Hyperspectral Imaging Assessment for Radiotherapy
Induced Skin-Erythema: Pilot study
Skin cancer (SC) is a widely spread type of cancer in USA, Canada, and Australia. Patients, in the senior age, of skin
cancer is usually referred to radiation therapy, rather than surgery due to old age complications, for treatment.
However, radiation therapy induces side effects that may vary between tissue necrosis down to skin erythema. As
erythema is the primary evidence of skin health perturbation due to radiation exposure, it needs to be precisely
assessed. Visual assessment (VA) is the gold standard for erythema evaluation. Nevertheless, VA is anything but
ideal, due to being dependent on experience and varying human sensations. Hyperspectral imaging (HSI), far from
human dependency, is an optical technique that provides an opportunity for objective investigation of erythema. HSI
spectral data permitted the computation of erythema indices using Dawson’s technique. Dawson relative erythema
index proved to be highly correlated (97.1%) to clinically visual score in monitoring skin erythema. In addition, on
the 7th session of radiation therapy, relative erythema index differentiates with 99% significance between irradiated
and non-radiated skin regions. In this study, HSI is compared to digital photography for skin erythema statistical
Skin cancer (SC) is classified into non-melanoma skin
cancer (NMSC) and melanoma skin cancer (MSC).
The latter is the life-threatening one, but fortunately it
accounts for only 2% of SC patients . SC can be
treated via several approaches , two of them are
more common: surgery and local radiotherapy . In
surgery, facial skin cancer has destitute healing
chances particularly in the curved surfaces of human
body such as cheeks and noses. In addition, senior SC
patients are frequently excluded from the surgical
approach due to age-related complications. Thereby,
disqualified patients for surgery are referred to local
radiotherapy treatment , . Radiotherapy
treatment employs ionizing radiation (X-rays photons/
electrons)  to produce a series of fatal DNA
destructions to control lesions growth. Frequently, the
prescribed dose of radiation is divided into a number
of smaller doses, called “fractions” , . These
fractions have upper and lower limits for daily dose
(Gy/day) need to be taken into consideration in order
to achieve maximum benefit with minimum toxicity
Toxicity of the healthy skin is the bad side of radiation
therapy. Ionizing radiation is the trigger of substantial
tragic consequences that take place for skin and
inherent subcutaneous layers. Particularly, some
studies – have articulated that at least 85-95%
of radiotherapy treated patients countered detrimental
side effects. These side effects vary in intensity from
mild to severe ones. The variation depends on several
human-related parameters , , –. The
parameters which are highly correlated with the skin’s
adverse reaction were studied formerly , –
, –, ,. Age, gender, radiation type,
and the treated body part of the patient were found to
be the most dominating parameters, while the dose
level was less dominant , , . The surveys
done on the irradiated patients stated that skin toxicity
might rise up via the following reactions: , 
1- The ephemeral skin redness (skin erythema) which
needs no longer than couple of hours to show up as
the earliest skin response, or it can linger visibly
unseen up to 15 fractions .
2- The soreness of the skin that takes place due to the
fluctuating death level of epidermal basal cells. It
typically occurs after 15 fractions. The consequence
of skin soreness is either a dry or a wet
3- A long-term skin response (dermal ischemia and
necrosis) may come about 2- 4 months after
4-Late skin damage, such as the development of
telangiectasia and severe necrosis might ensue a
In optical scope, erythema was picked out to be our
interesting symptom to study skin temporal reaction to
irradiation. Skin erythema is believed to be a
secondary inflammatory response. The inflammation
changes the skin color to be red. The skin redness is
the result of the damage to the basal cell population of
the epidermal layer. The former damage yields a
vasodilatory reaction which changes the blood content
within the dermis layer. The light within the visible-
NIR spectra penetrates the human skin with 2 mm
depth  and thus, imaging in VNIR region will be
probing the epidermal layer to capture the erythema.
Great endeavor were exerted in adopting several
techniques for objectively evaluating radiation-
induced erythema , –. Despite the
diversity of the former techniques, visual assessment
is still considered the gold standard for skin erythema
evaluation . However, visual assessment is far
from being flawless since it is subjective and strongly
depends on accumulated experience and human
factors including vision acuity. Two out of all the
utilized techniques, Diffuse reflectance spectroscopy
(DRS) – and digital imaging –, were
intimately involved in skin erythema assessment as
objective measures. Yet, both techniques have
inadequate satisfaction of clinicians. Clinicians
criticize DRS because of small inspected region and
need for contact. For digital imaging, clinicians are
dissatisfied with the limited spectral details and the
dependency on the photographer skills. Hence, a
technique that overcomes the former weaknesses
would be interesting for clinical work.
Hyperspectral imaging (HSI) is an optical technique
emerged rapidly and easily found its way to distinct
fields of research including the medical one .
Basically, HSI captures multiple frames, for a well-
defned region, on sequential spectral bands. Thus, a
radiance spectrum could be collected for each pixel
recorded in the stack of frames. Thereby, HSI rises
above the limitations of DRS and digital imaging and
overcomes the subjectivity and experience-based
disadvantages of the gold standard but at the expense
of simplicity and time consumption.
This study is investigating the potential of HSI for
objectively quantify radiotherapy induced skin
erythema. The milestones of this study include: (1)
analyze the visual assessment scores for the recruited
patients, (2) compute the skin diffuse reflectance along
the days of treatment, (3) estimate the skin’s
chromophores concentration using least square fitting
4) Apply statistical classification upon HSI and
corresponding digital color images for skin erythema
and compare the results.
2 Materials and Methods
2.1 Clinical protocol
The entire clinical work was performed likewise our
research ethics board committee protocol for human
experiments. A written consent was given by all
participating subjects along with a schedule of both
study and treatment sessions. The patients were
instructed not to apply any topical agent or skin
dressing except upon recommendation of the
oncologist. In case of any skin agent or dressing there
are requested to notify the study’s main investigator.
Clinical work, patient recruitments, and experimental
measurements commenced went on for six months.
