A Novel Use of Hyperspectral Images for Human Brain Cancer
Detection using in-Vivo Samples
Himar Fabelo
1
, Samuel Ortega
1
, Raúl Guerra
1
, Gustavo Callicó
1
, Adam Szolna
2
, Juan F. Piñeiro
2
,
Miguel Tejedor
1
, Sebastián López
1
and Roberto Sarmiento
1
1
Institute for Applied Microelectronics, University of Las Palmas de Gran Canaria, Las Palmas de Gran Canaria, Spain
2
Department of Neurosurgery, University Hospital Doctor Negrín, Las Palmas de Gran Canaria, Spain
Keywords: Brain Cancer Detection, Hyperspectral Imaging, Random Forest, Classification Algorithms.
Abstract: Hyperspectral Imaging is an emerging technology for medical diagnosis issues due to the fact that it is a non-
contact, non-ionizing and non-invasive sensing technique. The work presented in this paper tries to establish
a novel way in the use of hyperspectral images to help neurosurgeons to accurately determine the tumour
boundaries in the process of brain tumour resection, avoiding excessive extraction of healthy tissue and the
accidental leaving of un-resected small tumour tissues. So as to do that, a hyperspectral database of in-vivo
human brain samples has been created and a procedure to label the pixels diagnosed by the pathologists has
been described. A total of 24646 samples from normal and tumour tissues from 13 different patients have
been obtained. A pre-processing chain to homogenize the spectral signatures has been developed, obtaining
3 types of datasets (using different pre-processing chain) in order to determine which one provides the best
classification results using a Random Forest classifier. The experimental results of this supervised
classification algorithm to distinguish between normal and tumour tissues have achieved more than 99% of
accuracy.
1 INTRODUCTION
Malignant brain tumours, with a global incidence
around 3.5 per 100,000 people, are among the most
lethal and challenging cancers for treatment. Surgical
resection is one of the most important pillars in the
treatment of these tumours, but due to their locations,
sometimes arising from very eloquent areas of the
brain, and their diffuse and infiltrating limits, the total
excision is sometimes cumbersome or impossible to
achieve.
Modern Neurosurgery for these tumours relies on
image-guided resection, but it needs expensive and/or
invasive techniques, such as the Neuronavigation,
intraoperatory Magnetic Resonance Imaging (MRI),
injection of reactive for immunofluorescence, etc.
The goal of this investigation is to apply an innovative
and non-invasive technology tool for image-guided
brain tumour resection: Hyperspectral imaging.
This technology is a non-contact, non-ionizing
and non-invasive sensing technique very suitable for
medicine (Lu et al., 2014); (Akbari et al., 2012). It
consists in collecting and processing information
across the electromagnetic spectrum creating a
hyperspectral data-cube with the values of the
reflectance of the light captured in the scene for
different frequencies. This kind of images increases
the amount of information acquired in a scene
compared with the conventional RGB image or a
multispectral image (which has around ten bands), by
capturing data in a large number of contiguous and
narrow spectral bands over a wide spectral range.
Using the information generated by hyperspectral
imaging, it is possible to obtain a spectral signature of
each pixel. This spectral signature allows
differentiating the material or substance that is
presented in the pixel. It is expected that tumours will
be detected as changes in the spectral signatures
compared with normal tissues (Fei et al., 2012);
(Martin et al., 2006).
The work presented in this paper has been
developed within the HELICoiD (HypErspectraL
Imaging Cancer Detection) project. HELICoiD is a
European FET (Future Emerging Technologies)
project that has the aim of discriminating between
normal and tumour tissues in the surface of the human
brain during neurosurgical operations in order to
Fabelo, H., Ortega, S., Guerra, R., Callicó, G., Szolna, A., Piñeiro, J., Tejedor, M., López, S. and Sarmiento, R.
A Novel Use of Hyperspectral Images for Human Brain Cancer Detection using in-Vivo Samples.
