3D Data Hiding for Enhancement and Indexation on

Multimedia Medical Data

N. Tournier

1,2

, G. Subsol

, W. Puech

and J-P. Pedeboy

Universit

e de Montpellier 2, LIRMM, CNRS

161 rue Ada, Montpellier, France

STRATEGIES S.A., 41-43 rue de Villeneuve

Parc des Affaires SILIC - BP 80429, Rungis, France

Abstract. In medical applications, large quantities of multimedia data are ex-

changed such as 3D data acquired either by volume (CT-scan) or surface (laser

range) scanning. The increasing of the numerical data using raises some unsolved

issues. As for us, we are interested in the protection and the enhancement of mul-

timedia content by insertion of hidden message. According to Koller et al. [7]

some of these challenges are:

– Metadata embedding;

– Indexing and searching in database.

Data hiding may be a solution for these main applications in the medical domain.

It is possible to embed metadata, with security for conﬁdential data or for index-

ing area in a media, without increasing the size of the ﬁle.

1 Introduction

The main interest of data hiding is to embed additional information without increasing

the ﬁle size and keeping the compatibility with the norms and the standards such as

JPEG for image, DICOM for medical image, MPEG for video, etc.

In this Section, we propose to introduce brieﬂy the data hiding principle and we

present some medical application.

1.1 Data Hiding Principle

The principle is to embed a message into the useful data of the multimedia ﬁle and to

be able to extract or recognize it from the watermarked ﬁle. We can divide the insertion

process in two parts: the synchronization and the embedding.

Synchronization allows us to know where the encoded message can be embedded

in the host signal. It can be considered as a preparation of the media to the embed-

ding. The host signal can be represented in other domain, such as a frequential domain

(Discrete Cosines Transform, Fourier Transform), a wavelet domain (Discrete Wavelet

Transform) as a function of the application and the chosen method. The aim is to be

able to recognize the same subspace, called insertion subspace, before and after the

watermarking.

Tournier N., Subsol G., Puech W. and Pedeboy J..

3D Data Hiding for Enhancement and Indexation on Multimedia Medical Data.

DOI: 10.5220/0003311800430051

In Proceedings of the 2nd International Workshop on Medical Image Analysis and Description for Diagnosis Systems (MIAD-2011), pages 43-51

ISBN: 978-989-8425-38-6

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Fig. 1. Watermarking Scheme.

The message is encoded as a function of a secret key and the host signal. The em-

bedding consists in merging the encoded message and the host signal in the subspace

domain. The embedding principle is illustrated on Fig. 1.

On the watermarked data, we need to extract the embedded message. For the extrac-

tion process, ﬁrst we synchronize the watermarked data to ﬁnd the insertion subspace

again and we extract the message depending on the insertion method. The extraction

principle is illustrated on Fig. 2.

Fig. 2. Extraction Scheme.

1.2 Data Hiding Applications in Healthcare

The applications are various and depend on compromises between the capacity of the

message to embed, the insertion robustness, the modiﬁcation imperceptibility [11], the

algorithm security [6] and the complexity. As the data are voluminous, the algorithms

must be fast with a low complexity.

One of the main application of data hiding is communication, a transmission of a

secret message. It looks like cryptography approaches. In data hiding we do not want to

secure the media to transmit, the media is a carrier for the message we want to send. In

this condition, the watermark algorithm must be robust and secured. For example, in a

medical context, it could be interesting to embed the identity of the patient into the 3D

image acquired by a CT-scan. For privacy, we need to keep anonymous its identity for

a secured transmission of the data. We can note cryptography and data hiding can be

combined for these applications [12].

Data hiding can be used for enhance the media with metadata. It is quite easy to add

metadata in a ﬁle, but we do not want to increase the size of the media because more

the data are voluminous; more it is difﬁcult to store them. Such as the embedding of

indexing information [3] in order to identify the media. It can be also possible to carry a

diagnostic resume, a description of the image, some quotes or remarks in order to have

all the main information available on the same ﬁle. But there are some constraints for

example in ROI medical image, where the image must not be disrupt in the ROI areas

[14].