Eight skin cancer patients were consented to
participate in the study, however, only five patients
completed the entire study successfully. The
remaining three patients’ data was ignored due to the
study’s exclusion criteria. The demographic data of
the successful five patients is displayed in Table 1. The
average age of the involved patients was
approximately 75 years, with 2/3 female: male ratio,
respectively. All the recruited patients were diagnosed
with NMSC; either basal cell carcinoma (BCC) or
squamous cell carcinoma (SCC) and received at least
10 radiation fractions during treatment. The treated
location, for the recruited patients, varied between the
arms, legs and face regions. The prescribed radiation
type was either electrons or x-rays.
Table 1: Demographic data of the recruited skin cancer patients during experiments
The clinical session, for this study, was scheduled
routinely right before the irradiation session for each
patient in order to perform the study daily
hyperspectral imaging and skin assessment. Day-to-
day, one out of two radiotherapists scored, and
documented the patient’s skin relevant symptoms/
features. A special chair-bed was used to seat the
patients in order to fix the pose of the imaging
procedure every day. A typical platform for the
imaging instrumentation was designed and built to
maintain the relative patient-instrumentation
displacement. For each patient the treated region of
interest (ROI) circumference was marked with an
indestructible marker to ensure reproducibility. A
transparent film was used as a backup template for the
ROI marks considering permanent landmark features
like moles and freckles. The purpose of this template
is to keep tracking of marks whenever it is washed out
after weekends. Some information were jotted down
on the template such as direction symbols at the
template edges, which helps later in registering
consequent images . An example of the template
for one patient’s ROI is shown in Figure 1.
Figure 1: A skin cancer patient’s leg ROI is marked for
daily reference during irradiation treatment. A
transparent film is created with a copy of the patient
marks to be used to as a backup in case of real marks
washed out to verify clinical procedure reproducibility.
2.2 Visual assessment
On a daily basis, a simple questionnaire, designed by
the radiotherapist, was given to the patients to respond
verbally. The therapist is interested in documenting
the patients’ previous day activity along with any
possible skin reaction raised up from the same day
irradiation session. Imperative information to be
recorded was the case of applying topical agents or
dressing to the patient’s ROI within the three hours
preceding the imaging session. All the patients’
answers and erythema scores were securely recorded
in an established patient’s paper-based and digital
record. Skin-erythema level were planned to be graded
on a 5-steps scale to express the described skin status
as shown in Table 2Error! Reference source not
found.. The scale is designed based on a consultation
between the radiotherapists in Juravinski Cancer
Center. Each level is assigned to a percentage score.
Frequently, the patient ROI is not developing single
erythema level, but inhomogeneous distinct erythema
Table 2: Clinical assessment of skin erythema 5-steps
2.3 Digital Color Imaging
A digital camera was used to capture images for the
recruited patients’ ROI. The captured frame/ FOV was
not limited to the marked ROI of the patients but it also
incorporated a non-irradiated skin region as well. The
non-irradiated skin region within the captured frames
enabled tracking of the normal skin status on daily
basis. Digital imaging was repeated for three times for
each session to reduce human-based errors.
The digital color images were captured by a high
definition digital video camera (HYUNDAI-HDMI-
768) which has a 16 Mpixels sensor along with a 3”
LCD display to adjust the FOV. The camera is
equipped with an objective lens F/3.2, f=7.5 mm. The
camera dimensions are 110 x 58 x 50 mm, and weighs
240 g. The illumination and the perspective view were
considered to be the same along the entire study. A
montage of a patient’s images is shown in Figure 2.
Figure 2: Digital color images montage captured for a
patient diagnosed with skin cancer on her left tibia. The
images are arranged from left to right and from up to
down on time basis. The patient had originally a lot of
red skin features in her leg before receiving any
irradiation. However, she started to develop a visible
faint skin erythema at the 8th session of irradiation which
had been evolved to dry desquamation in the 4th week
after receiving 15 sessions of irradiation.
No visible erythema
Skin has a very light pink color
Skin reaction is more apparent
with clear borders but is still
pink with more intensity.
Erythema is apparent in bright
pink and borders are clearly
Skin is bright red, borders are
very well defined, capillaries
and bruising may be visible.
Color images were used to visibly assess the skin
status. In addition, it facilitated the cross-validation of
the clinical erythema scores with other, non-attending
2.4 Hyperspectral imaging
A custom-made HSI system, developed and
characterized elsewhere , was used to acquire the
patients’ datacubes. For this purpose, the patient’s
FOV is illuminated with a halogen light source
emitting in the visible-near infrared (VNIR) spectra.
The light was shined on the patients from different
directions to reduce the shadow effect.
The patient’s ROI back reflected light is collected at
the focal plane of a zoom lens. An optical relay is used
to convey the captured FOV through the optical
configuration using an achromatic lens pair. The
collimated beam output of the primary relay lens is
transmitted through a polarizing beam splitter (PBS),
which permits the linearly vertical polarized light
component to go through while reflects the
horizontally polarized component. The horizontally
polarized component is folded toward a half-wave
plate for a 900 degrees polarization rotation. By then,
both PBS light components are directed onto an
acousto-optic tunable filter (AOTF). Tweaking the
horizontal component of the polarized input ray
contributes in raising the system throughput since the
involved AOTF cell is optimized for linearly vertical
polarization. AOTF crystal derived by an RF
synthesizer diffracts tunable light bandwidth angularly
separated from the original beam. The second optical
relay lens is aligned to receive nothing but the
separated diffracted light band and focus it on the
The HSI imaging detector (Ximea MQ042rg-CM
Enhanced-IR) incorporated in the system is a
monochromatic CMOS sensor-based camera. It has an
active sensor area of 11.24 mm2 with 2048 x 2048
pixels and a pixel pitch of 5.5 µm. The camera
exposure time extends from ~25 µsec up to 1 second
and supports up to 90 frames per second. A skin region
of approximately 11 x 11 cm2 was considered to
involve both irradiated and non-irradiated regions,
resulting in a spatial resolution of 0.05 mm/ pixel.