DOI: 10.5220/0005849803110320
In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 4: BIOSIGNALS, pages 311-320
ISBN: 978-989-758-170-0
Copyright
c
2016 by SCITEPRESS Science and Technology Publications, Lda. All r ights reserved
311
provide the neurosurgeons with a real-time guide that
can help in the adequate surgical resection. As a
second goal, the project will try to obtain
hyperspectral signatures from different tumours, so it
might give clues to the neurosurgeon about the
tumour histology.
The results of this discrimination process will be
shown to the neurosurgeons by using a false colour
map where tumour and healthy tissues will be clearly
differentiated. This colour map will help them to
accurately determine the tumour boundaries in the
process of brain tumour resection, avoiding excessive
extraction of healthy tissue and the accidental leaving
of small tumour tissues.
2 MATERIALS AND METHODS
This section provides an overview of the
instrumentation and the methodology used to collect
the in-vivo hyperspectral data of human brain
samples.
2.1 Hyperspectral Imaging
Instrumentation
In order to obtain the hyperspectral images of the in-
vivo human brain surface during the neurosurgical
operations, the HELICoiD project has built a
demonstrator capable of simultaneously obtaining
two hyperspectral cubes. The two hyperspectral
cameras selected are the Hyperspec
®
VNIR A-Series
and the Hyperspec
®
NIR X-Series, manufactured by
HeadWall Photonics, Massachusetts, USA. The
VNIR (visible and near infrared) camera ranges
between 400 nm to 1000 nm. The NIR (near infrared)
camera ranges between 900 nm to 1700 nm.
Figure 1 shows the main parts of the
demonstrator. The most important elements of the
system are located in the acquisition scanning
platform. Table 1 presents the specifications of the
two push-broom hyperspectral cameras. These
cameras are fixed in a scanning unit composed by a
stepper motor and a screw with a maximum path of
230 mm and a step resolution of 6.17 µm.
Furthermore, a cold light emitter is located together
with the cameras. The cold light emitter is connected
to a 150 W Quartz Tungsten-Halogen system (QTH)
(Figure 2.c), which offers broadband emission in the
VIS (visible) and NIR spectral ranges (400 nm to
2200 nm), through an optical fibre. This system
isolates the high temperatures produced by the
halogen lamp, avoiding a direct emission to the brain
surface.
Data pre-processing system is composed by a
high performance computer which manages the entire
system, especially the acquisition scanning platform
and the interaction with the user through the graphical
user interface (GUI).
Figure 1: HELICoiD demonstrator main parts.
Finally, the processing sub-system platform has
the goal of performing the hyperspectral
classification in order to achieve the results in real-
time. The platform selected for this issue is the Kalray
many-core processor that features MIMD (Multiple
Instruction Multiple Data) architecture (B. D. de
Dinechin et al., 2013). This platform is focused on
intensive computing, low power and embedded
applications.
Table 1: Camera Specifications.
Hyperspec
®
VNIR
Hyperspec
®
NIR
Spectral range (nm)
400 – 1000 900 – 1700
Spectral resolution (nm)
2 – 3 5
Slit (μm)
25 25
Spatial bands
1004 320
Spectral bands
826 172
Frame Height (FOV) (mm)
129.21 153.6
Pixel Dimensions (IFOV)
(mm)
0.1287 0.4800
Max Pixels per Frame
1004 320
Max Frames per Capture
1825 489
Dispersion per pixel (nm)
0.74 4.8
Detector array
Silicon CCD InGaAs
Frame rate (fps)
90 100
Smart-BIODEV 2016 - Special Session on Smart Embedded Biomedical Devices for In Situ Physiological Signal Processing
312
Figure 2.b shows the HELICoiD demonstrator
inside the pre-operative area at the University
Hospital Doctor Negrín in Las Palmas de Gran
Canaria, Spain. Figure 2.a presents the acquisition
platform where the cameras and the cold light
element are located. On the left side of the platform,
the VNIR camera is located, and on the right side the
NIR camera is placed. In the middle of the two
cameras, the cold light emitter is located. These three
elements are correctly aligned in order to obtain the
images properly illuminated. Figure 2.d displays the
stepper motor controller, which is in charge of
managing the scanning platform shift.