For further information on data hiding application in healthcare, [5] made a review

of image watermarking applications.

Data hiding is useful in medical application. Nowadays, there are a lot of well-

known techniques in 2D watermarking. In the 3D domain, the topic is more difﬁcult

because the synchronization issue is more difﬁcult. There is not any way to scan the 3D

object obviously rather than in 1D scanning the audio with a pseudo-random clock; or

in 2D scanning the image line by line for example.

In the following Section, we propose data hiding system for 3D objects. First, we

introduce a watermark technique on 3D images acquired by CT-scan in Section 2 and in

Section 3 we deal with a watermark technique applied on 3D meshes. The presentation

of the techniques is validated by experimental results in the respective sections.

2 3D Image Data Hiding

2.1 2D Data Hiding on CT-scan before 3D Reconstruction

For this approach, the idea is to use well-known 2D watermarking technique for 3D

object. First, the object is acquired by volume (CT-scan) in order to have a 3D image.

Each slice is equivalent to a 2D image that we can watermark by classical techniques.

We illustrate on Fig. 3 some slices of a 3D image acquired by CT-scan.

From the 3D image we are able to build a 3D mesh by computing the iso-surface.

We illustrate the reconstruction of the skin (Fig. 4.a.) and the skeleton (Fig. 4.b.) from

our 3D image. The aim is to watermark the 3D image such as there is no perceptual

difference on the 3D object.

In our condition, we have 3D images

composed by 249 slices (512 × 512 pixels).

We can watermark each slice with the same message, or independently with a different

message for each one. It depends on the ﬁnal application. For example, if we prefer to

have a robust watermarking scheme, we will watermark the same message in each slice;

else if we want to enhance the media with metadata, we prefer to have an high capacity

watermarking scheme so we embed different message in each slice.

In that case, we just want to prove that the watermark keep a good quality on the 3D

reconstruction. We choose a well-known technique, by substitution of the LSB value of

the pixels [2].

When the images are watermarked, we reconstruct the object from the 3D image

into a mesh [8].

http://pubimage.hcuge.ch:8080/.

(a) (b)

Fig. 3. An example of some 3D image slices: a) A represention of the slice #1, b. the slice #130,

c. the slice #200 and d. the slice #230.

(a) (b)

Fig. 4. An example of 3D mesh reconstruction from the 3D image: a) The skin reconstruction

composed by 663

072 vertices and 1

126

466 faces; b) The skeleton one composed by 804

073

vertices and 1

577

853 faces.

For the extraction procedure, we just need to extract the message from the 3D image

by reading the LSB value of each watermarked pixel. It is a fragile technique with high

capacity, fast (low complexity) and quite efﬁcient to enhance 3D images with metadata.

2.2 Experimental Results

From the 3D image composed by 249 slices (512 × 512 pixels) we embed in each slice

30ko. With the LSB watermarking method, associated with a pseudo-random image

scanning; we can embed one bit per pixel. In these conditions, the capacity is ﬁxed at

0.9265 bit per pixel.

First, we are looking for the quality of the watermark on each image. We compute

the PSNR (Peak Signal Noise Rate) between the original slice and the watermark one.

The results are very good, the PSNR equals to 51.63 dB in average for an high ca-

pacity application. There is no perceptual difference between a slice and the respective

watermarked one (Fig. 5).

(a) (b)

(c)

Fig. 5. Comparison between: a) The original slice #230; b) The watermarked image; c) The

difference image.

The main aim is to know what is the impact of the watermark on the 3D mesh

reconstruction. We build the 3D mesh, by computing iso-surface on the watermarked

3D image and we compare the quality of the mesh obtained by the original image and

this one.

To quantify the quality we use the Hausdorff distance, that computes the longest dis-

tance of the shortest distance between a vertex in the ﬁrst mesh to a vertex in the second

one. For the skeleton mesh and for the skin mesh, the Hausdorff distance equals to 0.00.