A developed code by C-language was used to
synchronize the AOTF and camera operation
simultaneously through a user-friendly interface. The
single hyperspectral data set involves three datacubes:
the first for the patient ROI, the second is for a
standard white diffuse reflectance surface (SRS-99-
010, Labsphere, North Sutton, New Hampshire), and
the third for the camera dark current. The white
standard surface datacube is collected right after the
patient imaging to maintain the same position and
illumination conditions. The camera dark current data
is collected while the zoom lens cap is on and all lights
are off. Each datacube includes eighty-nine narrow
band images bounded by the spectral range of 450:850
nm. The integration time for capturing one band image
is one second, and thus a single datacube takes 89
seconds. Spectral data sets, likewise digital color
imaging, were repeated for 3 times to minimize human
errors. A simulation for the study’s imaging session is
shown on a volunteer in Figure 3.
Figure 3: A volunteer simulating patient is seated in
order to be prepared for hyperspectral imaging. The
treatment field is supposed to be in the right leg. The
treatment field is illuminated from two sides to avoid the
shadow effect. Sunglasses were worn to protect the
subject eyes during imaging time.
3 Data Analysis
For each individual patient, the skin erythema clinical
score day-to-day were tabulated in his/ her file. Beside
the clinical score, both the captured digital color
images and the spectral datacubes, for the
corresponding day, were filed and labeled by date.
3.1 Visual Assessment Score
The erythema score (S) is computed as the summation
of the visual erythema grade (G), assessed by the
radiotherapist, multiplied by the corresponding
occupied percentage area within the treatment field, as
shown in equation 1. To give an example, A score of
1.0 is equivalent to a bright erythema present on the
entire treatment field.
3.2 Spectral Reflectance Computation
The patients’ daily datacube is preprocessed for
analysis in two consecutive steps. The first step is to
subtract the dark datacube pixel intensity () from
both the patient’s raw datacube () along with the
reference white standard pixel intensity (). The
dark datacube represent the existing background. The
second step is to divide the patient data by the
corresponding background-free white reference data.
The former preparation steps are mathematically
displayed in equation (2), and applied using ImageJ
The resultant output, of equation (2) is the spectral-
resolved diffuse reflectance datacube. Since each data
cube is composed of 89 band-images, image
registration is a must step for processing. For image
registration purposes, an intermediate band (central
wavelength λ=600 nm) was selected as the reference
band. The former band is selected since it lies on the
highest point of the spectral performance for our
developed hyperspectral imaging system. The
reference band is used for image registration to
eliminate the patient’s movements using ImageJ.
Selected band-images out of the spectral-resolved
diffuse reflectance datacube for a patient in the third
week (13th session) of treatment is displayed in Figure
4 to exhibit the development of erythemaError!
Reference source not found..
Figure 4: Sample of the datacube band images at selected
wavelengths (560, 580, 650, 700, and 840 nm), along with
a digital color image for the same patient ROI. All the
images, spectral and color, were taken at the same time
point (13th session) of irradiation. The 560, 580 nm bands
are highlighting the hemoglobin absorption, while the
650, 700 nm are representing the melanin correction
bands, and the 840 nm NIR band is showing the low
absorption of the majority of skin chromophores
(hemoglobin, melanin, and water).
In erythema, the skin content of the hemoglobin
chromophore is majorly affected. This particular
chromophore, hemoglobin, is distinguished with two
spectral absorption peaks at 560 and 580 nm bands.
Thus, the highly concentrated hemoglobin regions
appear darker , . The penetration depth at the
green spectra, including the 560 and 580 nm bands, is
limited to fractions of mm range , which restricts
the visibility to superficial capillaries. However,
deeper skin information could be achieved at longer
wavelengths since the skin visible chromophores have
less absorption preference and thus, light can travel
further down to a couple of millimeters. For instance,
the longer wavelength bands show a pale shadow, due
to erythema, in the region of increased hemoglobin
content as seen at 650 nm, and 700 nm bands. At the
band of 840 nm central wavelength, all the
chromophores display absorption minima.
The reason for erythema is not only because of the
more scattered red light out of the skin, but also due to
the back reflected green light was intensely absorbed.
Hence, wherever the more hemoglobin is incremented
in a skin location, the less green light is scattered back
as shown in Figure 5. As the melanin concentration of
the epidermal layer is undefined, the estimated
hemoglobin components concentrations, via detecting
the back diffuse reflected light, are considered to be
apparent ones for a single layered tissue model. The
apparent concentrations are indicative but not
equivalent to the true concentrations found, for
example, in x-vivo dermal skin tissue. However, it
could still be a quantitative index for erythema percent
variation if a hemoglobin baseline concentration is
Figure 5: Patient diffuse reflectance spectra for the
treatment field of skin on treatment days (reddish dots),
and at the end of treatment (black dots). A reference
diffuse reflectance spectrum for non-irradiated skin
region is plotted (blue dots). The amount (percent) of
reflected light is temporally decreasing in the green
region in a gradual style rather than the corresponding
arbitrary behavior in the red one.
3.3 Erythema Index Computation
The erythema index, proposed by Dawson  and
cited more than 420, is a common computational
method, for probing skin chromophores change in
concentration out of measured diffuse reflectance. The
method is based on computing the area under the curve
500 550 600 650 700 750 800 850
(AUC) of the logarithm of the skin diffuse reflectance
reciprocal (LRR) in the visible region (510-610 nm).
This spectral region comprises two of the hemoglobin
distinguished absorption peaks. Dawson’s erythema
index (DEI) is computed according to equation (3):
Hereby, p, q, r, s, and t are symbols expressing the
LRR values of the measured data at the five
wavelengths 510, 540, 560, 580, and 610 nm,
respectively. According to the HSI instrument
specifications the former wavelengths were not
exactly reachable, so we replaced the p, q, r, s, and t
LRR exact values by the corresponding LRR average
value for the adjacent spectral range
Melanin, beside hemoglobin, chromophore is an
influential component in the visible diffuse reflectance
for skin. For that reason, melanin effect needs to be
compensated in erythema index computation. For
melanin correction factor, reflectance data at two
wavelengths (650&700 nm) is used to compute
Dawson melanin index (DMI) . The former
wavelengths were selected because melanin spectral
absorbance is proportional to its concentration at these
two spectral points, as well as the low absorbance of
the hemoglobin in this spectral region . DMI is
computed with the same averaging range
for the used wavelengths applying equation (4).