(a)
(b)
(c)
(d)
Figure 2: (a) Acquisition scanning platform, (b) complete
HELICoiD demonstrator, (c) Quartz Tungsten-Halogen
system and (d) stepper motor controller.
2.2 Hyperspectral Image Dataset
Using the HELICoiD demonstrator, an in-vivo
human brain hyperspectral image database has been
created. The hyperspectral cubes have been obtained
from 13 different patients at the University Hospital
Doctor Negrín. The disease of the tissues captured
during this study involves both primary brain tumours
and secondary tumours (metastasis). All primary
tumours captured have been diagnosed as grade IV
glioblastoma. For secondary tumours two different
types of metastasis, lung and renal, have been
collected.
From this database, a dataset formed by normal
brain tissue and tumour tissue (primary and
secondary) of hyperspectral samples have been
collected and labelled. The work presented in this
paper is only focused in the VNIR hyperspectral
cubes as they have provided better results. Table 2
shows the number of the labelled in-vivo human brain
samples available from the VNIR hyperspectral
cubes.
Table 2: HELICoiD Labelled Spectral Signature Data Base.
Tissue Type Patients # Samples
Normal 9 12604
Tumour
Primary 8 10059
Secondary 4 1983
In order to obtain the samples correctly labelled,
the four steps flowchart presented in Figure 3 has
been followed. First, when the neurosurgeons have
the brain surface exposed, they place two sterilised
rubber ring markers over it. One marker is placed
over the zone where clearly the tumour lesion is
located. The other marker is placed over an area far
from the tumour lesion, where the neurosurgeon can
be quite confident that the brain tissue is healthy.
After that, the operator of the HELICoiD
demonstrator captures the hyperspectral image of this
exposed brain surface.
Figure 3: Data capture and labelling process.
So as to identify the location of the markers over
the brain, the neuronavigator pointer is used. Figure 4
illustrates the use of the neuronavigator to identify the
position of the markers in a MRI.
Next, neurosurgeons remove the tissue inside the
tumour marker. This tissue is sent to the pathologists,
which are the experts who can determine the real
A Novel Use of Hyperspectral Images for Human Brain Cancer Detection using in-Vivo Samples
313
diagnosis of the tissue inside the marker. If the brain
tissue is tumour, pathologists specify the grade and
the type of the tumour.
(a)
(b)
Figure 4: (a) Neuronavigator pointer over the tumour
marker located on the brain surface exposed and (b),
neuronavigator screen capture with the coordinates of the
tumour marker in a MRI.
Finally, with this information, the pixels inside the
markers are cropped manually, avoiding pixels which
could have specular reflections produced by the non-
uniformity of the brain surface. These selected pixels
are labelled and stored with the information provided
by the pathologists. Labelled pixels will be used as
inputs in a supervised classification algorithm
scheme.
Figure 5 presents the most representative bands of
the VNIR hyperspectral image of the patient 12’s
brain surface captured by the demonstrator.
3 CLASSIFICATION SYSTEM
For performing a spectral classification using the
hyperspectral images captured, a classification
system based on a Random Forest (RF) classifier has
been defined. Figure 6 shows an overview of this
classification system.
The first stage of the proposed classification
system is the acquisition step, where the labelled
dataset of the tumour and normal samples are
collected. The procedure followed to collect these
data has been previously described in section 2.2.
After the acquisition stage, a pre-processing chain
is applied to the labelled dataset. In this pre-
processing stage an image calibration is done in order
to address the problem of the spectral non-uniformity
of the illumination device and the dark current.
Furthermore, a set of steps with the goal of removing
the noise of the spectral signatures and to reduce the
number of bands of the samples without lose the main
spectral information are applied.
Finally, so as to homogenize the spectral
signatures in terms of reflectance level, a pixel bright
correction step and a normalization step are
performed.
In the classification stage, the labelled dataset is
partitioned into two different datasets. Training
dataset is used to generate the classifier model while
test dataset is used to validate this model, obtaining
the results of the classification. Sensitivity, specificity
and overall accuracy are the evaluation metrics
chosen in order to know the goodness of the classifier
model. These evaluation metrics will be described
later.
