Temporally Synchronized Reversible Data Hiding of EEG to MREG

Angelos Fylakis

1, 2

, Anja Keskinarkaus

1, 2

, Vesa Kiviniemi

3, 2

and Tapio Seppänen

1, 2

Center for Machine Vision and Signal Analysis, University of Oulu, Oulu, Finland

MRC Oulu, Oulu University Hospital, Oulu, Finland

Oulu Functional Neuroimaging, MIPT Research Group, University of Oulu, Oulu, Finland

Keywords: Data Hiding, Data Management, Brain Research, Multimodality.

Abstract: Simultaneous MREG and EEG recordings are vastly used in neurobiology, but so far they are stored and

handled as separate files. This paper proposes a method to combine those two entities with the objective of

establishing data management efficiency, while secondary objectives are confidentiality, availability and

reliability in data. To be more specific, it is a reversible data hiding method for hiding EEG in MREG with

the ability of fully recovering MREG and the embedded EEG signal. It is based on histogram shifting,

exploiting data quantization and Region of Interest segmentation. The embedding procedure maintains

temporal synchronization between EEG and 32-bit MREG making it a novel data hiding application. It is

demonstrated through experiments that MREG maintains high perceptual fidelity and also verified that after

EEG extraction and acquisition of every electrode’s sample, MREG is fully reversed to its exact initial state.

1 INTRODUCTION

Modern ultrafast Magnetic Resonance Imaging

(MRI) sequences combined with multimodal data

produce vast amounts of data creating efficiency and

security related problems. Considering this together

with Human Connectome projects it is clear that

demands on data storage and analysis are constantly

increased. In such applications, data is stored in big

databases called electronic healthcare records. They

are either based on local hospital networks or cloud

networks. This paper presents a method to increase

data management efficiency and data security, which

are the main problems in most e-health applications.

The proposed method applies data hiding

techniques and embeds Electroencephalography

(EEG) data in MR-Encephalography (MREG)

recordings. MREG sequences are similar to fMRI

enabling even faster and more sensitive monitoring

of functional activation of the brain sampled every

25-100 msec. EEG-fMRI/MREG signal recordings

appear to have great importance in neurobiology

enabling researchers to understand neural behaviour.

fMRI/MREG provides detailed spatial resolution

showing activated brain areas but not details as to

temporal resolution. EEG, on the other hand,

provides information related to temporal resolution

promoting study on the dynamics of brain function,

while its poor spatial resolution restricts

identification of the neural sources (Menon and

Crottaz-Herbette, 2005). Overall, this is what makes

EEG and fMRI/MREG complementary data.

The proposed method will focus on the

management of temporally simultaneous EEG-

MREG recordings. Compared to non-simultaneous,

the simultaneous recordings have the advantage that

the two data types reflect the same neuronal

processes. This is because for both recordings the

condition of the subject is the same. Simultaneous

EEG-MREG recordings are used for instance in

localizing epileptic seizure. There are multiple

applications in the research area other than the

clinical one as researchers try to make a better

understanding of the neural processes. Makeig et al.,

(2002) and Czisch et al., (2004) are examples of

research papers that use simultaneous recordings for

clinical and developmental studies. Another paper is

the one by Jacobs et al., (2014). In their paper, they

analyse epileptic spikes from EEG-MREG to

determine the yield of fast MRI in the analysis of

intrinsic brain signals. A more recent paper is the

one by Rajna et al., (2015) detecting patterns of

brain activity by exploiting the superior spatial

accuracy of MREG data and the temporal dynamics

provided by EEG signals being 500 times faster than

MREG.

As the use of multimodal data such as EEG-

Fylakis, A., Keskinarkaus, A., Kiviniemi, V. and Seppänen, T.

Temporally Synchronized Reversible Data Hiding of EEG to MREG.

DOI: 10.5220/0005665700580067

In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 5: HEALTHINF, pages 58-67

ISBN: 978-989-758-170-0

MREG with an increasing amount of details

recorded becomes more commonly used, efficiency

in data management becomes crucial. Efficiency

first refers to the storage and transfer of multimodal

data. The problem with multimodality is that when

data is transferred through the Internet and hospital

servers, there is risk of data loss due to the transfer

and storage of multiple files. Secondly, efficiency

refers to the high capacity of the data and the time to

access and analyse it. In our case, the use of EEG

and MREG as separate entities requires that access

to specific segments over time has to be done

separately. Also, as data is commonly stored

separately, it reserves more space and within

filesystems it requires that linkages between files

need to be handled manually.