The corrected erythema index (DEIc) is computed
using DMI by applying the formula shown in equation
(5), where γ = 0.04 is an empirically derived balancing
constant to avoid negative values as a result of other
non-significant skin chromophores.
In this study, recruited patients were seen on daily
basis to monitor the development of erythema due to
radiation sessions. For this reason, day-to-day
erythema assessment might be needed for clinical
purposes. Thus, relative Dawson erythema index
() is calculated. is computed by subtracting
for one day from an earlier one. The relative
index is a more indicative value toward skin color
changes since it avoids the effect of repeated systemic
3.4 Estimation of Skin Chromophores
Diffuse reflectance data acquired by spectroscopic
based instrumentation need to be interpreted using a
model to account for the propagation of the light inside
biological tissues. In our study, the model can be
approximated by a perpendicular light source
illuminating a semi-infinite, homogeneous slab.
Quantitatively estimating the concentrations of the
main chromophores can be accomplished with a
conditional least-square fitting of the skin-absorbers’
extinction coefficients. The logarithmic reciprocal of
the measured diffuse reflectance data is
optimally fitted by a sum of the dominant skin
chromophores’ extinction coefficients in the
applicable spectral region, multiplied by their
concentrations as shown in equation (6) :
Three chromophores are dominant in the visible
spectrum (500-840 nm; the oxygenated fraction of
hemoglobin (HbO2), the deoxygenated fraction of
hemoglobin (Hb), and melanin. A minor effect might
be attributed to other skin chromophores such as water
and fats. This effect can be compensated by adding a
constant as in equation (7):
Where the terms:
are the concentrations and extinction
coefficients of HbO2, Hb, and melanin, respectively.
3.5 Image Classification
Classification, in HSI, is the process of allocating
distinct data classes within a hyperspectral datacube.
For allocating the separate classes, a step-wise
procedure is followed. The first step is data
normalization to a reference (the white standard in our
case). The second step is to reduce the noise in the
collected data via a smoothing technique. Various
approaches are available to reduce the background
noise in an image including software enhancement or
involving high end optical filters. The later approach
is an expensive one. However, software enhancement
using for example Wiener filter is considered efficient
noise reduction approach . Wiener filter
statistically estimates the meaningful information out
of the original noisy data with a lossless amount of real
information. It uses the minimum mean square error
method to exclude less useful information . It is
selected as it improves the relative signal to noise
counts through applying the convenient parametric or
nonparametric techniques . To make it simple,
Wiener filter can be considered as a low-pass imaging
filter that is able to be tweaked to the input image’s
local variance and can be mathematically expressed as
in equation (8) . The terms: is the filtered
pixel corresponding to the noisy input pixel ,
expresses the noise variance, and and are the
image locally computed mean and standard deviation
Truthfully, hyperspectral bands are not equally
informative. One main reason for certain bands to be
less informative, is systemic reduced performance.
The system low performance might be due to non-
uniform illumination spectra or lower detector’s
quantum efficiency or the former reasons combined.
Hence, band/ variable selection (VS) is a
computational technique assists in excluding the least
informative bands in the captured datacubes. VS is not
arbitrary but it is executed given a threshold of 99%
information within datacube is retained. VS is
different from dimensionality reduction techniques, as
the reduction techniques provide its output with no
capability of extracting the significant bands out of the
original ones. In this study, matrix low-rank
representation (MLRR)  is our VS criterion.
MLRR technique is based on sorting the variables in
terms of the band self-contained information. To put it
simply, MLRR is clustering data pixels drawn from
larger multiple subspaces to delineate outliers.
Computing the bands’ rank, within a datacube, is
accomplished using the Frobenius (element-wise)
matrix norm, as follows
MLRR clustering starts with calculating the
summation for each squared matrix, band-image,
columns and arrange the columns in a descending
order. Then, specifying a threshold percentage value
P = 0.99 (significance interval), a comparison between
each band, , and the average norm
of all the
bands will highlight outlier bands among the rest:
In summary, VS determines the least informative
bands preserving 99% of the contained information in
the datacube. By then, the data analysis computation
time, for HSI, is reduced while preserving maximum
contained information. To give an example, Figure 6
displays two rows of images which belongs to a single
datacube. The images in the top row have a very low
signal to noise ratio and provide less than 1% of the
datacubes’ information. Elimination of these bands
will have a negligible effect on the imperative spectral
information. In this case, the excluded bands (485 nm,
487 nm, 491 nm) are in a less interesting spectral
region in erythema context, physiologically, has no
distinguishable features relevant to dominant skin
chromophores. On the other side, a sample of the
highest informative, spectral bands is shown in the
bottom row of Figure 6.
Figure 6: A display for the least informative bands (485
nm, 487 nm, 491 nm), (top-row), and the highest
informative bands (696 nm, 728 nm, 736 nm), (bottom-
row) in a single datacube captured for one patient. Both
groups of bands were determined by the matrix low-rank
For erythema classification purposes, skin annotations
for the distinctly developed skin erythema are
required. Annotations were generated using Image-J
software, by an expert radiotherapist based on
attending the study’ daily imaging session. The
annotations serve as ground truth data for erythema
assessment. The ground truth is used to train and
cross-validate the developed classifier. In fact, skin
erythema is a non-homogenous symptom and thus its
annotations do not form mutually exclusive contours
which could be a complication for classification
purposes. To come over this complication, we applied
a logic X-OR operation to separate distinct erythema
levels to end up with exclusive contours as shown in
Figure 7. For reference, non-erythematous skin was
added to express normal skin.
Following the annotations process, the data becomes
ready for classification. Among various techniques,
we picked up linear discriminant analysis (LDA) 
method for data classification in this study. In
principle, LDA technique is established on developing
a linear transformation matrix H that reduces the
spectral k-dimensional raw data of vector F to an s-
dimensional vector provided that s < k .