Figure 5: Most representative bands of the VNIR hyperspectral image (400 nm to 1000 nm) from patient 12.
Smart-BIODEV 2016 - Special Session on Smart Embedded Biomedical Devices for In Situ Physiological Signal Processing
314
Figure 6: Classification system overview.
3.1 Data Pre-processing
A pre-processing chain composed by three main steps
(image calibration, spectral noise and band reduction
and data normalization) has been developed in order
to homogenize the spectral signatures of the labelled
samples obtained from the in-vivo hyperspectral data-
cubes.
3.1.1 Image Calibration
The first step in the pre-processing chain is the image
calibration, where the significant signal variations
caused by the non-uniform illumination over the
surface of the captured scene are corrected. The
acquired raw image is calibrated using the white and
dark reference images.
White and dark reference images are acquired by
the demonstrator inside the operating theatre under
the same illumination conditions used to acquire the
in-vivo brain surface images. The white reference
image is obtained from a standard white reference tile
and the dark reference image is obtained by keeping
the camera shutter closed. The hyperspectral
calibrated image is calculated using the equation (1),
where CI is the calibrated image, RI is the raw image
and WR and DR are the white and dark reference
images respectively. Figure 7 shows the spectral
signature of a grade IV glioblastoma tumour tissue
before the calibration step (raw pixel) and Figure 8
after the applied calibration.
 =100


(1)
3.1.2 Noise and Dimensionality Reduction
The second step in the pre-processing chain is to
apply a series of filters in order to remove the noise
existing in the spectral signatures, mainly due to the
CCD sensor of the VNIR camera.
First of all, the noise filter which conforms the
first step of the HySIME algorithm is applied,
reducing a large amount of noise from the spectral
signatures. This function, which is named
Hyperspectral Noise Estimation, infers the noise in a
hyperspectral data set, by assuming that the
reflectance at a given band is well modelled by a
linear regression on the remaining bands (Bioucas-
Dias and Nascimento, 2008); (Nascimento and
Bioucas-Dias, 2015). Figure 9 shows the same
spectral signature after having applied this noise
filter.
Figure 7: Raw spectral signature of a grade IV glioblastoma
tumour tissue.
Figure 8: Calibrated spectral signature of a grade IV
glioblastoma tumour tissue.
A Novel Use of Hyperspectral Images for Human Brain Cancer Detection using in-Vivo Samples
315
After this step, the bands from 0 to 50 and the
bands from 750 to 826 are removed since these bands
contain too much noise due to the limited
performance of the CCD sensor, the grate and the
light scattering in the extreme bands. This fact can be
seen in Figure 9. Additionally, this step reduces the
number of bands in the spectral signatures from 826
to 700 bands.
Afterwards, a smoothing technique is
independently applied to each pixel of the image. This
technique modified each pixel
of the spectral
signature of the pixel under analysis, =
(
,
,
…,
), where k is the selected pixel and N
B
is the original number of bands. The new value of the
“smoothed point” (
)
is the average of the values
corresponding to predefined number of its
surrounding points, as shown in equation (2), where
is number of bands to be combined.
(
)
=

/(2+1)


(2)
Due to the extremely high spectral resolution of
the images, it has been observed that consecutive
bands are correlated, providing redundant
information. In order to avoid this redundancy and
speed up the hyperspectral analysis of the data set, a
few bands have been removed. Moreover, it is not
needed to perform the smooth filter for those bands
that are going to be removed, which eases the filtering
process in terms of computational burden. In
particular 129 spectral bands, from the 700 spectral
bands previously processed, have been totally
filtered, uniformly covering the spectral range from
400 to 1000 nm as shown in Figure 10.