Beyond the efficiency requirements are the

security requirements. Constant transfer of

biomedical data through networks and storage in

cloud databases raises security issues. Some among

the most important ones are confidentiality,

availability and reliability in biomedical data

(Coatrieux, et al., 2006). Concerning confidentiality,

biomedical data is private patient information and

thus direct access to all data by non-authorised users

would violate privacy. Availability refers to the

ability for direct access to all inconsistent entities.

Last, reliability has to do with the problems of

verifying integrity of data, as well as tracing and

validating authentic data. Tampered data can

mislead and cause errors in the diagnosis and similar

problems can occur when non authentic material is

used, so mechanisms for tracing and validating

authentic content are also required.

Data hiding has proved to significantly satisfy

those requirements by enriching data with metadata

and thus providing a new layer of security. Data

hiding is defined as the practice of imperceptibly

altering an object to embed a message about it (Cox

et al., 2007). In this case, we exploit data hiding

principles to firstly improve efficiency in data

management by embedding EEG signals in MREG

sequences. Secondly, confidentiality, availability

and reliability of data can be guaranteed. Access to

the hidden data can be restricted only to entitled

users, to ensure confidentiality. Then, as multimodal

data co-exist in a single package, availability is

ensured. Also, to improve reliability, hiding IDs or

digital signatures along host data can be used as a

proof of data’s ownership and authenticity (Grover,

1997) and in a similar manner, tamper-proofing is

solved. The target is to prove integrity with

embedded data and locate any possible tampering.

Embedded digital signatures in the host object are

capable after extraction and reversion of confirming

that no tampering has taken place and further use is

safe.

Integrity also requires reversibility in data hiding

techniques where modifications cause certain non

visible or slightly visible artefacts on host data. In

biomedical data, it is essential to use intact

information in the analyses. The solution comes with

reversible data hiding techniques. Reversible data

hiding refers to the recovery of the host medical

object to its exact initial state (Dharwadkar et al.,

2010).

Here, we propose a reversible data hiding

method based on histogram shifting. The original

histogram shifting method was proposed by Ni et al.

(2006) followed by improvements by Tsai et al.

(2009) and Fallahpour et al. (2011). The method

was implemented for 8-bit image applications; while

in this paper, we consider a multi-dimensional video,

that is, the 32-bit MREG as host. This is the main

novelty of this paper differentiating it from past

research which was focused on static images (Pan et

al., 2009). This paper further incorporates data

quantization in order to apply histogram shifting on

32-bit data.

Another novelty is temporal synchronization.

The EEG signal and/or other signals to be embedded

in the MREG sequences are synchronized in the

time domain. In practice, accessing an MREG

segment guarantees extraction of the simultaneously

recorded EEG. This is significant both time and

capacity wise as MREGs are big files requiring high

access time which is inefficient in the analyses.

Moreover, from the extracted EEG data, it is

possible to match samples from specific EEG

electrodes as they were embedded in specific order.

In medical imaging, the image can usually be

defined having two regions; the one of interest that

contains the imaged tissue and the variable extent

regions of no interest. A ROI (Region of Interest)

and a RONI (Region of Non Interest) can be selected

either by a medical doctor or automatically,

assuming for example that the background is of less

importance. There are unavoidably always voxels

outside the targeted tissue which have no meaning

for the diagnostic purposes. On the contrary,

distortions on the ROI area can cause errors in the

diagnosis. A novelty in the current method is the

option to restrict the histogram shifting idea to the

ROI, this way, elimination of the background in any

processing steps does not destroy hidden data, while

reversibility guarantees a recovery of the ROI.

The paper is organized as follows. Section 2

includes the description of the data hiding methods,

Temporally Synchronized Reversible Data Hiding of EEG to MREG

the data quantization, ROI segmentation and EEG

compression. Then, it includes the method to avoid

side information and the digital signature

exploitation. Section 3 explains the data formats and

the experimental results, Section 4 is the discussion

part of the paper and last Section 5 includes the

conclusion and suggestions for future work.

2 METHODS

The following subsections describe the embedding

and extracting-reversing methods of the proposed

technique. It can be described as the problem of

embedding a set of data w in a digital object I. Thus,

producing a new object I

, such that w can be reliably

located and extracted from I

(Collberg and Nagra,

2010), reversing at the same time I

to its original

state I. In this case, the message w is the EEG signal

and the host I is the MREG data where w is

embedded in I maintaining temporal synchronization

between I and w.

