Computer Aided Medical Image Diagnosis Enabling Electrophysiological Simulation Studies

Tomi Heinonen^ab, Prasun Dastidar^c

^aRagnar Granit Institute, Tampere University of Technology, Tampere, Finland
^bNokia Corporation, NVO, Tampere, Finland
^cTampere University Hospital, Department of Diagnostic Radiology, Tampere, Finland

Correspondence: T Heinonen, Ragnar Granit Institute, Tampere University of Technology
P.O. Box 692, FIN-33101 Tampere, Finland.
E-mail: tomi.heinonen@nokia.com, phone +358 50 3738 611, fax +358 3 247 4013

Abstract. Due to increasing number of PACS installations and efficient computing environments the popularity of computer-aided diagnosis is increasing rapidly. There are numerous applications available for evaluating different disease signs. Many PACS manufacturers develop multipurpose PACS workstation software enabling interactive image displaying and analysis tools. In addition to obtaining useful quantitative information for the analyzed images, the resulting data can be applied in electrophysiological studies, i.e., segmented anatomical structures and lesions can be applied in ECG/EEG simulations. Presenting the results of analyses requires efficient image processing. The purpose of this research project was to map appropriate image analysis features, and to develop a group of software components aiming to complete image diagnosis environment. Several software components were developed by our research team and applied in clinical research projects. Our findings indicate that computer aided diagnosis is an important tool for future radiological and clinical work improving the quality of diagnosis. Furthermore, the applicability of medical image processing in the field of electrophysiological simulations is essential in order to obtain new research material.

Keywords: Radiology; Computer Aided Diagnosis; Image Processing; Segmentation; Volumetry; 3D; PACS; DICOM

1. Introduction

The popularity of digital imaging devices and PACS installations has increased during the last decade [Reiner et al., 2000; Greinacher et al., 1990]. This is partially due to savings in costs but also advantages in better image quality, physical storage size, tele-radiological possibilities, interactive consultation, and the speed in sending/receiving images. Nevertheless, images are often diagnosed using conventional techniques: Films are studied on an illuminator board, visual segmentation and classification are carried out by the radiologist, lesion and organ diameters are measured using a ruler, cross-sectional areas are approximated using grids, and volumes are approximated using various formulas together with e.g., diameters. Furthermore, 3D nature of section images (e.g., MRI and CT) is applied only by looking the images slice by slice. When compared to digital image analysis, conventional techniques suffer in accuracy, interactivity, and quantitative possibilities. Digital images can be presented on a computer display and modified in real time. Image enhancement (IE), image restoration (IR) and feature extraction can be carried out. Anatomical structures can be segmented and the resulting information utilized in volumetry and weight calculation. In addition, cross-sectional images, such as MR and CT can be presented in 3D. Even though digital image analysis appears to be superior, conventional techniques still have several advantages. The image resolution on the films is high and several films can be studied simultaneously. At present, the size and resolution of digital displays is reduced hence several displays are required in order to study more than one digital X-ray image at a time.

The main reasons why digital diagnosis has not become yet popular are mainly due to technical limitations with computer displays / high price, the lack of appropriate teaching in medical schools, and the lack of standardization. Furthermore, the older generations of radiologists are not willing to change their habits and techniques. Nevertheless, digital diagnosis will become popular during the next decade as PACS installations have during the last decade. Another possibility of digital diagnosis is the utilization of analyzed images in visualization and electrophysiological studies. For example, electroencephalographic signals of an epileptic patient can be combined with segmented MR images. Based on electrical conductivities of segmented tissues, mathematical simulations can be applied in locating epileptic focus. In addition to this inverse problem [Laarne et al., 2000] simulation, it is possible to carry out forward problem simulations aiming to calculate electric fields caused by an artificial electric source.

The objective of this study was to evaluate, estimate, and implement appropriate image processing tools aiming to complete image diagnosis environment. Special attention was paid to features applicable to electrophysiological simulations. However, simulations and their usefulness were not studied.

2. Material and Methods

In order to develop useful and clinically functional image diagnosis tools our research team carried out several studies in co-operation with numerous clinicians representing expertise in radiology, neurology, surgery, otorhinolaryngology, gynecology, and clinical neurophysiology (Tampere University Hospital, Tampere, Finland). The aim was to map essential image processing tools in image diagnosis, and in addition, to find out what requirements are set for the user interfaces and general functionality. In the results section, we describe the findings based on our research studies.

Based on the co-operation, our research team developed several software prototypes [Heinonen et al., 1998; Heinonen et al., 1998^b; Heinonen et al., 1999] and user interfaces which were applied in clinical work and research projects associated with several disorders. All software was developed in Windows NT environment using C++ language. In addition to general image diagnosis, our group participated in a study developing electrical simulation methods for the hart, brain and stomach, applying the developed image processing tools.

3. Results

Based on the co-operation in hospital research projects, the research results can be divided to four classes, which are Hardware requirements, Basic Visualization requirements, Advanced Visualization requirements, and Special Image Processing requirements. In addition to these, we present the advantages of developed software prototypes.