3.1.3 Data Normalization
Due to the surgery procedure, the pixels are captured
at different height, and hence, at different radiation
intensity. This fact typically causes that pixels
labelled as tumour and normal tissue have very
different radiation intensities. If these pixels are
introduced without any pre-processing in a classifier,
the pixels could be classified according to its
brightness, without really taking into account their
spectral signatures. In order to avoid this fact, a pre-
processing step which normalizes the brightness of
the pixels in the image needs to be included. This
process calculates the brightness of each pixel of the
hyperspectral image and divides each pixel by its
brightness, as shown in equation (3). In this equation

is the pixel with the brightness correction, is the
pixel to be corrected and
is the i-th component of
this pixel. With this pre-processing step, the
brightness of each pixel is homogenized without
modifying its spectral signature. Figure 10 illustrates
the final spectral signature with the full pre-
processing chain applied.

=


(3)
Figure 11 presents the VNIR RGB image of the
patient 12, with the tumour area remarked (surgeon
prior evaluation), and the most representative features
of the final pre-processed data-cube. As it can be
appreciated, in feature 45 veins and tumour tissue
have a low brightness regarding to the normal tissue.
However, in feature 55 and 65, veins and normal
tissues have approximately the same brightness level
while the area where tumour is located exhibit lower
brightness. Feature 80 allows seeing veins in high
brightness conditions while tumour and normal
tissues have the same brightness. Finally, the feature
125 is relevant because where the tumour area is
located there is a high level of brightness. This fact
suggests that this pre-processed chain additionally
can obtain high level of contrast to distinguish
between veins, normal tissues and tumour tissues.
Figure 9: Spectral signature with the HySIME filter applied
to a grade IV glioblastoma tumour tissue.
Figure 10: Spectral signature with the noise and band
reduction step and the normalization applied to a grade IV
glioblastoma tumour tissue.
Smart-BIODEV 2016 - Special Session on Smart Embedded Biomedical Devices for In Situ Physiological Signal Processing
316
Figure 11: Most representative features of the final pre-
processed image from patient 12.
3.2 Classification Algorithm
In this research work a supervised learning algorithm
has been employed, where the input features of the
classifier consist in the spectral signatures extracted
from brain tissue. The data mining algorithm chosen
for classifying data is Random Forest (RF). This
algorithm has been already used in the classification
of hyperspectral data (Ham et al., 2005). Random
Forest is an ensemble of Decision Trees (DTs), where
each tree has been generated with the same training
set, but is growing using different random vectors
(Breiman, 2001). A single Decision Tree handles
high-dimensional data well, has the ability to ignore
irrelevant features and provides an easy model
interpretation. However, DT usually has relatively
low prediction accuracy. Due to the advantages
provided by DT, many efforts to improve its
prediction accuracy has been proposed. It has been
discovered that one of the best ways to improve the
performance of Decision Tree-based algorithms is
using ensembles of DT, like Random Forest classifier
(Svetnik et al., 2003). The output of RF classifier is
calculated as the most popular class voted by the
trees.
Advantages of RF compared to other statistical
classifiers include very high classification accuracy;
a novel method of determining variable importance;
ability to model complex interactions among
predictor variables; flexibility to perform several
types of statistical data analysis, including regression,
classification, survival analysis, and unsupervised
learning; and an algorithm for imputing missing
values (Cutler et al., 2007).
3.3 Experimental Setup
The experimental setup chosen for this study merges
all available hyperspectral labelled data (from 13
different patients) in a single dataset. The dataset
employed in this research work has been created by
joining each single operation hyperspectral labelled
data, even if a unique class is given for a certain
operation. As the training and testing stages for
classification have been performed using data from
all the operations, the inter-patient variability of the
data will be taken into account.
The labelling of data has been performed using
two different abstraction levels of the diagnosis of the
tissue. In the first level, tissues have been grouped in
'Normal' tissues and 'Tumour' tissues, and the
classification using this labelling scheme has been
named 'Tag Level 1'. For the second labelling scheme,
'Tumour' tissues have been divided in 'Primary' and
'Secondary' tissues, attending to the diagnosis
provided by pathologists. This labelling scheme has
been named 'Tag Level 2'. The summary of the
dataset is given in Table 3.