Not only reversibility is guaranteed, but also

imperceptibility of hidden data can be maintained

with specific settings related to data quantization

described in the following subsection. In our method

we hand data hiding using histogram shifting.

Because of the fact that the original method by Ni et

al. was designed for use in 8-bit images, which is

discrete data having a 256 bin histogram we had to

develop extensions that make use of data

quantization. This outputs a discrete histogram

applicable for the above scheme. MREG unlike

regular images is typically consisted of 32-bit data

and thus does not produce the common histogram of

an 8-bit image but a nearly flat histogram as values

are spread over 2

bins. For this reason the capacity

which is equal to the maximum histogram peak value

is limited.

Furthermore, another extension is a mode of

embedding only in the extracted important areas of

the image to increase robustness. This mode is also

tested and described in more detail within the

subsections that follow. EEG lossless compression

techniques are also suggested which could be of great

benefit in the extended mode where the ROI is

isolated because this restriction reduces capacity

significantly.

Following the algorithmic description of the

methods, subsection 2.5 presents a solution making

the extracting method blind, which means that it does

not require any side information as input. The last

subsection exploits the possibilities of digital

signature exploitation for integrity control, but also

as a measure to confirm the successful run of the

reversion along extraction.

2.1 Data Quantization

In all digital images, in order to form a digital

function, the gray-level values have to be converted,

i.e. quantized into discrete quantities. This process of

assigning gray levels to discrete intensity levels is

called quantization (Gonzalez and Woods, 2002).

The process of quantization can also refer in certain

cases to down-sampling existing discrete values.

Here, we deal with the second case as float single

precision values are actually discretised values but

with a very small quantization step.

In the MREG samples tested in this paper, image

data was as described in 32-bit single-precision

floating-point format, which means 2

histogram

bins while the original histogram shifting method

was optimized for 8-bit grayscale images and thus

for 2

histograms bins. A solution to solve this

format problem was to develop a method that down-

samples the histogram from 2

bins to 2

bins or

lower, as seen in Figure 1.

In our methods, we have tested varying data

quantization options getting histograms with less

than 256 bins in order to increase capacity,

something necessary for certain applications.

Specifically, results include tests from 256 down to

8 histogram bins.

Figure 1: Effect on histogram by exploiting quantization

getting down from 2

(top) to 256 bins (bottom).

Quantization should not downgrade image

quality, otherwise reversion is not possible. For that

purpose the quantization used in current methods

HEALTHINF 2016 - 9th International Conference on Health Informatics

can be better described as “grouping”. Thus, it does

not actually down-sample data but it creates

histogram bins, each one corresponding to an equal

range of data values. The quantization step is

calculated as follows:

Q = (maxv - minv) / bns, (1)

where maxv is the maximum intensity value, minv

the minimum intensity value and bns the target

number of histogram bins.

2.2 Restriction in the ROI

It is likely that modifications occur in the dark

background of the MREG, i.e. the RONI of the

MREG. For instance, voxel values of the

background can be removed or set to 0. Especially,

since this area outside the brain is usually extracted

from any analysis in order to increase the speed and

accuracy of brain analytics. Those modifications

would mean that data embedded there is lost without

making any visually distinctive difference to the

MREG. If we take the background of the brain as an

example, elimination of the whole background

would not make any visible or functional difference

for the analyses. In contrast, any modification in the

brain area affects the analyses, but those

modifications can be detected with methods using

digital signatures such as those described in

Subsection 2.6. As RONI elimination would affect

hidden data, we propose an extension restricting the

hiding target area in the ROI. For the ROI extracting

procedure, we use the BET2 brain estimation

algorithm by Jenkinson et al. (2005) creating a mask

of the imaged tissue, as Figure 2 depicts.

Figure 2: ROI masking using BET2.

2.3 EEG Compression

EEG lossless compression can also be incorporated

before embedding in the MREG to compensate

capacity loss caused by the segmentation and usage

as host of the ROI area only. Some recent examples

include the lossless multichannel compression

method proposed by Wongsawat et al. (2006) and

the lossless method making use of wavelet transform

and neural network predictors proposed by Sriraam

(2012). The above methods can achieve ratios of

2.77 to 1 and 2.99 to 1, respectively.

In the experiments presented in Section 3,

capacity comparisons can be made between

embedding raw and compressed EEG data using the

method proposed by Sriraam. All the experiments

have been performed embedding raw data, while

capacity using compressed EEG has been estimated

using the ratio referred at the paper.