That means there is no difference between these two meshes and we are able to extract

the watermark from the 3D image. It is a very high capacity technique, nevertheless we

can not extract the message from the 3D mesh.

3 3D Mesh Data Hiding

3.1 3D Data Hiding using Euclidean Minimum Spanning Tree

In Section 2, we have presented a 2D watermark technique used for 3D objects. We

remark the extraction process must be done from the 3D image and not from the 3D

mesh. In this Section, we want to watermark the mesh. Various techniques exist, most

of them are robust to geometrical rigid transformations. In this Section, we present an

original 3D watermarking approach that does not move any vertex [1].

(a) (b)

Fig. 6. An example of a mesh: a) The EMST of the mesh; b) We can remark connections in the

EMST are not necessarily edges in the mesh.

Few 3D watermarking approaches do not move any vertex [1], [10], [9]. The origi-

nality of [1] is to watermark a message by editing the connections in the mesh without

increasing the ﬁle size and without moving the vertices. In order to synchronize the

message they compute an Euclidean minimum spanning tree (EMST) illustrated Fig. 6.

The EMST is unique, and from a seed vertex v

, the path of the vertices is also unique.

This can be an interesting synchronization tool because we can cover the mesh with a

unique path.

In the EMST, they are searching quadruples, a father-vertex and three sons, to em-

bed one bit. The insertion is quite simple, if we want to embed a 0-bit the edge of the

EMST is chosen. Otherwise, if we want to embed a 1-bit, the other edge is in the EMST.

The insertion process is illustrated Fig. 7.

In order to not create important visual distortions and synchronizations issues, the

quadruples must verify the following conditions:

– Coplanarity: the measure of the angle formed by the two triangles (Fig. 8);

– Convexity: the quadruple must be convex in order to cover quite the same surface

(Fig. 9);

– Covering: if two quadruples are neighbours only one of them is used for the em-

bedding.

(a) (b) (c)

Fig. 7. Illustration of the EMST (a), the mesh when a 0-bit is embedded (b) and when a 1-bit (c).

Fig. 8. Representation of the coplanarity condition.

(a) (b)

Fig. 9. Representation of the watermarking impact on a quadruple not convex before the water-

mark (a) and after (b). The covered area is different and there is a topological problem: v

must

not be inside of the triangle (v

, v

3.2 Experimental Results

We present some results on our 3D Skeleton. We illustrate, Fig. 10.a the simpliﬁed

3D mesh and its EMST Fig. 10.b. During the watermarking process, the vertices are

not moving, the cloud is the same, only the connectivity is changed. Indeed, there is

not any distortion due to the vertex displacement. So we compute the Hausdorff error,

including the connectivity distortion. The Hausdorff distance is computed on the center

of the triangles and not on the vertices and normalised as a function of the bounding

box diagonal computed with METRO [4].

As a function of the coplanarity threshold, we can embed more information in the

mesh. More the threshold is low, more the connectivity distortion is better but the capac-

(a) (b)

Fig. 10. An example of EMST applied on a simpliﬁed 3D mesh Skeleton (48

711 vertices and

614 faces).

ity maximum decrease because we select fewer quadruples. For a coplanarity threshold

ﬁxed at 10

◦

we can embed a message of 575 bits maximum in Skeleton and for 5

◦

the

maximum capacity is up to 264 bits.

In order to compare the distortion in the mesh, we ﬁx the capacity of insertion to

4 · 10−3 bits per vertex (i.e. 216 bits on Skeleton). After the embedding we compute

the Hausdorff distortion. For a coplanarity thresholds equals to 10

◦

we get 3.94 · 10

−4

and 5

◦

we have 1.08 · 10

−4

. That means for a diagonal of 20 cm, the maximum error is

0.08 mm.

4 Conclusions

We have presented a watermarking overview for medical application, in order to protect

the 3D images or to enhance the 3D object by embedding metadata. In medical appli-

cation the data are large for each patient, so we need to develop fast algorithms without

increasing the size of the data.

We have introduced 2D watermarking techniques used for 3D images. It is a low

complexity algorithm with high complexity to enhance the 3D model. But it is very

fragile to noise addition and it is impossible to extract the message from the 3D mesh.