In order to estimate the classifier performance,
and for obtaining the optimal configuration of the
selected algorithm, a three-way cross validation has
been employed. Three-way cross-validation consists
in two different stages for splitting the available
dataset: in the first stage, k-fold cross-validation is
used in order to get training and testing subsets. Test
data will be used to estimate the classifier
performance, and the training data are partitioned
again into training data and validation data (Figure
12). The training subset of the second cross-
validation stage is used to create the model of the
classifier, and the validation data are used to evaluate
the performance of the classifier. With the second
stage partition, the model fitting will be
accomplished, and the parameters of the classifier
will be modified in order to obtain the optimal
configuration of the algorithm. The test set obtained
in the first dataset split, is used to perform the
performance evaluation of the algorithm by using
unknown data for the classifier. The k value selected
for both cross-validations is 10, which is a typical
value used in data mining
Figure 12: Three-way cross-validation experimental setup
overview.
A Novel Use of Hyperspectral Images for Human Brain Cancer Detection using in-Vivo Samples
317
Table 3: HELICoiD Labelled Dataset with two
classification tag levels.
#Operation
Diagnosis
#Samples
Tag Level 1 Tag Level 2
1
Normal 408
Tumour Secondary 578
4
Normal 1939
Tumour Secondary 522
5
Normal 832
Tumour Secondary 493
6 Normal 806
7 Normal 768
8
Normal 1484
Tumour Primary 3259
10 Tumour Primary 425
12
Normal 806
Tumour Primary 1424
13 Tumour Secondary 390
14 Tumour Primary 1139
15
Normal 648
Tumour Primary 800
16
Normal 4913
Tumour Primary 1987
17 Tumour Primary 1025
Total 24646
3.4 Evaluation Metrics
The goodness of the classifier has been measured
using sensitivity, specificity and overall accuracy
metrics. Sensitivity measures the test ability to
identify a condition correctly. It is computed as
follows:
 =

+
(5)
where TP is the number of true positives and FN is
the number of false negatives in a population.
Specificity measures the test ability to exclude a
condition correctly. It is expressed as follows:
=

+
(6)
where TN is the number of true negatives and FP is
the number of false positives in a population.
Finally, the equation (7) shows the accuracy
metric that represents the percentage of total correctly
classified samples in a population:
 =
+
+++
(7)
4 EXPERIMENTAL RESULTS
The hyperspectral classification experiments have
been performed using the three different pre-
processed data previously described. The calibrated
data are the labelled samples with only the white and
dark calibration applied. HySIME filtered data are the
previous calibrated samples with the HySIME filter
applied over them. The last set of data has the
complete pre-processed chain applied, this set of data
is called pre-processed samples.
As it was commented previously, two different
levels of diagnosis detail have been evaluated.
4.1 Tumour Vs Normal: Tag Level 1
In this section will be presented the results obtained
using the Random Forest classification system over
the three different set of data taking into account the
first tag level (normal vs tumour tissues).
The results of the classification in this case study
shows that an automatic discrimination between
'Normal' tissue and 'Tumour' tissue is possible using
the hyperspectral signature of the tissues. Sensitivity
and specificity lie in the same range, which means
that the algorithm is capable to identify both kinds of
tissues.
Although the classification results provide
accurate discrimination rates in terms of accuracy,
specificity and sensitivity, varying the pre-processing
stage results in an improvement of the classification.
From data shown on Table 4, it can be seen that the
accuracy improves from around 93%, when the pre-
processing chain consists only of the calibration of
the hyperspectral image, to 99% when using the
whole pre-processing chain presented in Figure 6.
Figure 13 presents these results in a bar chart.
Figure 13: Comparison between the classification results of
the three data sets at tag-detail level 1.
93,67%
96,06%
99,68%
94,05%
95,67%
99,68%
93,24%
96,49%
99,67%
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
Calibrated HySIME Filtered Pre-processed
Accuracy (%) Sensitivity (%) Specificity (%)
Smart-BIODEV 2016 - Special Session on Smart Embedded Biomedical Devices for In Situ Physiological Signal Processing
318
Table 4: Comparison between the classification results of
the three data sets at the tag level 1.