2.4 Data Hiding and Extraction

The idea is applying the histogram shifting scheme

for an MREG sequence hiding an EEG or any other

physiological signal. That has to be done in a

temporal synchronized manner to make benefit of

the techniques that combine simultaneously sampled

recordings of EEG and MREG. Assume that the

sampling rates for EEG and MREG are s

and s

samples per second, respectively. In practice, that

means that for each MREG frame over time, the

embedding algorithm will hide data that corresponds

to F EEG samples.

F = s1 / s2 (2)

Figure 3: Embedding - block diagram.

The following embedding pseudocode describes

step by step a case with F EEG samples per MREG

frame where quantization is performed creating a

histogram with bns bins. An illustration is depicted

at the block diagram of Figure 3.

Temporally Synchronized Reversible Data Hiding of EEG to MREG

EMBED(w[L,B],I[M,N,P,T])

//Step 1

Q = GetQ(I,bns) (eq.1)

c←1

for vmin≤i≤vmax,i=i+Q

Qp[c]←i,c←c+1

//Step 2

for 1≤t≤T

hq[1…bns]←0

for 1≤i≤bns

for 1≤x≤M,1≤y≤N,1≤z≤P

if Qp[i]≤I[x,y,z,t]<Qp[i+1]

hq[i]←hq[i]+1

(mx,mxi)←max(hq) //(value,index)

(mn,mni)←min(hq) //(value,index)

//Step 3

for 1≤x≤M,1≤y≤N,1≤z≤P

if Qp[mxi+1]≤I[x,y,z,t]

if I[x,y,z,t]<Qp[mni]

I[x,y,z,t]←I[x,y,z,t]+Q

//Step 4

s←(t-1)*F+1 (eq.2)

for s≤l<s+F,1≤b≤B

(xi,yi,zi,t)←GetNextIdx(mxi)

if w[l,b]=1

I[xi,yi,zi,t]←I[xi,yi,zi,t]+Q

//Step 5

Iw←I

return Iw

END

First, the algorithm reads the input data w of L

samples over time, each of B bits, converted into a

binary stream and the MREG sequence I of size

MxNxP and temporal resolution T. I might either be

the whole MREG or the segmented ROI area

acquired as described in Subsection 2.2.

In Step 1, the quantization step Q is calculated,

given the MREG’s maximum and minimum

intensity values vmax and vmin and the number of

bins bns according to equation 1.

In Step 2, a loop for all the frames over time is

initiated and for each one the histogram Hq is

generated. Qp stores the positions of quantization’s

threshold points. Then, in the histogram Hq, the

maximum peak value and the minimum (zero) value

are located, which are hq(mxi), mxi ∈ [1,bns] and

hq(mni), mni ∈ [1,bns], respectively, as

demonstrated in Figure 4. Note that in case the

minimum value hq(mni) > 0, then, simply the

coordinates of those voxels and its greyscale values

are kept as overhead information to be hidden along

w and hq(mni) is set to 0.

In Step 3, assuming without loss of generality

that mxi < mni, bins following the peak up before the

zero location, i.e. hq(x), for every x ∈ (mxi,mni) are

shifted to the right by one histogram unit creating

one empty bin next to the maximum peak location’s

bin. This is performed by adding Q to each voxel

within this value range.

In Step 4, the algorithm accesses the voxels

which have intensity value within the range which

corresponds to the peak location histogram bin.

Voxel indexes in this range are returned sequentially

using function GetNextIdx(). As the binary

information to be embedded is accessed

sequentially, the voxel in order maintains its value to

store a 0 bit or is increased by the quantization step

value Q to store a 1 bit.

Finally, at Step 5, the modified MREG Iw is

returned.

Figure 4: Maximum peak location and minimum (zero)

location.

Figure 5: Extracting - block diagram.

For the extracting procedure, the MREG

containing hidden information Iw is entered as input

while the original maximum peak location and

minimum (zero) location are found without the need

of input side information. A solution for blind

extraction without side information is given in the

next subsection while the extracting method is

described in the following pseudocode and also

HEALTHINF 2016 - 9th International Conference on Health Informatics

illustrated at the block diagram of Figure 5.

