Visualization and Post-Processing of 5D Brain Images - IEEE Xplore

Proceedings of the 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference Shanghai, China, September 1-4, 2005

Visualization and Post-processing of 5D Brain Images Yan Zhang1, Peter J. Passmore1, and Richard H. Bayford 2 1

2

School of Computer Science, Middlesex University, London, UK, N17 8HR School of Health and Social Sciences, Middlesex University, London, UK, EN3 4SA

Abstract—Visualization plays a central role in the presentation and interpretation of medical image data. Radiologists and surgeons must be able to accurately interpret the data for diagnosis and surgical planning. The data obtained from many imaging systems can contain functional as well as structural information producing 4D datasets. In some cases this can extend to 5D when the image provides spectral information. Generally speaking, more information can be revealed in 5D than 4D imaging. Although several approaches are available to visualize 4D medical data, there is limited research on the visualization of 5D medical data. To present 5D medical datasets efficiently on a 2D screen provides considerable challenges to visualization. In this paper, a 5D brain EIT (Electrical Impedance Tomography) dataset is used as a case study. The relationship and differences between multiple dimensional dataset visualization in different areas are analysed. A statistical post–processing method is then adopted to concentrate information included in the fifth dimension. A scheme to visualize 5D medical dataset is proposed and results are shown based on a simulated dataset.

different impedances. EIT imaging exploits this property (i.e. impedance) by injecting a small current through sensors encompassing the area to be imaged. EIT imaging is cheap, safe, and portable. Compared with other functional imaging approaches, EIT has high temporal resolution and poor spatial resolution. Currently, EIT is not in routine biomedical use for any purpose, but studies have applied EIT to assess cardiac function, pulmonary hypertension, regional lung function, brain function, breast cancer, and gastrointestinal tract. The first generation of EIT imaging measures impedance changes over a few seconds or minutes using a current applied at a single frequency of about 50 kHz [2]. Because different tissues have different spectral properties, EIT may also be performed at multiple frequencies at a time. This has the advantages that tissue may be characterized much better. Advanced EIT hardware is able to measure 30 frequencies simultaneously [3]. Using multi-frequency EIT hardware to trace impedance changes over time results in 5D EIT image data: three for space, one for time and one for frequency.

I. INTRODUCTION

II. METHODS

Visualization can play a central role in the presentation and interpretation of medical image data for clinicians. The need for new approaches to image visualization and analysis becomes increasingly important and pressing as improvements in imaging technology enable more complex objects and processes to be imaged and simulated. At present, the evolution of different visualization approaches in medicine can be grouped into five generations: 1D waveform display, 2D image display, 3D dataset visualization, multidimensional dataset visualization and virtual reality type visualization [1]. A multi-dimensional medical image traditionally means a dynamic volume dataset, e.g. fMRI (functional Magnetic Resonance Imaging), PET (Positron Emission Tomography) images, which includes four-dimensional information: three for space and one for time. With the development of medical imaging methods, 5D medical images are produced: one typical example is to include spectral information in the image dataset. To present 5D medical data efficiently on a 2D screen provides considerable challenges to visualization. As a case study, this paper concentrates on 5D brain EIT (Electrical Impedance Tomography) images. However the method presents in this paper could be transferred to other imaging methods. EIT is a relatively new medical imaging method. The physiological basis of EIT is that different tissues have

0-7803-8740-6/05/$20.00 ©2005 IEEE.

A. Visualization of multi-dimensional data In the field of medicine, most of the research on multiple dimensional visualization concerns the Visible Human Project, which is based on the Visible Human Dataset (VHD). VHD is a high resolution, multi-modality, 3D anatomical database of an entire male and female body. Some work has been done to visualize 4D dynamic volume visualization. Few efforts have been made in the visualization of 5D data. From a computing science point of view, multidimensional visualization usually means visualization of multi-dimensional data in databases. It also includes visualization of multi-dimensional flow data, geographical data etc. However few researchers focus their work on the visualization of medical image data in more than four dimensions. B. Characteristics of multi-dimensional medical images To understand the relationship and difference among multi-dimensional data visualization in different areas, we propose to divide the dimension (S) of dataset into two parts: S = m+n (1) In which m dimensions form an m-dimensional space where samples are taken, variables corresponding to these m