Then, we have presented a 3D watermark method based on a mesh. The originality

is to watermark a message without moving any vertices and without increasing the ﬁle

size. The synchronization is made by the localization of quadruples in the EMST of the

vertex cloud. The method is also fragile with low capacity of insertion, but the 3D mesh

is watermarked. Nevertheless, the complexity is very high (quadratic as a function of

the number of vertices).

It is very fragile to noise addition, so we focus on this method to improve it by

studying the possible displacement of the vertex in order to be more robust. As a func-

tion of the Prim’s algorithm and the seed vertex, we are able to know how can a vertex

be moved without editing the connections in the EMST [13]. Indeed, we watermark our

message in the most robust areas.

We study the EMST structure in order to ﬁnd new robust criterion in order to select

robust areas and improve the computational time.

References

1. P. Amat, W. Puech, S. Druon, and J.P. Pedeboy. Lossless 3D Steganography Based on MST

and Connectivity Modiﬁcation. Signal processing. Image communication, 25(6):400–412,

2010.

2. C.K. Chan and LM Cheng. Hiding Data in Images by Simple LSB Substitution. Pattern

Recognition, 37(3):469–474, 2004.

3. S. Cheng, Q. Wu, and K.R. Castleman. Non-Ubiquitous Digital Watermarking for Record

Indexing and Integrity Protection of Medical Images. In Image Processing, 2005. ICIP 2005.

IEEE International Conference on, volume 2. IEEE, 2005.

4. P. Cignoni, C. Rocchini, and R. Scopigno. Metro: measuring error on simpliﬁed surfaces. In

Computer Graphics Forum, volume 17, pages 167–174. Wiley Online Library, 1998.

5. G. Coatrieux, L. Lecornu, B. Sankur, and C. Roux. A Review of Image Watermarking Ap-

plications in Healthcare. In Proc. of the 28th International Conference IEEE Engineering in

Medicine and Biology Society, 2006.

6. A. Kerckhoffs. La Cryptographie Militaire. Journal des sciences militaires, 9(5-38):161–191,

1883.

7. D. Koller, B. Frischer, and G. Humphreys. Research Challenges for Digital Archives of 3D

Cultural Heritage Models. Journal on Computing and Cultural Heritage (JOCCH), 2(3):1–

17, 2009.

8. W. E. Lorensen and H. E. Cline. Marching Cubes: A High Resolution 3D Surface Construc-

tion Algorithm. In Proceedings of the 14th annual conference on Computer graphics and

interactive techniques, page 169. ACM, 1987.

9. X. Mao, M. Shiba, and A. Imamiya. Watermarking 3D Geometric Models Through Triangle

Subdivision. In Proc. SPIE, volume 4314, pages 253–260, 2001.

10. R. Ohbuchi, H. Masuda, and M. Aono. Watermaking Three-Dimensional Polygonal Models.

In Proceedings of the Fifth ACM International Conference on Multimedia, pages 261–272.

ACM New York, NY, USA, 1997.

11. B. Planitz and A. Maeder. Medical Image Watermarking: A Study on Image Degradation.

In Proceedings of the Workshop on Digital Image Computing: Technqiues and Applications,

pages 3–8, 2005.

12. W. Puech and J. M. Rodrigues. A New Crypto-Watermarking Method for Medical Images

Safe Transfer. In Proc. of the 12th European Signal Processing Conference, pages 1481–

1484, 2004.

13. N. Tournier, W. Puech, G. Subsol, and J-P. Pedeboy. Sensitivity Analysis of Euclidean Min-

imum Spanning Tree. In SPIE’10: 3D Image Processing (3DIP) and Applications, San Jose,

USA (CA), 2010.

14. A. Wakatani. Digital Watermarking for ROI Medical Images by Using Compressed Sig-

nature Image. In System Sciences, 2002. HICSS. Proceedings of the 35th Annual Hawaii

International Conference on, pages 2043–2048. IEEE, 2002.