EXTRACT(Iw[M,N,P,T])

//Step 1

Get Q,Qp (see, embed Step 1)

//Step 2

for 1≤t≤T

Get hq (see, embed Step 2)

Get mxi,mni (see, 2.5)

//Step 3

s←(t-1)*F+1 (eq.2)

for s≤l<s+F,1≤b≤B

(xi,yi,zi)←GetNextIdx(mxi)

if Qp[mxi]≤Iw[xi,yi,zi,t]

if Iw[xi,yi,zi,t]<Qp[mxi+1]

w[l,b]←0

if Qp[mxi+1]≤Iw[xi,yi,zi,t]

if Iw[xi,yi,zi,t]<Qp[mxi+2]

w[l,b]←1

Iw(xi,yi,zi,t)←Iw[xi,yi,zi,t]–Q

//Step 4

for 1≤x≤M,1≤y≤N,1≤z≤P

if Qp[mxi+2]≤Iw[x,y,z,t]

if Iw[x,y,z,t]<Qp[mni+1]

Iw[x,y,z,t]←Iw[x,y,z,t]-Q

//Step 5

Ir←Iw

return Ir, w

END

Following the same idea as in the embedding

method, data is extracted scanning the MREG

sequentially. Binary information is extracted

depending on whether greyscale voxel intensities

have retained the value corresponding to the

maximum peak location or they have been increased

by Q and thus moved to the next histogram bin.

MREG is reversed by returning those values back to

the peak location by subtracting intensities by Q and

shifting the histogram back by reducing the intensity

of the voxels corresponding to bins between the

maximum peak location mxi and the minimum

(zero) location mni by Q. Thus, the histogram is

shifted back to its original state and the MREG is

fully recovered. Last, the recovered MREG I

and

the extracted data w are returned.

2.5 Side Information

Side information refers to an extra input that the

extracting algorithm requires in order to perform its

operations. This input might be a secret key or some

information required for the extraction. In our case,

it refers to the maximum peak and minimum (zero)

histogram locations that have already been indexed

by running the embedding procedure.

In order to avoid the use of side information,

instead of detecting the location of the maximum

peak histogram value, the method detects the

location of maximum sum of consecutive pairs of

histogram values. Of course, this does not guarantee

detection of the right location. Thus, supplementary,

an identification code is also embedded in the first

bits along the payload, including the position of the

zero location. In case an invalid identification has

been extracted, then, the next largest pair of bins is

tested and so on until the right maximum peak and

minimum (zero) locations have been found.

Note that side information can be also used for

confidentiality purposes. Either the extracting

procedure is restricted being accessible only by

specific authorized users, or both the embedding and

extracting procedures require a secret key as side

information. Of course, this key is only known by

the authorized users (Cox et al., 2007). In this case,

the extracting procedure can be designed so that to

require the maximum peak and minimum (zero)

locations in the form of a key. Using an invalid key,

only irrelevant data is extracted.

2.6 Digital Signatures

For integrity control, an option is embedding a

digital signature of the original image along the

embedded payload. The digital signature will be

later proof as it is able to reveal data tampering.

Here, we use the Secure Hash Algorithm 2 (SHA-2)

(National Institute of Standards and Technology,

2014). It is a hash function designed by the U.S.

National Security Agency and published in 2001 by

the National Institute of Standards and Technology.

Different inputs produce different hashes and thus

different digital signatures for the input MREGs.

Here, the digital signatures are produced before

the embedding procedure, meaning that the original

MREG is used as input. After extraction and

reversion, the extracted signature is compared with

one produced from the reversed MREG. If the

signatures are identical, data integrity is confirmed.

This procedure is also useful in order to confirm that

reversion runs successfully. An invalid signature can

reveal errors of the process as the reversed MREG is

not identical to the original host object. This means

that it is not safe to use it for a diagnosis.

3 DATA AND EXPERIMENTS

For the experimental purposes, EEG data has been

collected through open databases available online at

physionet (Goldberger et al., 2000). All data samples

acquired had been already anonymized.

Temporally Synchronized Reversible Data Hiding of EEG to MREG

The MREG samples are also anonymous data

which was acquired from our university hospital

following research procedures with informed

consent.

In all experiments, the host MREG is consisted

by frames of 64x64x64 voxels, running time is 60

sec. and frame rate is 10 frames per second. Data is

in 32-bit single precision RAW NIfTI format.

The embedded EEG is 64 channel RAW data

with sampling rate of 160 samples per second. Each

data sample has 16-bit resolution. In total, in a one

minute run, we have 9,600 samples of 1,024 bits. Let

us note here that testing was not restricted to those

64 channels of embedded data but to the maximum

available capacity figures as well. This is performed

by simply repeating the available 64 channel data.