Figure 7: A sample of annotations created by the
radiotherapist to mark two distinct erythema regions as
ground truth data. The annotations are overlaid on both
spectral data (left) and digital color image (right). The
eye shield appeared in the images was put on the patient
face for the elimination of light inconvenience during
The spectral dimensional reduction is placed on the
concept of achieving the maximum possible
separation among data classes dissemination . To
make it clear, LDA rule of thumb is to maximize the
displacement in between the central mean point (M) of
the classes, meanwhile preserving interclass variance
(V) to be minimum . The challenge is to find the
axis that projects the data cloud into separate clusters
with means and small
variance . For instance, in the case of two
classes’ mode the term , in equation (12),
represents the plane of best separation, should be
maximal to achieve superlative separation between
each class and the other.
4 Results and Discussion
4.1 Visual assessment
The total visual erythema assessment score is
determined along the prescribed radiotherapy
treatment for all patients. Figure 8 displays the
computed visual assessment average score for the
recruited patients who successfully completed the
entire study’s sessions. The scores were presented
against the treatment completion percentage due to
different prescribed number of sessions among
patients. Error bars represent the standard deviation of
the visual assessment scores among patients.
Accumulation of radiation doses, for patients, along
the plan of treatment guided the erythema score to
increase proportionally with time. The standard
deviation is relatively large since human factors are
distinct between patients including age, cancer type,
Figure 8: Visual assessment average score for the
recruited patients, who successfully completed the study,
during the plan of treatment. The scores are presented
against the treatment completion percentage . Error bars
represent the standard deviation among patients. A
horizontal dotted line is drawn for referencing purpose.
4.2 Skin Reflectance and Erythema
Once the skin spectral reflectance for patients is
computed, it was thought-provoking to monitor the
behavior for skin chromophores indices. Regarding
melanin, the computed index displayed a fluctuating
behavior from one session to the next. Figure 9 shows
the average values for melanin index for the recruited
patients while the error bars represent the standard
deviation among the patients. A dotted black
horizontal line is intersecting with the initial level of
apparent melanin prior to radiation treatment. The
purpose of this line is to monitor the melanin index
value with respect to the original level. The oscillation
in melanin index value all along the treatment proves
that there is no consistent relation between the skin
melanin content and accumulation of radiation.
Figure 9: the study successful Patients’ average melanin
apparent presence behavior for the entire radiotherapy
treatment time are scattered on the data points while the
010 20 30 40 50 60 70 80 90 100
Therapist Skin Erythema Score
010 20 30 40 50 60 70 80 90 100
error bars are expressing the variability recorded among
In addition to melanin, corrected and relative erythema
indices were calculated for the involved patients by
applying Dawson’s  formulas. The computed
melanin corrected erythema index is exhibited in
Figure 10 while the corresponding relative index is
displayed in Figure 11. A healthy individual
volunteered to be imaged at the same time on a daily
basis for 10 consecutive days maintaining the same
conditions of imaging likewise the patients. The
volunteer data was captured according to the study
protocol. Corrected and relative erythema indices
were computed and presented in Figure 10 and Figure
11 for contrast purpose. Error bars in the volunteer
case are expressing the standard deviation of
Figure 10: displays, in red points, the computed average
for melanin corrected erythema index. the error bars
express the variability in computed erythema among
patients. The red dotted horizontal line is acting as a
baseline for all the subsequent measurements. The blue
points and dotted horizontal line is displaying the
volunteer melanin corrected erythema index and
Figure 11: displays, in red points, the computed average
for relative erythema index. The error bars express the
variability in computed relative erythema among
patients. The red dotted horizontal line is acting as a
baseline for all the subsequent measurement. The blue
points and dotted horizontal line display the volunteer
relative erythema index and baseline, respectively.
The corrected erythema index showed no constant
tendency during the early treatment days, but it
increases persistently, with a variable rate, in the
second half of the treatment itinerary. We expect that
the oscillating behavior for corrected erythema index
is due to the fluctuation in the rate of basal cells
destruction in the epidermal layer of the skin .
Consequently, the inflammatory response of the skin
is not consistent. This inconsistent inflammation might
be due to the stops in treatment because of weekends,
or statutory holidays within the treatment time period.
These stops contributed in large error bars. On the
volunteer side, the melanin corrected erythema index
showed an intermittent alteration with respect to the
initial base and small error bars. The volunteer
alteration reflected the daily changes with no clear
The relative erythema index is comparing the skin
reflectance change for a radiated body part with
respect to a non-radiated one. For patients, relative
erythema index displayed a consistent increase against
the days of treatment. To confirm the former result,
the relative erythema was computed to a healthy
volunteer for two distinct regions in his arm. The
healthy volunteer data displayed, in Figure 11, similar
values in all the days of imaging except for the
baseline measurement. For verification purposes, the
former results out of the computed erythema indices
need to be statistically significant.
For significance check, Wilcoxon nonparametric two
tail paired test was used to examine the hypothesis of
a meaningful difference between the skin’s diffuse
reflectance at subsequent time points of radiation
treatment session against a prior one. Wilcoxon test
was utilized in this study, seeing that it goes well with
a small number of observations. Basically, Wilcoxon
test is analyzing the difference in the computed
median for two groups of population before and after
exposure to a certain experiment . Relative
erythema indices, corrected erythema indices, and
melanin indices were elected since they are dependent
parameters on the skin diffuse reflectance at nine time-
points along the radiation treatment plan. Each data
point was examined with respect to the initial data
point. The examination score is compared to the
tabulated Wilcoxon statistical value for the
corresponding number of observations. The tabulated
Wilcoxon value is plotted as a black horizontal line in
Corrected erythema index (arbitrary units)
0 1 2 3 4 5 6 7 8 9 10
Relative Erythema index (arbitrary units)
Figure 12 (a). This line is the marginal threshold,
above which the null hypothesis is rejected. The
significance of Wilcoxon results is illustrated in Figure
Wilcoxon test showed that relative erythema,
corrected erythema, and melanin indices reached and
crossed the null rejection threshold at a certain time
point through the prescribed radiation treatment period
for patients. Markedly, the melanin index has a single
time point in the mid-way of treatment period which
crossed the null rejection threshold. As a result,
Wilcoxon statistical examination confirmed the
inconsistent change in the apparent melanin
concentration within the skin due to receiving
Figure 12: Wilcoxon (a) statistical test results against null
hypothesis rejection limit (no change), (b) corresponding
significance percentage for relative erythema, corrected
erythema, and melanin indices, along the radiotherapy
The relative erythema index was the early parameter
in terms of time to cross the null hypothesis rejection
limit with 90% significance at the third time point.
