is
the double Debye theory to predict the dielectric response
[50 − 150] femtoseconds [5], [18]. The extracted parameters
of human skin in the terahertz regime. This approach is
of the double Debye model corresponding to all tumorous
based on the molecular interaction between water molecules
samples are grouped by tumor percentage. The average values
of skin tissue and THz radiation. The double Debye model
of the double Debye parameters together with their standard
incorporates two Debye relaxation processes that represent the
errors in three groups including samples with less than 30% ,
impact of an external electric field on water molecules. Under
30 − 50% and more than 50% of tumor are recorded in Table.
the excitation caused by incident T-rays, natural tetrahedral
I.
structures of water molecules are reoriented by the breakage
and reformation of multiple hydrogen bonds surrounding cen-
tral water molecules in the slow relaxation process (on a time III. S TATISTICAL A NALYSIS
scale of picosecond) [11], [17]. The fast relaxation process A. Correlation Analysis
is attributed to the reorientation of the single central water Reference [11] analysed the correlation between the ter-
molecules with a short moment (hundreds of femtoseconds). ahertz absorption coefficients and tumor content of tissue
The double Debye model used for analytically describing the samples for the purpose of evaluating the senility of TPI
relative complex permittivity of human skin is [5] for detecting tumor. The author found the highest correlation
εs − ε2 ε2 − ε∞ r = 0.75 of the absorption at 0.5 THz for each BCC
ε̃r (ω) = ε ∞ + + . (1)
1 + ωττ 1 1 + ωττ 2 samples with the percentage of tumor. Pearson correlation
Here ε s is static permittivity at low frequency, ε ∞ is the lim- between each parameter of the double Debye model and
iting permittivity at high frequency, and ε 2 is the transitional tumor percentage for all samples were also presented in
dielectric constant between two relaxation process. εs −εε2 and Table II. These parameters only provide the correlation up
ε 2 − ε ∞ respectively represent the dispersion in amplitude of to r = 0.62 that is still far lower than r = 0.75. However, it
the slow and fast relaxation processes corresponding to their is worth considering the fact that the correlations of τ 1 with
relaxation time constants τ 1 and τ 2 . two parameters ε s , ε 2 are significant (0.99, 0.65 respectively).
This suggests certain combinations of these parameters could
B. Experimental Data and Parameter Extraction improve the correlation with the BCC percentages of skin
The data used in this study were published by [11]. In samples. Therefore, several empirical non-linear relationships
particular, tissue samples were excised from mainly head and of ε s , ε 2 , ε ∞ , τ 1 were considered as follows,
neck of ten patients undergoing MMS. Each sample includes ε s − ε2
a certain proportion of BCC and adjacent normal tissue. There δ1 = , (3)
τ1
are totally 23 samples which include 13 containing tumor and εs − ε 2 − ε∞
10 without tumor. A TPI system was used for spectroscopy δ2 = , (4)
τ1
measurements in transmission mode. The time-domain signals
measured from this device are transformed into frequency- δ1BCC − ∆normal
1
∆1 = , (5)
domain spectra through Fourier transformation. Amplitude δ1normal
and phase recovered from these spectra facilitate directly
δ2BCC − ∆normal
2
measuring frequency-dependent refractive index of tissue n ∆2 = , (6)
and absorption coefficient α. Further details of this experiment δ2normal
procedure is fully described in [11]. The measured frequency- Here, δ1BCC and δ2BCC respectively stand for the values of
dependent complex permittivity is calculated through the fol- δ1 and δ2 corresponding to a BCC-contained sample while
lowing relationship: δ1normal and δ2normal are values of δ1 and δ2 for the healthy
ε̃(ω) = (n(ω) − κ(ω))2 . (2) sample of the same patient with that tumor. δ1 and δ2 are
continuously calculated for all 23 skin samples before applying
The fitting procedure is based on minimizing the total square (5)-(6) to find ∆1 and ∆2 of each BCC sample. (5)-(6)
XN
error function |ε̃r (ωi ) − ε̃(ωi )|2 over N sampled frequen- are proposed based on the fact that the hydration profile of
i=1
skin varies between patients while the double Debye model
cies in order to obtain the best fit to the permittivity data represents the dielectric response of water content inside skin