When ROI was segmented and used as input, the

capacity was significantly reduced as demonstrated

in the following subsection. In order to improve

those capacity figures, the quantization step Q is

increased accordingly, down-sampling data and

consequently decreasing the number of histogram

bins. This reduces marked images’ quality producing

even visible artefacts in certain cases, but

reversibility is guaranteed. Capacity is considerably

increased, but in some cases, it is clear that

compression of the EEG is also required to reach the

64 channel capacity of the testing set. That is why

lossless EEG compression techniques are applied.

The following subsections include information

about the capacity provided in the host MREG

thanks to the data hiding technique, as well as

information on data fidelity comparing the original

MREG with the one that carries hidden data.

3.1 Capacity

Table 1 shows the results acquired from eight

experiments over 1 minute of data measuring the

capacity both in bits of maximum available hiding

space, including standard deviation and in maximum

number of EEG channels that can be hidden.

Capacity refers to hosting capabilities per single

MREG frame. Note that temporal synchronization

was maintained meaning that for each MREG frame,

16 EEG samples over time are hidden.

At the first experiment, the whole MREG image was

used as a host. At the following six experiments, the

ROI of the MREG has been segmented restricting it

as the host area. In this case, different quantization

steps were tested, resulting from 256 to 8 histogram

bins. Restricting to the ROI also means that the

homogenous intensity of the image's background is

lost, reducing capacity significantly in a histogram

shifting technique. That is why higher quantization

is essential to get fair capacity figures. For example,

in our case where the EEG samples were consisted

of 64 channels, in order to get sufficient capacity

quantization step has to be significantly increased

while EEG compression is essential for smaller

quantization steps.

Table 1: MREG’s capacity.

Host area,

# bins

Capacity

# bits

Mode1

# channels

Mode2

# channels

Entire, 256 187958±0.49 734 2185

ROI, 256 888±0.00 3 10

ROI, 128 1366±26.29 5 15 – 16

ROI, 64 2672±42.91 10 30 – 31

ROI, 32 5264±63.09 20 60 – 61

ROI, 16 9846±99.79 38 113 – 115

ROI, 8 18126±185.03 70 – 71 208 – 212

The impact of compressing EEG data is depicted

as follows: Mode 1 shows the number of embedded

uncompressed EEG channels, while Mode 2 the

number of compressed channels estimated according

to the Sriraam’s (2012) method in which a

compression ratio of 2.99 can be achieved.

3.2 Fidelity

For the perceptual quality experiments, Table 2

includes the fidelity figures for the same tests

performed in the previous subsections. In every case,

a 1 minute EEG signal segment was hidden in a 1

minute MREG segment maintaining temporal

synchronization. In the first row, once again, the

entire image was used as host, while at the second

row and below, the ROI was segmented and used as

host. Gradually, the quantization step is increased

and thus the number of histogram bins is decreased.

In every case, the figures include the Peak Signal

to noise ratio (PSNR) for exploitation of the

maximum number of channels that capacity enables,

as seen on Table 1. Input MREGs for the PSNR

function were down-sampled to 8 bits. Concerning

the current test set and being restricted to 64 EEG

channels, PSNR was 51.25 dB using the entire

MREG as host and 256 bins and 30.42 dB using the

segmented ROI as host and 8 histogram bins.

Figure 6 demonstrates the result of embedding 64

EEG channels in 256 bins at the entire image, while

Figure 7 shows another case of embedding 10 EEG

channels using the ROI and the maximum available

capacity when 64 bins are produced. Last, Figure 8

depicts an extreme case of embedding 64 EEG

channels in 8 bins of the ROI.

HEALTHINF 2016 - 9th International Conference on Health Informatics

Table 2: MREG’s fidelity.

Host area, # bins

PSNR (dB) for max. # channels

Entire, 256 53.34

ROI, 256 78.34

ROI, 128 48.94

ROI, 64 42.79

ROI, 32 36.82

ROI, 16 32.92

ROI, 8 30.54

Figure 6: Original (left) and marked (right) MREG

intersection (64 EEG channels, Entire MREG, 256 bins,

PSNR: 51.25 dB).

Figure 7: Original (left) and marked (right) MREG

intersection (10 EEG channels, MREG’s ROI, 64 bins,

PSNR: 42.79 dB).