Furthermore, it reached a significance ratio of 99% at
the seventh time point earlier than the corrected
erythema index. This result highlighted the potential
of the relative erythema index in delineating the skin
response to cancer radiation treatment in a consistent
way. It also gives an evidence of the insignificant of
melanin chromophore concentration variability during
the radiation therapy treatment. So as to approve the
benefit of computing erythema indices, it needs to be
correlated with visual assessmentError! Reference
source not found.. For that purpose, Pearson
correlation was computed. The correlation between
the visual assessment scores and the corrected
erythema index was found to be 79.1%, and for the
relative erythema index was 97.1%. The former
correlation results emphasized the potency of relative
erythema index in objectively quantifying the skin’s
reaction toward exposure to radiation.
4.3 Estimation of skin Chromophores
The main chromophore responsible for the skin’s
redness/ erythema is the hemoglobin with its both
parts, oxygenated and non-oxygenated. Probing
hemoglobin apparent concentration is an analytical
step in developing deep understanding of the
physiological phenomena accounting for the induced
skin inflammation due to radiation exposure. For that
purpose, least square fitting for the computed
reflectance data was accomplished using the
extinction function for each of both hemoglobin
components. This is the way used in this study in order
to quantitatively estimate hemoglobin apparent
concentration during treatment.
0 1 2 3 4 5 6 7 8 9 10
Radiation Time Points
Wilcoxon Test R esults
Rel. Er. Ind.
Cor. Er. Ind.
0 1 2 3 4 5 6 7 8 9 10
Treatment Time Points
Rel. Er. Ind.
Cor. Er. Ind.
0 2 4 6 8 10 12 14 16 18 20
Oxyhemoglobi n (arbitrary units)
Figure 13: Mean Hb-oxy (a), Hb-deoxy (b), and total
hemoglobin (c), as a function of time during irradiation.
Error bars are displayed representing standard
deviation. Baseline values are drawn in each plot in
dashed lines, respectively.
The apparent concentration of hemoglobin parts for
involved patients, in this study, were averaged and
scattered along the days of treatment as illustrated in
Figure 13 (a, b). Error bars are representing the
variability between the patients. Oxyhemoglobin
apparent concentration shows a clear trend of upturn
throughout the days of irradiation dose exposure.
Deoxyhemoglobin, oppositely, was interchangeably
altering between upturns and downturns. Adding both
hemoglobin components was the way to obtain the
total hemoglobin apparent concentration displayed in
Figure 13 (c). The estimated concentration of the total
hemoglobin showed a sporadic trend of ups and downs
in the treated body part versus the original
concentration before irradiation with a general
incremental trend. This result is expected due to the
unlike temporal apparent concentration behavior acted
by both hemoglobin components.
4.4 Image classification
Applying the developed LDA classifier on both the
digital color photographs and the spectral datacubes
was the way to contrast both techniques. The contrast
was held through a supervised classification for color
and spectral data relevant to the ground truth
annotations provided by the radiotherapist. Distinct
ratios between the training sets and the test sets were
examined. Based on examination, it is found that 1:1
ratio between training and test samples were superior
to different other options such as the Pareto principle
for 20:80 percentage, as well as other 60:40 %
percentage. The 1:1 ratio produced satisfactory results
for most of the assessment parameters. The
hyperspectral system shows a comparable
performance with digital photography in parameters
such as accuracy and sensitivity while it outperforms
RGB imaging in specificity and the geometrical mean.
We report the classification results for six parameters,
for ten repeated times of randomly selecting training
and test sets out of the population data. The results of
classification are displayed in
Table 3: Classification results for comparison between HSI and RGB imaging regarding skin erythema assessment.
91.08 ± 0.83
86.65 ± 0.89
88.68 ± 1.01
32.89 ± 3.39
47.23 ± 2.63
86.2 ±0. 94
90.10 ± 0.94
16.85 ± 4.63
We expect that this pilot study provides a good
approach toward objective assessment for skin
induced erythema due to radiation therapy. In
particular, HSI permitted contactless visualization and
quantitative assessment of relatively wide skin region
0 2 4 6 8 10 12 14 16 18 20
Deoxyhemogl obin (arbitrary units)
0 2 4 6 8 10 12 14 16 18 20
Total Hemoglobin (arbitrary units)
in terms of oxygenation and perfusion in radiotherapy
which saves time in contrast with a skin limited region
investigation by diffuse reflectance spectroscopy. We
conclude the results of this study as follows:
1- Relative erythema index value, at the 7th radiation
session, varies with 99% significance compared
with the non-radiated skin. Relative erythema
index is verified to be the most sensitive
parameter to prob the skin response due to
irradiation, as it achieved a 97.1% correlation
ratio relative to visual assessment of erythema.
2- Oxyhemoglobin chromophore concentration
showed a gradual increase along with radiation
accumulation in the areas adjacent to the lesion
This increase of hemoglobin concentration is
probably due to the skin healing procedure after
damage. Oppositely, deoxyhemoglobin showed
an oscillating behavior upon the radiation period.
3- The supervised erythema classification based on
HSI data tremendously outperformed RGB
imaging in terms of specificity and geometric
We are convinced that future progress in producing a
more even illumination as well as a reduction of the
instrumentation noise level will pave the way for a
fast, reliable, and, above all, objective quantitative
method for erythema classification. In contrast, HSI
may serve as a good candidate to the fast recognition
of individuals attacked by harmful radiation. Further
experimental, clinical, and analytical work need to be
done to make HSI a real-time clinical instrument.