Calibrated HySIME Filtered Pre-processed
Sensitivity (%)
94.05 95.67 99.68
Specificity (%)
93.24 96.49 99.67
Accuracy (%)
93.67 96.06 99.68
This trend is kept for the rest of the evaluation
metrics: specificity improves from 93% to 99% and
sensitivity improves from 94% to 99%, when the full
pre-processing chain is used, instead of performing
only the calibration.
4.2 Normal Vs Primary Vs Secondary:
Tag Level 2
The same data have been classified with a different
tag scheme, where 'Tumour' tissues have been divided
into 'Primary' and 'Secondary' tumour tissue labels.
Figure 14 illustrates the accuracy results between the
different pre-processed datasets and Table 5 to Table
7 show the classification results in terms of sensitivity
and specificity for each class. These data have been
obtained by calculating the confusion matrix of each
dataset that can be seen in Table 8. The error
estimation of this classification shows that the
algorithm can keep a good performance on
discriminating between normal and tumour tissue,
even using a more complex labelling scheme. The
sensitivity and specificity values show also high
values, which indicates that all classes have been
properly classified. Results show again that the used
pre-process chain improves the results of the
classification.
Figure 14: Accuracy comparison between the classification
results of the three data sets in tag level 2.
Table 5: Classification results of the calibrated dataset in
tag level 2.
Sensitivity (%)
Normal Primary Secondary
Specificity (%)
Normal
- 94.67 99.67
Primary
94.63 - 94.34
Secondary
97.64 89.21 -
Table 6: Classification results of the HySIME filtered
dataset in tag level 2.
Sensitivity (%)
Normal Primary Secondary
Specificity (%)
Normal
- 96.07 99.51
Primary
95.49 - 96.4
Secondary
99.45 91.88 -
Table 7: Classification results of the pre-processed dataset
in tag level 2.
Sensitivity (%)
Normal Primary Secondary
Specificity (%)
Normal
- 99.92 100.00
Primary
99.31 - 99.90
Secondary
100.00 100.00 -
5 CONCLUSIONS
In this paper, it has been described a hyperspectral
acquisition system used to create a hyperspectral
database of human brain tissues previously diagnosed
as tumour or normal. In each surgical procedure, a
few rubber ring markers have been placed by the
neurosurgeons to get assessed diagnosis from
pathology. Some of these markers were located in
areas of brain where neurosurgeons were quite sure
that the tissues were healthy, and the other markers
were placed where the resection was going to be
performed. A biopsy from the rejected tissue was sent
to pathology, providing assessed diagnosis of the
tissues inside the marker. These samples were used as
the ground truth for classification.
Table 8: Confusion matrix of the three types of datasets in tag level 2.
Predicted Results
Calibrated HySIME Filtered Per-Processed
Normal Primary Secondary Normal Primary Secondary Normal Primary Secondary
Gold
Standard
Normal 1225
52 3
1223
43 1
1249
7 0
Primary
69
917
15 50
910
16 1
1013
0
Secondary
4 55
124
6 34
181
0 1
193
91,96%
93,91%
99,63%
0%
20%
40%
60%
80%
100%
Calibrated HySIME Filtered Pre-processed
Accuracy (%)
A Novel Use of Hyperspectral Images for Human Brain Cancer Detection using in-Vivo Samples
319
Taking into account the diagnostic information
provided by pathologists, the pixels inside markers
were extracted from the hypercubes and labelled
according to the diagnosis. Due to the complexity of
the diagnostic information, a labelling scheme
consisting in two abstraction levels of disease details
had been proposed.
The classification results shown in section 4 show
that it is possible to obtain an accurate and automatic
discrimination between different types of tissues
using the labelling schemes proposed. Although the
three proposed pre-processing chains provided
accurate classification results (accuracy higher than
89% for all the classifications), the more complex one
provided the best classification results in all the
experiments exposed in this paper.
In the near future, some additional research is
foreseeable to be done. Firstly, the complexity of the
diagnosis can be further explored. For instance,
primary tumours could be classified according to its
Grade, and Secondary tumours (metastasis) could be
differentiated attending to their origin (breast, lung,
etc.). The next step will be to define a more complex
labelling scheme to better classify the type of tumour.