Figure 8: Original (left) and marked (right) MREG

intersection (64 EEG channels, MREG’s ROI, 8 bins,

PSNR: 30.42 dB).

4 DISCUSSION

Nowadays, it is very common for medical data to be

stored in clouds and to be transferred between

databases through open hospital networks. A typical

example of such an application can be seen in

biobanking. Biobanks are electronic repositories

where big data collections with relevant consents of

the donors are stored and furthermore made

available to research through common availability

services.

In biobanking applications, efficiency in using

the vast amounts of increasing data, as well as the

fact that there might always be disputes over

originality and authenticity of data creates big issues

(Ruzzo, 2014). Furthermore, it is expected that a

great amount of development in applications and

research related to information collected in biobanks

will take place in the near future. All things

concerned provide a solid motivation towards the

research in data hiding with applications in

biomedical databases.

The main novelty of the paper is the introduction

of a reversible data hiding method for a

multidimensional data sequence that is the MREG.

To our knowledge, this is the first paper to feature

the temporal synchronization of EEG and MREG

recordings. Past literature focuses on static

biomedical images; a thorough review of which is

presented by Pan et al. (2009).

In the introduction, the paper points out

numerous benefits of utilizing the data hiding

method in EEG-MREG applications. It is

demonstrated that, in general, packaging MREG and

EEG in one entity is very significant for efficiency

in data management and storage. Moreover, it

should be pointed out that this is also beneficial for

the visualization of MREG sequences combined

with EEG data. The existence of temporally

synchronized data in one package enables the

potential for significant visualization improvements

as there can be instant view of the corresponding

EEG for a given MREG sequence segment over

time. A user can select an MREG segment to view

or analyse, and then by the use of the data hiding

method, the user may also efficiently acquire an

output of the reversed MREG data accompanied

with the temporally equivalent EEG signal’s

segment. So far, this is a process that has been done

manually, and thus our method demonstrates data

hiding’s benefits on data management efficiency in

visualisations.

In every experiment, the embedded data was 16-

bit EEG data of 64 channels and 160 samples per

second but properties can vary. For different cases,

capacity figures can be easily approximated thanks

to the column that shows the capacity in number of

bits in Table 1.

The proposed data hiding method is reversible,

Temporally Synchronized Reversible Data Hiding of EEG to MREG

so the original data can always be retrieved.

However, with the tuneable quantization stepsize

parameter, the fidelity of the MREG sequences

carrying the EEG signal can be controlled. Best

quality figures are attained with low down-sampling.

ROI segmentation reduces capacity so different

quantization stepsizes were tested finding out that

fair fidelity is maintained with the use of 64

histogram bins in embedding or higher. In those

cases, PSNR was maintained over 40 dB which is

generally considered a threshold when it comes to

imperceptibility.

There are certain cases where the image requires

absolute fidelity which practically means that it

should be visually identical to the original MREG

throughout all its phases of use. This also concerns

usage before data is extracted and reversed for the

analyses’ purposes. For instance in applications

requiring preview of the MREG, it is important that

higher quantization is avoided, and thus ROI

segmentation should preferably not be considered.

Otherwise visible artefacts which can be disturbing

might appear on the MREG. Alternatively, EEG

compression can be a solution for this problem

because lower quantization can be enough for the

required capacity.

5 CONCLUSIONS

This paper presented a method for hiding EEG or

other physiological signals into MREG with a main

purpose of providing efficiency in data management

and storage. Furthermore, the paper addressed

security issues, i.e. confidentiality, availability and

reliability of content. Tamper proofing capabilities

are additionally provided as small alterations on the

host image affect hidden data and thus illegitimate

extracted data or digital signatures imply data

tampering. The data hiding method can guarantee

with the proper quantization settings high fidelity

between the original MREG sequence and its

version that carries hidden data. Moreover

reversibility is available. Along extraction of data,

the method reverses the MREG that carries hidden

data to its original state. Last, temporal

synchronization between EEG and MREG data is

always maintained.

In future work, we will develop methods for

increasing data hiding capacity. The purpose is to

combine more data modalities hosted in medical

data.

REFERENCES

Collberg, C., and Nagra, J., 2010. Surreptitious software.

Addison-Wesley, Upper Saddle River, NJ, 1

edition.

Coatrieux, G., Lecornu, L., Sankur, B., and Roux, C.,

2006. A review of image watermarking applications in

healthcare, In Engineering in Medicine and Biology

Society, 28th Annual International Conference of the

IEEE, pages 4691–4694.