 A. C. Society, “Cancer Facts and Figures
 M. Veness and S. Richards, “Role of modern
radiotherapy in treating skin cancer.,”
Australas. J. Dermatol., vol. 44, no. 3, pp.
159-166; quiz 167-168, 2003.
 N. P. Nguyen et al., “Effectiveness of
radiotherapy for elderly patients with non-
melanoma skin cancer of the head,” Geriatr.
Gerontol. Int., vol. 15, no. 5, pp. 601–605,
 A. Sanabria, A. L. Carvalho, J. G. Vartanian,
J. Magrin, M. K. Ikeda, and L. P. Kowalski,
“Comorbidity is a prognostic factor in elderly
patients with head and neck cancer.,” Ann.
Surg. Oncol., vol. 14, no. 4, pp. 1449–1457,
 D. M. Brizel, R. K. Dodge, R. W. Clough, and
M. W. Dewhirst, “Oxygenation of head and
neck cancer: Changes during radiotherapy and
impact on treatment outcome,” Radiother.
Oncol., vol. 53, pp. 113–117, 1999.
 M. Caccialanza, R. Piccinno, L. Kolesnikova,
and L. Gnecchi, “Radiotherapy of skin
carcinomas of the pinna: a study of 115 lesions
in 108 patients.,” Int. J. Dermatol., vol. 44, no.
6, pp. 513–7, 2005.
 W. Maddocks-Jennings, J. M. Wilkinson, and
D. Shillington, “Novel approaches to
radiotherapy-induced skin reactions: A
literature review,” Complement. Ther. Clin.
Pract., vol. 11, no. 4, pp. 224–231, 2005.
 D. Porock and L. Kristjanson, “Skin reactions
during radiotherapy for breast cancer: The use
and impact of topical agents and dressings,”
Eur. J. Cancer Care (Engl)., vol. 8, no. 3, pp.
 S. K. MacBride, M. E. Wells, C. Hornsby, L.
Sharp, K. Finnila, and L. Downie, “A case
study to evaluate a new soft silicone dressing,
Mepilex Lite, for patients with radiation skin
reactions.,” Cancer Nurs., vol. 31, no. 1, pp.
 S. D’haese, M. Van Roy, T. Bate, P.
Bijdekerke, and V. Vinh-Hung, “Management
of skin reactions during radiotherapy in
Flanders (Belgium): a study of nursing
practice before and after the introduction of a
skin care protocol.,” Eur. J. Oncol. Nurs., vol.
14, no. 5, pp. 367–372, 2010.
 J. W. Denham et al., “Factors influencing the
degree of erythematous skin reactions in
humans,” Radiother. Oncol., vol. 36, no. 2, pp.
 M. McQuestion, “Evidence-Based Skin Care
Management in Radiation Therapy,” Semin.
Oncol. Nurs., vol. 22, no. 3, pp. 163–173,
 S. Balter, J. W. Hopewell, D. L. Miller, L. K.
Wagner, and M. J. Zelefsky,
“Fluoroscopically guided interventional
procedures: a review of radiation effects on
patients’ skin and hair.,” Radiology, vol. 254,
no. 2, pp. 326–341, 2010.
 M. S. Chin et al., “Hyperspectral imaging for
early detection of oxygenation and perfusion
changes in irradiated skin.,” J. Biomed. Opt.,
vol. 17, no. 2, p. 026010, Feb. 2012.
 N. Lee et al., “Skin toxicity due to intensity-
modulated radiotherapy for head-and-neck
carcinoma,” Int. J. Radiat. Oncol. Biol. Phys.,
vol. 53, no. 3, pp. 630–637, 2002.
 T. Olschewski, K. Bajor, B. Lang, E. Lang,
and M. H. Seegenschmiedt, “[Radiotherapy of
basal cell carcinoma of the face and head:
Importance of low dose per fraction on long-
term outcome].,” J. Dtsch. Dermatol. Ges.,
vol. 4, no. 2, pp. 124–30, 2006.
 R. I. J . W . HOPEWELL (CRC Normal Tissue
Radiobiology Research Group and H.
(University of Oxford), The Churchill
Hospital, “The skin : its structure and response
to ionizing radiation,” INT. J. RADIAT. BIOL
., vol. 57, no. 4, pp. 751–773, 1990.
 L. Diana L. Glennie, Joseph E. Hayward,
Orest Z. Ostapiak, James Wright and T. J. F.
Doerwald-Munoz, “Diffuse reflectance
spectroscopy for monitoring erythema in head
& neck intensity modulated radiation
therapy,” J. Radiat. Oncol., 2014.
 K. J. Zuzak, M. T. Gladwin, R. O. C. Iii, I. W.
Levin, J. Karel, and R. O. Can-, “Imaging
hemoglobin oxygen saturation in sickle cell
disease patients using noninvasive visible
reflectance hyperspectral techniques : effects
of nitric oxide,” vol. 0510, pp. 1183–1189,
 S. Wan, J. a Parrish, and K. F. Jaenicke,
“Quantitative evaluation of ultraviolet induced
erythema.,” Photochem. Photobiol., vol. 37,
no. 6, pp. 643–648, 1983.
 B. L. Diffey, R. J. Oliver, and P. M. Farr, “A
portable instrument for quantifying erythema
induced by ultraviolet radiation,” Br. J.
Dermatol., vol. 111, no. 6, pp. 663–672, Dec.
 C. Balas, “An imaging colorimeter for
noncontact tissue color mapping.,” IEEE
Trans. Biomed. Eng., vol. 44, no. 6, pp. 468–
 J. Nyström, P. Geladi, B. Lindholm-Sethson,
J. Rattfelt, A. C. Svensk, and L. Franzen,
“Objective measurements of radiotherapy-
induced erythema,” Ski. Res. Technol., vol. 10,
no. 4, pp. 242–250, 2004.