Secondly, we are working in the design a case study
where the automatic diagnostic of a new patient could
be computed by using a model that had been created
using the hyperspectral data from previous (and in
consequence different) patients. Thirdly, it could be
interesting to test the performance of other different
machine learning algorithms, like the support vector
machines (SVM), the neural networks (NN), etc.
Finally, due to the large experience that the research
group has in hardware implementations, we are
considering the implementation of the pre-processing
and classification algorithms in some hardware
platform (FPGA, GPU, ASICs, many cores, etc.) to
accelerate its execution.
ACKNOWLEDGEMENTS
This work has been supported in part by the European
Commission through the FP7 FET Open programme
ICT-2011.9.2, European Project HELICoiD
“HypErspectral Imaging Cancer Detection” under
Grant Agreement 618080.
REFERENCES
G. Lu, B. Fei, 2014. Medical hyperspectral imaging: a
review. Journal of biomedical optics, vol. 19, no. 1, pp.
010 901–010 901.
H. Akbari, L. V. Halig, D. M. Schuster, A. Osunkoya, V.
Master, P. T. Nieh, G. Z. Chen, and B. Fei, 2012.
Hyperspectral imaging and quantitative analysis for
prostate cancer detection. Journal of biomedical optics,
vol. 17, no. 7, pp. 0 760 051–07 600 510.
B. Fei, H. Akbari, L. V. Halig, 2012. Hyperspectral imaging
and spectral-spatial classification for cancer detection.
Biomedical Engineering and Informatics (BMEI), 5th
International Conference on. IEEE, pp. 62–64.
M. E. Martin, M. B. Wabuyele, K. Chen, P. Kasili, M.
Panjehpour, M. Phan, B. Overholt, G. Cunningham, D.
Wilson, R. C. DeNovo, & T. Vo-Dinh, 2006.
Development of an advanced hyperspectral imaging
(HSI) system with applications for cancer detection.
Annals of Biomedical Engineering, 34(6), pp. 1061–
1068.
B. D. de Dinechin, R. Ayrignac, P.-E. Beaucamps, P.
Couvert, B. Ganne, P. G. de Massas, F. Jacquet, S.
Jones, N. M. Chaisemartin, F. Riss, T. Strudel, 2013. A
clustered manycore processor architecture for
embedded and accelerated applications. High
Performance Extreme Computing Conference (HPEC),
IEEE, pp.1-6.
J. M. Bioucas-Dias and J. M. Nascimento, 2008.
Hyperspectral subspace identification. Geoscience and
Remote Sensing, IEEE Transactions on, vol. 46, no. 8,
pp. 2435–2445.
J. M. Nascimento and J. M. Bioucas-Dias, 2015.
Hyperspectral noise estimation.
https://github.com/jhausser/ParTI/blob/master/mvsa_d
emo/estNoise.m, last accessed: November 2015.
JiSoo Ham, Yangchi Chen, Melba M. Crawford, 2005.
Investigation of the Random Forest Framework for
Classification of Hyperspectral Data. IEEE
Transactions On Geoscience and Remote Sensing, Vol.
43, No. 3, pp. 492 – 501.
Breiman, L., 2001. Random forests. Machine learning, Vol.
45, No 1, pp. 5-32.
Svetnik, V., Liaw, A., Tong, C., Culberson, J. C., Sheridan,
R. P., & Feuston, B. P., 2003. Random forest: a
classification and regression tool for compound
classification and QSAR modeling. Journal of chemical
information and computer sciences, Vol. 43, No 6, pp.
1947-1958.
R. Cutler, T. C. Edwards, K. H. Beard, A. Cutler, K. T.
Hess, J. Gibson and J. J. Lawler, 2007. Random Forests
for Classification in Ecology. Ecology, Ecological
Society of America, Vol. 88, No. 11, pp. 2783-2792.
Smart-BIODEV 2016 - Special Session on Smart Embedded Biomedical Devices for In Situ Physiological Signal Processing
320