3.1. Hardware Requirements

Because new PC based computers are relatively efficient they can be applied in any image analysis procedures required. The only real hardware based requirements and also limitations are associated with the display quality and graphics adapter. Even though a human can distinguish between 64 gray scales at a time, normal 8bit gray scales are not sufficient (256 gray scales) because a trained radiologist can distinguish between several hundred grayscales by studying the image piecewise. In order to solve this problem, 12bit gray scale graphic cards can be applied. It is also possible to use virtual gray scales with 24bit RGB graphics adapters in order to display more than 256 gray scales at a time. Such techniques appear to produce very realistic gray scales. A larger problem is the resolution requirement because an accurate X-ray resolution can be e.g., 4000 * 4000 pixels. The most accurate general displays are able to present 2048 * 1536 pixels. Special displays are capable of displaying more accurate information, but the price is high. However, it is possible to apply lower resolution by zooming in/out the image and studying it piecewise. In order to study several images simultaneously, more than one monitor and graphics adapters can be used.

3.2. Basic Visualization Requirements

The general requirements for digital image visualization are simplicity, speed, and easiness. Even though the software must be simple, it can include several advanced operations, which are hidden from a normal user. The basic operation requirements are presented in Table 1.

Table 1.    Basic image visualization requirements.

3.3. Advanced Visualization Requirements

In order to utilize more the digital nature of medical images, advanced image processing and visualization techniques can be applied (see Table 2). These tools can be applied in normal radiological diagnosis by selecting appropriate preset windowing functions. In addition, various details on the images can be efficiently emphasized using interactive adaptive histogram equalization; original image and the equalized image are superimposed and the transparency of the equalized image is altered interactively. Such presentation combines the original appearance of the image and is capable of emphasizing small low intensity lesions and details. Another important tool is measurement; lines, circles and polygons can be drawn on top of the images and their dimensions can be calculated based on imaging parameters.

Table 2.    Advanced image visualization requirements.

Segmentation appears to be a key technique in order to enable volumetric analysis, weight calculation, and 3D visualization. In general, the purpose of segmentation is to recognize and classify different tissues, organs and lesions on the image by using computer. Because images consist of voxels with known dimensions, it is possible to calculate volumes and weights of segmented structures. It is also necessary to store segmented images to the image archive in order to estimate e.g., disease/medication progression based on changes in volumes or structures. Segmentation is also a key issue in electrophysiological simulations. Appropriate anatomical models can be reconstructed based on segmentation.

3.4. Special Requirements

Even though segmentation and volumetry are efficient tools in estimating disease/medication progression, the total volume of segmented lesions does not always correlate with clinical findings. This is clearly seen in neurological diseases, such as multiple sclerosis (MS). The poor correlation can be explained on the basis of the locations of lesions; one small lesion can be fatal but another larger lesion does not cause any symptoms. One small lesion located in an important nervous pathway can create grave symptoms but at the same time a very large lesion in not so important pathway gives rise to only minimal symptoms. In order to improve the estimation of real lesion load, anatomical atlases must be applied. Segmented patient images are compared to atlas images in such way that the anatomical/functional locations of the lesions are found out. Utilization of anatomical atlases involves numerous problems, from which the varying anatomy of the human brain is the most difficult one – atlas images must be matched with the patient images.

Table 3.    Special image visualization requirements.

3.5. Test Studies in Clinical Environment

The developed software Anatomatic™ and Medimag™ were applied in numerous patient studies, which are presented in Table 4. For further information for the results on these studies, please refer to the references in the Table.

Table 4.    Test studies in clinical environment.

4. Discussion

According to our studies, 8bit gray scale hardware can be applied efficiently in computer aided diagnosis when the entire image dynamics are utilized in interactive windowing – even though only 256 separate intensities can be presented at a time, region of interest can appear with better contrast when compared to conventional 12-16 bit visualization. However, this type of interactive windowing is not yet approved for general clinical radiology.

Automatic segmentation is not reliable due to varying anatomy hence semiautomatic approach is recommended. However, such approach suffers from human errors causing inter- and intraobserver variability. In addition, some normal structures and pathological lesions have similar kind of appearance, so it is essential to have experienced radiologist performing the segmentation. At present, some semiautomatic segmentation routines require fairly long time therefore the popularity of these techniques has not yet reached daily clinical practice. In long-term longitudinal studies as e.g., MS disease, the use of same imaging device is recommended due to variability in resulting images. Our experience has shown that the measured parameters at the time of imaging and that at the time of operation differ somewhat due to existing different circumstances on the operation table. This leads to difficulties in calibration of the volumetric techniques.

We were able to apply the developed software in numerous different disease studies and also in clinical practice. In addition to the diseases mentioned earlier these software could be found useful in various other diseases, such as gall stone diseases, tumors and inflammatory diseases of the alimentary system, respiratory system, and the peripheral nervous system. Also structures and tumors of the musculoskeletal system can be quantified. Aneurysms and arteriovenous malformations can be volumetrically estimated before and after interventional therapy/surgery.