719
Authorized licensed use limited to: Zhejiang University. Downloaded on February 27,2024 at 05:54:29 UTC from IEEE Xplore. Restrictions apply.
tissue. The tumor is generally believed to be more highly ROC curve
1
hydrated than normal tissue but this context may be only
available when both the former and the latter belong to the 0.9
same patient and maybe even the same examined body part.
0.8
As a result, the subtraction in (5)-(6) facilitates removing the
effect of different skin hydration backgrounds between patients 0.7
on the changes due to the difference in tumor content. The
0.6
correlation values of δ1 , δ2 , ∆1 , ∆2 with BCC percentages are

Sensitivity
provided in Table II. 0.5 εs
ε2
Table II 0.4
ε∞
T HE P EARSON CORRELATION VALUES OF MODEL PARAMETERS
0.3 τ1

εs ε2 ε∞ τ1 (ps) τ2 (ps) δ1 δ2 ∆1 ∆2 τ2
0.2
0.44 0.63 -0.1 0.34 0.2 0.74 0.79 0.92 0.85 δ1

0.1 δ2

0
B. Classification Analysis 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
1−Specificity
The applied Pearson correlation in III-A, which only ex-
amine the linear dependence between two variables, is readily
Fig. 1. ROC curves of ε s , ε2 , ε ∞ , τ 1 , τ 2 , δ1 , δ2 and the absorption
not a complete measure for identifying potential indicators coefficient α at 0.5 THz.
in classification. Thus, the receiver operating characteristic
(ROC) plot is commonly employed to explore the classification
accuracy in medical community [19]. The ROC plots also IV. D ISCUSSION
demonstrate the trade-offs between the sensitivity and speci-
ficity. A threshold value chosen by averaging two consecutive As can be seen in Table I, the impact of increased tumor
instances of original data is used to split the data into two sepa- proportion inside skin samples on the variation of the double
rate sets which are compared with pathology to determine the Debye parameters generally results in the increases of these
correspondent sensitivity and specificity. Fig. 1 demonstrate values. In particular, average values of ε s consistently increase
the ROC curves of ε s , ε 2 , ε ∞ , τ 1 , τ 2 , δ1 , δ2 and the absorption due to the increase of BCC percentage while the similar trend
coefficient α at 0.5 THz respectively. Furthermore, the area is also observed in the remaining cases of ε 2 , ε ∞ and τ 1 . The
under the ROC curve (AUC) of a classification feature is an high variance of ε s , τ 1 , which are well-correlated, could be
independent index from any particular threshold value and due to overfitting. Constraining the variation range of τ 1 will
commonly used to assess the general performance of this be helpful to reduce the variance of these two parameters. The
feature. An AUC closer to 1 indicates a better classification differences in ε 2 , ε ∞ are far smaller as compared to those in
ability. Specific AUC values of all ROC curves in Fig. 1 can ε s , τ 1 and τ2 . In fact, high frequency parameters like ε 2 , ε ∞
be seen in Table III. According to Fig. 1, it is possible to are more dependent on refractive index and low frequency
parameters including ε s , τ 1 and τ2 are more sensitive to
Table III absorption coefficient [5]. Thus, the more prevalent difference
T HE AREA UNDER ROC CURVES (AUC).
between BCC and normal skin in absorption coefficient than
εs ε2 ε∞ τ1 τ2 δ1 δ2 α refractive index in THz regime can be reflected using the
0.89 0.69 0.46 0.727 0.46 0.95 0.93 0.91 double Debye model.
Despite that Pearson correlations of the five double Debye
determine the ROC point which provides the highest correct parameters with tumor percentage are quite low, the combina-
classification accuracy (CCR) with respect to each parameter. tions of these parameters even demonstrate far higher corre-
Table IV only demonstrates specific threshold values responsi- lations than the previous study by [11]. Table II particularly
ble for the best CCRs of ε s , ε 2 , δ1 , δ2 and α together with their demonstrates the higher sensitivity of δ1 and δ2 to BCC per-
sensitivity and specificity due to the prominent AUC values of centage than that of absorption coefficient at 0.5 THz, which
these parameters. indicates prominent potential of these combinations for skin
cancer delineation. The very high standard errors of the Debye
Table IV
S TATISTICAL MEASURES OF THE BEST CLASSIFICATION PERFORMANCE parameters as can be seen in Table I limits the sensitivity
BASED ON ROC ANALYSIS . of each individual parameter to BCC content of samples.
However, the dependencies of the double Debye parameters on
Parameter Threshold Sensitivity Specificity CCR
tumor content may be non-linear whilst the Pearson correlation
εs 28.23 84.62 90 86.96
ε2 4.71 69.23 70 69.57 test is only based on a linear regression analysis. Thus, there
δ1 5.80 92.31 90 91.30 may require further studies to clarify this assumption. On
δ2 5.24 84.62 90 86.96 the other hand, the aforementioned correlations between the
α 119.88 84.62 90 86.96 parameters are beneficial to the combinations of the double
Debye parameters. The exceptional correlations provided by