 C. K. Kraemer, D. B. Menegon, and T. F.
Cestari, “Determination of the minimal
phototoxic dose and colorimetry in psoralen
plus ultraviolet A radiation therapy,”
Photodermatol. Photoimmunol. Photomed.,
vol. 21, no. 5, pp. 242–248, 2005.
 G. N. Stamatas, B. Z. Zmudzka, N. Kollias,
and J. Z. Beer, “In vivo measurement of skin
erythema and pigmentation: new means of
implementation of diffuse reflectance
spectroscopy with a commercial instrument.,”
Br. J. Dermatol., vol. 159, no. 3, pp. 683–90,
 S. Taylor, W. Westerhof, S. Im, and J. Lim,
“Noninvasive techniques for the evaluation of
skin color,” J. Am. Acad. Dermatol., vol. 54,
no. 5, pp. S282–S290, 2006.
 K. Vishwanath, D. Klein, K. Chang, T.
Schroeder, M. W. Dewhirst, and N.
Ramanujam, “Quantitative optical
spectroscopy can identify long-term local
tumor control in irradiated murine head and
neck xenografts.,” J. Biomed. Opt., vol. 14, no.
5, p. 054051, 2009.
 D. L. Glennie, J. E. Hayward, D. E. Mckee,
and T. J. Farrell, “Inexpensive diffuse
reflectance spectroscopy system for
measuring changes in tissue optical
properties,” J. Biomed. Opt., vol. 19, no.
December, pp. 105005-1-105005–6, 2014.
 D. Yohan et al., “Quantitative monitoring of
radiation induced skin toxicities in nude mice
using optical biomarkers measured from
diffuse optical reflectance spectroscopy.,”
Biomed. Opt. Express, vol. 5, no. 5, pp. 1309–
 M. Setaro and A. Sparavigna, “Quantification
of erythema using digital camera and
computer-based colour image analysis: a
multicentre study.,” Skin Res. Technol., vol. 8,
pp. 84–88, 2002.
 S. G. Coelho, S. A. Miller, B. Z. Zmudzka, and
J. Z. Beer, “Quantification of UV-Induced
Erythema and Pigmentation Using Computer-
Assisted Digital Image Evaluation,”
Photochem. Photobiol., vol. 82, no. 3, pp.
 E. Rizova and A. Kligman, “New
photographic techniques for clinical
evaluation of acne,” J. Eur. Acad.
Dermatology Venereol., vol. 15, no. SUPPL.
3, pp. 13–18, 2001.
 G. Lu and B. Fei, “Medical hyperspectral
imaging: a review.,” J. Biomed. Opt., vol. 19,
no. 1, p. 10901, Jan. 2014.
 R. Abdlaty, J. Orepoulos, P. Sinclair, R.
Berman, and Q. Fang, “High Throughput
AOTF Hyperspectral Imager for Randomly
Polarized Light,” Photonics, vol. 5, no. 1, p. 3,
 C. a Schneider, W. S. Rasband, and K. W.
Eliceiri, “NIH Image to ImageJ: 25 years of
image analysis,” Nat. Methods, vol. 9, no. 7,
pp. 671–675, 2012.
 G. N. Stamatas and N. Kollias, “In vivo
documentation of cutaneous inflammation
using spectral imaging.,” J. Biomed. Opt., vol.
12, no. 5, p. 051603, 2007.
 T. Lister, P. a Wright, and P. H. Chappell,
“Optical properties of human skin.,” J.
Biomed. Opt., vol. 17, no. 9, pp. 90901–1, Sep.
 F. H. Mustafa and M. S. Jaafar, “Comparison
of wavelength-dependent penetration depths
of lasers in different types of skin in
photodynamic therapy,” Indian J. Phys., vol.
87, no. 3, pp. 203–209, Nov. 2012.
 J. B. Dawson et al., “A theoretical and
experimental study of light absorption and
scattering by in vivo skin.,” Phys. Med. Biol.,
vol. 25, no. 4, pp. 695–709, 1980.
 D. L. Glennie, J. E. Hayward, D. E. Mckee,
and T. J. Farrell, “Inexpensive diffuse
reflectance spectroscopy system for
measuring changes in tissue optical
properties,” vol. 19, no. December, pp. 1–6,
 D. L. Glennie, J. E. Hayward, and T. J. Farrell,
“Modelling changes in skin blood
concentration with total diffuse reflectance
spectroscopy,” pp. 1–31.
 S. Tan, R. Young, D. Budgett, J. Richardson,
and C. Chatwin, “A pattern recognition
Weiner filter for realistic clutter
backgrounds,” Opt. Commun., vol. 172, no. 1–
6, pp. 193–202, 1999.
 M. Isaac N. BANKMAN (Applied Physics
Laboratory Johns Hopkins University Laurel,
Handbook of Medical Imaging. Academic
 T. Wang, Ö. Özdamar, J. Bohórquez, Q. Shen,
and M. Cheour, “Wiener filter deconvolution
of overlapping evoked potentials,” J.
Neurosci. Methods, vol. 158, no. 2, pp. 260–
 Fundametals of digital image processing.
Prentice Hall, 1989.
 G. Liu, Z. Lin, S. Yan, J. Sun, Y. Yu, and Y.
Ma, “Robust Recovery of Subspace Structures
by Low-Rank Representation,” Pattern Anal.
Mach. Intell. IEEE Trans., vol. 35, no. 1, pp.
 M. Li and B. Yuan, “2D-LDA: A statistical
linear discriminant analysis for image matrix,”
Pattern Recognit. Lett., vol. 26, no. 5, pp. 527–
 T. Bandos, “Classification of hyperspectral
images with regularized linear discriminant
analysis,” Geosci. Remote …, vol. 47, no. 3,
pp. 862–873, 2009.
 U. of T. Welling, Max (Department of
Computer Science, “Fisher Linear
Discriminant Analysis,” 2005.
 M. H. Shilts et al., “Minimally Invasive
Sampling Method Identifies Differences in
Taxonomic Richness of Nasal Microbiomes in
Young Infants Associated with Mode of
Delivery - Online First - Springer,” Microb.
Ecol., pp. 233–242, 2016.