1083

dimensions are location variables, which are independent to each other. The other n dimensions are either n dependent variables (dependent on the location variables) measured in the m-dimensional space, or n independent variables if no location variables exist. In a database, each record is a sample of S dimensional dataset. Commonly, these S dimensions are spatially incoherent. So in this case, m is zero, and n equals to S. Several approaches have been developed to visualize this kind of multi-dimensional data, for example: scatter plot, parallel coordinates, star coordinates etc. Multi-modality medical anatomical datasets tend to have three location variables to define a volume space. So in this situation m equals to 3. The other n variables are dependent on the space variables. To visualize this kind of multi dimensional data, volume visualization methods can be used to display the three location variables, other variables can be expressed with different glyphs, e.g., line, arrow, within the volume space. For a 4D or dynamic volume medical dataset, there are four location variables (space + time), and there is another dimension which is the property measure by the medical imaging method, e.g., in fMRI, it is the “BOLD” signal; in PET, it is emitted gamma ray photons; in EIT, this property is impedance. There are two common methods to visualize this kind of dataset. The first way is to display it in a matrix of tomography images. With this method, each tomography image is a two-dimensional slice of the 4D dataset; the two dimensions of the matrix present the other two dimensions of the 4D dataset. Although this method is simple, it tends to be difficult to interpret data in this format. The second approach is to visualize a 3D volume sample on the screen by volume rendering or simply display some multiple section planes (e.g. orthogonal section of the volume) of 3D volume sample; the fourth dimension, which is usually time, can be visualized by animation of the 3D sample. In a 5D EIT dataset, all these five dimensions (space + time + frequency) are location variables, which are used to locate the measured property – impedance. It is possible to visualize this kind of dataset by an animation of a tomography picture matrix. However considering the difficulty of interpreting 4D data displayed in this way, it is likely that clinicians will have even more difficulty with 5D data using this method. On the other hand, for a dataset with more than three location variables, it is always desirable to display three space dimensions together, as they are in the real world, and present other dimensions based on this space. Our visualization scheme for 5D brain EIT images are proposed based on this idea. C. Scheme of visualization for 5D brain images As clinicians usually possess a great deal of background knowledge about human anatomical structure, it should be ideal to visualize the three space dimensions as a volume brain image. The time dimension can be used to create the animation. The main problem in the visualization of 5D

brain EIT dataset is then how to add the frequency dimension effectively into the image. A common preliminary step in the analysis of medical images (for the human brain) is to process images so that areas of clinical interest are segmented out from the whole brain volume. We propose to display these sub-volumes or Regions Of Interest (ROIs) within the brain volume using different colours for different frequencies as a first visualization step. How such ROIs are identified is a matter of much research and depends on the type of data used. The method we adopted is outlined in the next subsection. Fig 1 shows the approach used in our system to visualize the 5D brain dataset.

Fig 1: Illustration of visualization method for 5D brain images

D. Post-processing of 5D brain images Because of the noise introduced during the measurement, the limitation of the reconstruction algorithm, and the physiological changes inside brain, functional images of the brain generally include changes almost everywhere inside the brain. If the clinician’s goal is just the identification of the ROI, then it is reasonable to highlight just this information. The location of ROI in functional brain images (4D) is a research area currently attracting a lot of interest. Several signal and image processing methods have been developed to date. SPM (Statistical Parametric Mapping) is one of the most successful methods in this area. SPM refers to the construction of spatial extended statistical processes to test hypotheses about regionally specific effects [4]. Currently, the SPM method and software are designed only for the analysis of SPECT, PET and fMRI images, which are all 4D dynamic volume medical images. We have previously demonstrated that it is reasonable to analyze 4D (space + time) brain EIT images with SPM [5]. In order to locate ROI in 5D (space + time + frequency) brain EIT dataset, we process every 4D (space + time) EIT dataset at a fixed frequency with SPM. Once the ROI for each frequency is located, it seems reasonable to use different colours to display the ROIs for different frequencies in one brain volume.

1084

Fig 2: Illustration of a small part of the simulated 5D brain EIT dataset. Each small image is a transverse plane across the brain; each row of the pictures presents impedance change by time course; each column shows impedance value at different frequencies measured simultaneously.

III.EXPERIMENTS AND RESULTS A. Introduction to the simulated 5D dataset In brain EIT imaging, functional impedance changes, with a time course of minutes, maybe caused by cells swelling or blood volume and flow increase. In this experiment, we created some datasets to model the conductivity change caused by blood volume and flow increase during visual stimulation, and assumed 12 frequencies are measured simultaneously. The time interval between each sample point is set to be 3 seconds and the whole measurement lasts 3 minutes. Visual stimulus is presented from the end of the first minute to the end of the second minute. The Balloon hemodynamic model [6] is adopted to model the blood volume and flow changes caused by visual stimulus. Impedance change inside the brain is calculated according to the blood volume and flow change and spectral properties of brain tissues. After the calculation of the impedance changes, an EIT reconstruction algorithm is used to produce the EIT image dataset. The size of this simulated dataset is 71×51×69×60×12 (X, Y, Z, Time, Frequency). Coordinates in the dataset is defined as: X increases from left to right, Y increases from posterior to anterior, and Z increases from inferior to superior. Fig 2 illustrates some 2D slices in the simulated 5D brain EIT dataset.