720
Authorized licensed use limited to: Zhejiang University. Downloaded on February 27,2024 at 05:54:29 UTC from IEEE Xplore. Restrictions apply.
∆1 (r = 0.92) and ∆2 (r = 0.85) suggest that the different R EFERENCES
nature of water profile between patients’ skin tissues does [1] S. Nakajima, H. Hoshina, M. Yamashita, C. Otani, and N. Miyoshi,
considerably affect the sensitivity of the double Debye model “Terahertz imaging diagnostics of cancer tissues with a chemometrics
to skin, especially tumorous tissue. Furthermore, the approach technique,” Applied Physics Letters, vol. 90, no. 4, p. 041102, 2007.
[2] R. M. Woodward, B. E. Cole, V. P. Wallace, R. J. Pye, D. D. Arnone,
proposed in (5)-(6) is worth considering for further applica- E. H. Linfield, and M. Pepper, “Terahertz pulse imaging in reflection
tions in order to separate the cancer-related contrast from the geometry of human skin cancer and skin tissue,” Physics in Medicine
natural differences between patients’ bodies. and Biology, vol. 47, no. 21, pp. 3853–3863, 2002.
[3] V. P. Wallace, P. F. Taday, A. J. Fitzgerald, R. M. Woodward, J. Cluff,
The potential Debye-related parameters for skin cancer clas- R. J. Pye, and D. D. Arnone, “Terahertz pulsed imaging and spec-
sification have been found through intensive analysis presented troscopy for biomedical and pharmaceutical applications,” Faraday
Discussions, vol. 126, pp. 255–263, 2004.
in section III-B. To be more specific, ε s , δ1 , δ2 possess the [4] E. Berry, G. C. Walker, A. J. Fitzgerald, N. N. Zinovev, M. Chamberlain,
highest AUC indices, which represent themselves as the most S. W. Smye, R. E. Miles, and M. A. Smith, “Do in vivo terahertz imaging
prominent explanatory variables for classification. The value systems comply with safety guidelines?,” Journal of Laser Applications,
vol. 15, no. 3, 2003.
of AUC corresponding to the absorption coefficient at 0.5 [5] E. Pickwell, A. J. Fitzgerald, B. E. Cole, P. F. Taday, R. J. Pye,
THz is also recorded in Table III for comparison with those T. Ha, M. Pepper, and V. P. Wallace, “Simulating the response of
of ε s , δ1 , δ2 , which highlights the supremacy of these Debye terahertz radiation to basal cell carcinoma using ex vivo spectroscopy
measurements,” Journal of Biomedical Optics, vol. 10, no. 6, p. 064021,
parameters in terms of classification potential. This superiority 2005.
is coherent due to the fact that the double Debye model math- [6] Z. D. Taylor, R. S. Singh, D. B. Bennett, P. Tewari, C. P. Kealey,
ematically incorporates the differences of both the frequency- N. Bajwa, M. O. Culjat, A. Stojadinovic, H. Lee, J. P. Hubschman, E. R.
Brown, and W. S. Grundfest, “THz medical imaging: in vivo hydration
dependent refractive index and absorption coefficient. In re- sensing,” IEEE Transactions on Terahertz Science and Technology,
ality, these optical properties also can be combined under vol. 1, no. 1, pp. 201–219, 2011.
the form of time-domain impulse functions or frequency- [7] R. M. Woodward, V. P. Wallace, R. J. Pye, B. E. Cole, D. D. Arnone,
E. H. Linfield, and M. Pepper, “Terahertz pulse imaging of ex vivo basal
domain power spectra. The contrast in these factors between cell carcinoma.,” The Journal of Investigative Dermatology, vol. 120,
cancerous and normal tissue may also include either unknown no. 1, pp. 72–78, 2003.
contributors besides water absorption or misleading features [8] A. J. Fitzgerald, V. P. Wallace, M. Jimenez-Linan, L. Bobrow, R. J. Pye,
A. D. Purushotham, and D. D. Arnone, “Terahertz pulsed imaging of
[7]. Deeper analysis of ROC curves in Fig. 1 results in the best human breast tumors,” Radiology, vol. 239, no. 2, pp. 533–540, 2006.
threshold values in Table IV for detecting the BCC samples. [9] K. Wang and D. M. Mittleman, “Metal wires of terahertz wave guilding,”
The best CCR with the sensitivity 92.31% and specificity 90% Nature, vol. 430, pp. 376–379, 2012.
[10] M. Mogensen and G. B. E. Jemec, “Diagnosis of nonmelanoma skin