Cox, I., Miller, M., Bloom, J., Fridrich, J., and Kalker, T.,

2007. Digital watermarking and steganography.

Morgan Kaufmann. Burlington, MA, 2

edition.

Czisch, M., Wehrle, R., Kaufmann, C., Wetter, T. C.,

Holsboer, F., Pollmacher, T., and Auer, D. P., 2004.

Functional MRI during sleep: BOLD signal decreases

and their electrophysiological correlates. In European

Journal of Neuroscience, volume 20, pages 566–574.

Dharwadkar, N. V., Amberker, B., Panchannavar, S., and

Panchannavar, P., 2010. Reversible fragile medical

image watermarking with zero distortion. In Int’l

Conference in Computer and Communication

Technology, IEEE, pages 248–254.

Fallahpour, M., Megias, D., and Ghanbari, M., 2011.

Reversible and highcapacity data hiding in medical

images. In IET Image Processing, 5(2):190–197.

Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff, J.

M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody

G. B., Peng C. K., and Stanley, H. E., 2000.

Physiobank, physiotoolkit, and physionet components

of a new research resource for complex physiologic

signals. In Circulation, 101(23):215–220.

Gonzalez, R. C., Woods, R. E., 2002. Digital image

processing, Prentice-Hall. Upper Saddle River, NJ, 3

edition.

Grover, D., 1997. The protection of computer software -

its technology and applications, Cambridge University

Press. NY.

Jacobs, J., Stich, J., Zahneisen, B., Assländer, J.,

Ramantani, G., Schulze-Bonhage A., Korinthenberg,

R., Hennig, J., and LeVan, P., 2014. Fast fMRI

provides high statistical power in the analysis of

epileptic networks. in Neuroimage, volume 88, pages

282–294.

Jenkinson, M., Pechaud, M., and Smith. S., 2005. BET2:

MR-based estimation of brain, skull and scalp

surfaces. In Eleventh Annual Meeting of the

Organization for Human Brain Mapping, volume 17.

Makeig, S., Westerfield, M., Jung, T. P., Enghoff, S.,

Townsend, J., Courchesne, E., and Sejnowski, T. J.,

2002. Dynamic brain sources of visual evoked

responses. In Science, volume 295, pages 690–694.

Menon, V., and Crottaz-Herbette, S., 2005. Combined

EEG and fMRI studies of human brain function. In

International Review of Neurobiology, volume 66,

pages 291–321.

National Institute of Standards and Technology, 2014.

Descriptions of SHA-256, SHA-384, and SHA-512.

csrc.nist.gov/groups/STM/cavp/documents/shs/sha256

-384-512.pdf.

HEALTHINF 2016 - 9th International Conference on Health Informatics

Ni, Z., Shi, Y. Q., Ansari, N. and Shu, W., 2006.

Reversible data hiding. In IEEE Transactions on

Circuits and Systems for Video Technology,

16(3):354-362.

Pan, W., Coatrieux, G., Montagner, J., Cuppens, N.,

Cuppens, F., and Roux, C., 2009. Comparison of some

reversible watermarking methods in application to

medical images, In 31st Annual International IEEE

Conference on Engineering in Medicine and Biology

Society, pages 2172–2175.

Rajna, Z., Kananen, J., Keskinarkaus, A., Seppänen, T.,

and Kiviniemi, V., 2015. Detection of short-term

activity avalanches in human brain default mode

network with ultrafast MR encephalography, In

Frontiers in Human Neuroscience, 9(448).

Ruzzo, A., 2014. Biobanking's big problem: how do you

manage all that data? blog.fisherbioservices.com/bid

/376102/Biobanking-s-Big-Problem-How-do-you-

manage-all-that-data.

Sriraam, N., 2012. A high-performance lossless

compression scheme for EEG signals using wavelet

transform and neural network predictors. In

International Journal of Telemedicine and

Applications, 2012(5): 1–8.

Tsai, P., Hu, Y. C., and Yeh, H. L., 2009. Reversible

image hiding scheme using predictive coding and

histogram shifting. In Signal Processing, 89(6):1129–

1143.

Wongsawat, Y., Oraintara, S., and Rao, K. R., 2006.

Integer sub-optimal Karhunen-Loeve transform for

multichannel lossless EEG compression. In 14

European Signal Processing Conference, IEEE, pages

1–6.

Temporally Synchronized Reversible Data Hiding of EEG to MREG