(a)

(b)

Fig 3: Post-processing results of two 4D datasets measured at two different frequencies. In these figures, the original EIT information is presented in greyscale in three orthogonal planes. The ROIs located by SPM are superimposed in colour.

B. Post-processing of the simulated dataset As mentioned above, SPM is used in our system to process 4D datasets for each frequency. SPM provides two statistical models: one for PET/SPECT data, and the other for fMRI data. We adopt the first model to analysis EIT data because it is more flexible. Fig 3 presents the processing results of two 4D dataset with different frequencies. More details on how to uses SPM analysis with 4D EIT data can be found in a previous paper [5].

C. Visualization of the 5D dataset After post-processing of 5D EIT data with SPM, the information revealed at each frequency has been defined as a ROI. Then it is possible to visualize these ROIs in one brain volume. In our prototype system, we use colour to identify the ROIs detected under different frequencies. The proposed 5D visualization system has been developed in VC++ with using VTK (Visualization ToolKit). VTK is a powerful, open-source, object-oriented system for computer graphics, visualization and image processing. In our system, volume rendering is adopted to present main view for radiologist or surgeon. 2D sections are provided to enable the observation of the detail in arbitrary planes. The user can choose to observe one frequency each time, or multiple frequencies together. Animation can be displayed according to time or frequency change. Using colour allows the user to investigate a number of frequencies at the same time, and gives the viewer a simultaneous view of different ROIs. We can further inform this impression by including quantitative measures, for example ROI volume, as shown in Fig 5 below. However having multiple ROIs in a volume could quickly overwhelm the viewer and thus we have to seek further methods to create smaller, higher order ROIs. One generic method of doing this is to implement Boolean spatial operators as used in Constructive Solid Geometry (CSG) such as union, intersection and difference. For example intersection shows areas which are significant at all considered frequencies, while the inverse of the union of all ROIs would identify

1085

Volume (voxels)

areas which are never change significantly for any frequency. Fig 4 shows some images produced from our prototype system including intersection and union of ROIs for two frequencies.

3500 3000 2500 2000 1500 1000 500 0 1

2

3

4

5

6

7

8

9

10

11

12

Frequency number

Fig 5: ROI volumes for different frequencies

IV.

(a)

(c)

(b)

(d)

DISCUSSION AND FUTURE WORK

Unlike other multi-dimensional datasets, medical multidimensional datasets normally have only one dependent variable, the other dimensions are all location variables. These characteristics make some common approaches, which are used to visualize multi-dimensional datasets in other fields, unsuitable for visualization of medical multidimensional dataset. In this paper a statistical postprocessing method is adopted to concentrate information in one dimension. Then a visualization method was proposed based on this processing with the resulting prototype system showing some encouraging results. In future work we propose to further implement and investigate spatial operators, various quantifications and localisation via coregistration with a brain atlas. REFERENCES

(e)

(f)

(g)

(h)

Fig 4: Illustrations of visualization of simulated 5D EIT dataset. Image (a) and (b) show volume rendering of ROIs detected at frequencies 1 and 2 respectively; Images (c) and (d) display ROIs included in (a) and (b) simultaneously from different viewing positions; Image (e) and (f) present the union and intersection of ROIs detected at frequencies 1 and 2; Images (g) and (h) visualize the ROIs of twelve frequencies together from different viewing positions.

[1] Solaiyappan, M. (2000). visualisation pathways in biomedicine. handbook of medical imaging: processing and analysis. i. n. bankman, academic press: 659-685. [2] Brown B H and Seagar A D. (1987) “The Sheffield data collection system” Clin. Phys. Physiol. Meas. 8 Suppl. A 91-97 [3] Wilson A J, Milnes P, Waterworth A R, Smallwood R H and Brown B H. (2001) “MK3.5: a modular, nulti-frequency successor to the Mk3a EIS/EIT system”. Physiol. Meas. 22 4954 [4] Friston KJ, Frith CD, Liddle PF, Frackowiak RSJ (1991): Comparing functional (PET) images: The assessment of significant change. J Cereb Blood Flow Metab 11:690–699. [5] Zhang Y., Passmore P. J., Yerworth R. J., Bayford R.H. (2005).”Statistical analysis for brain EIT images”. Proc. of MediVis05, IEEE Comp. Soc. Press, 60-67. [6] Richard B. Buxton, Eric C. Wong, Lawrence R. Frank (1998).“Dynamics of Blood Flow and Oxygenation Changes During Brain Activation: The Balloon Model”. Magnetic Resonance in Medicine 39, 855-864

1086