corresponding to δ1 , which has the highest AUC, is obtained cancer/keratinocyte carcinoma: A review of diagnostic accuracy of non-
using the threshold value δ1 = 5.80. δ2 is also a potential melanoma skin cancer diagnostic tests and technologies,” Dermatologic
feature for classification since it highly correlates with the Surgery, vol. 33, no. 10, pp. 1158–1174, 2007.
[11] V. P. Wallace, A. J. Fitzgerald, E. Pickwell, R. J. Pye, P. F. Taday,
tumor percentage of skin sample (r = 0.79) as well as achieves N. Flanagan, and T. Ha, “Terahertz pulsed spectroscopy of human basal
impressive sensitivity (84.62%) and specificity (90%). In spite cell carcinoma.,” Applied Spectroscopy, vol. 60, no. 10, pp. 1127–1133,
of only having very low Pearson correlation with BCC content, 2006.
[12] V. Wallace, A. Fitzgerald, S. Shankar, N. Flanagan, R. Pye, J. Cluff, and
ε s still expresses itself as a key classification feature with its D. Arnone, “Terahertz pulsed imaging of basal cell carcinoma ex vivo
statistical measures of performance similar to those of δ2 . and in vivo,” British Journal of Dermatology, vol. 151, no. 2, pp. 424–
At the bottom line, as the notable classification outcomes 432, 2004.
[13] E. Pickwell, B. E. Cole, A. J. Fitzgerald, M. Pepper, and V. P. Wallace,
are simply based on a very basic method using threshold “In vivo study of human skin using pulsed terahertz radiation,” Physics
values, applying more complex classifiers such as support in Medicine and Biology, vol. 49, no. 9, pp. 1595–1607, 2004.
vector machine promise significant improvement of accuracy. [14] C. B. Reid, E. Pickwell-MacPherson, J. G. Laufer, A. P. Gibson, J. C.
Hebden, and V. P. Wallace, “Accuracy and resolution of thz reflection
spectroscopy for medical imaging,” Physics in Medicine and Biology,
vol. 55, no. 16, p. 4825, 2010.
V. C ONCLUSION [15] B. C. Q. Truong, H. D. Tuan, H. H. Kha, and H. T. Nguyen, “System
identification for terahertz wave’s propagation and reflection in human
skin,” in Communications and Electronics (ICCE), 2012 Fourth Inter-
We have provided a comprehensive analysis of the double national Conference on, pp. 364–368, Aug 2012.
Debye model in terms of its sensitivity to tumor content of [16] B. C. Q. Truong, H. D. Tuan, H. H. Kha, and H. T. Nguyen, “Debye
human skin and the potential of using its parameters to classify parameter extraction for characterizing interaction of terahertz radiation
with human skin tissue,” IEEE Transactions on Biomedical Engineering,
tumors from normal skin tissue. While there is no significant vol. 60, pp. 1528–1537, June 2013.
correlation of each Debye parameter with the BCC percent- [17] C. Ronne and S. R. Keiding, “Low frequency spectroscopy of liquid
age of skin sample, the proposed combination coefficients water using thz-time domain spectroscopy,” Journal of Molecular Liq-
uids, vol. 101, no. 13, pp. 199 – 218, 2002. Molecular Liquids. Water
δ1 , δ2 , ∆1 , ∆2 are significantly dependent on the tumor content at the New Millenium.
with the corresponding Pearson correlation up to r = 92%. [18] J. T. Kindt and C. A. Schmuttenmaer, “Far-infrared dielectric properties
In addition, both δ1 and δ2 , our proposed combination of of polar liquids probed by femtosecond terahertz pulse spectroscopy,”
The Journal of Physical Chemistry, vol. 100, no. 24, pp. 10373–10379,
the double Debye parameters, acquire promising values of 1996.
AUC, sensitivity and specificity which confirm the striking [19] T. Fawcett, “Roc graphs: Notes and practical considerations for re-
potential of using these parameters to detect skin tumor. searchers,” ReCALL, vol. 31, no. HPL-2003-4, pp. 1–38, 2004.
Future studies will consider applying advanced classification
algorithms together with intensive tests on THz images of skin
cancer.

721
Authorized licensed use limited to: Zhejiang University. Downloaded on February 27,2024 at 05:54:29 UTC from IEEE Xplore. Restrictions apply.