Volumetric analysis using segmentation has become a routine procedure in phase three drug trials in diseases like multiple sclerosis and brain infarcts. It helps the researchers and experts to analyze their results with more volumetric accuracy. It provides the clinicians with measurement of the end points of drug treatment and after operations. It also provides them with the prognosis and thus acts as a surrogate for many clinical markers. The ease with which these software can be used nowadays promises a future where every clinician will have all these software at his disposal in the polyclinic room.

A decade back when computerized volumetric analysis was not yet introduced to the routine medical analysis, manual volumetric analysis took long tedious hours of hard work. In diseases like MS where there are hundreds of lesions the task seemed impossible. Nowadays with the help of these semi-automatic volumetric software, the analysis of effects of drug and the prognosis has become much more effective. The response to treatment in cancers like in the ovarian tumors has become much more accurate and thus options for different further post-operative and post-radiation treatment has become much more effective.

With the help of the 3D software, the 3D reformations of lesions and body structures have become more realistic and accurate. With the help of the 3D images the effects of different lesions and tumors on the surrounding parts can be determined in all three sagittal, coronal and axial planes and also in different sub planes. For the neurosurgeon the use of 3D images in navigational operations is very helpful and has become routine in many centers.

Segmentation appears to be a key technology in computer aided diagnosis of medical images. One of its applications is the generation of source data for electrophysiological simulations. Voxels labeled with tissue specific conductivities can be applied in calculating electric field distributions. However, very accurate conductivities have not yet been obtained, and in addition, conductivity in certain tissues is unisotropic, hence perfect models cannot be developed.

We believe that in the near future computer-aided diagnosis will become routine tool in radiology and other clinical fields. Before that, some standardization is required in integrating anatomical atlases, image analysis software, simulation tools, and PACS systems. Also teaching and clear change in thinking attitude towards more digital diagnostic world is required. As a conclusion for this study, the following picture illustrates the relations of different medical research components in order to create a complete diagnosis environment.

Acknowledgements

The authors wish to thank the Ragnar Granit Institute and the Ragnar Granit Foundation. Also the Department of Radiology of the Tampere University Hospital is greatly acknowledged.

References

Reiner B, Siegel E, McKay P. Adoption of alternative financing strategies to increase the diffusion of picture archiving and communication systems into the radiology marketplace. Journal of Digital Imaging, 13(2 Suppl 1): 108-13, 2000.

Greinacher CF, Bach EF, Herforth M, Luetke B, Seufert G. Related Articles Computer-assisted radiology--requirements and solutions for digital diagnostic imaging. Medical Informatics, 15(1): 21-9, 1990.

Heinonen T, Dastidar P, Kauppinen P, Malmivuo J, Eskola H. Semi-automatic tool for segmentation and volumetric analysis of medical images. Medical & Biological Engineering & Computing, 36: 291-296, 1998.

Heinonen T, Visala K, Blomqvist M, Eskola H, Frey H. 3D Visualization library for multimodal medical images. Computerized Medical Imaging and Graphics, 22(4): 267-273, 1998^b.

Heinonen T, Lahtinen A, Häkkinen V. Implementation of Three-Dimensional EEG Brain mapping. Computers and Biomedical Research, 32(2): 123-131, 1999.

Dastidar P, Heinonen T, Vahvelainen T, Elovaara I, Eskola H. Computerized volumetric analysis of lesions in multiple sclerosis using a new semiautomatic segmentation software. Medical & Biological Engineering & Computing, 37: 104-107, 1999.

Dastidar P, Heinonen T, Ahonen JP, Jehkonen M, Molnar G. Volumetric measurements of right cerebral hemisphere infarction: use of a semiautomatic MRI segmentation technique. Computers in Biology and Medicine. 30(1): 41-54, 2000.

Dastidar P, Kulkas T, Heinonen T, Lahtinen A, Ryymin P, Frey H. MR volumetry and digital neuroanatomic mapping in vascular dementia. Abstracts of the XVI World Congress of Neurology, Buenos Aires, Argentina, September 14-19, 1997. In: Journal of the Neurological Sciences, Supplement to vol 150. Elsevier. S1-S367, p. S328, 1997.

Dastidar P, Heinonen T, Virta T, Kuurne T. Use of semiautomatic segmentation and three-dimensional reformations in the evaluation of intracranial tumors. Medical & Biological Engineering & Computing, 37 (suppl 1): 268-269, 1999^b.

Dastidar P, Numminen J, Heinonen T, Ryymin P, Rautiainen M, Laasonen E. Nasal airway volumetric measurement using segmented HRCT images and acoustic rhinometry. American Journal of Rhinology, 13(2):97-103, 1999^c.

Dastidar P, Heinonen T, Numminen J, Höckert A, Rautiainen M. Semiautomatic segmentation of CT images in volumetric estimation of airways in nasal cavity and paranasal sinuses. European Archives of Otorhinolaryngology, 256(4):192-198, 1998.

Dastidar P, Mäenpää J, Heinonen T, Kuoppala T, Van Meer M, Punnonen R, Laasonen E. Magnetic resonance imaging based volume estimation of ovarian tumours: use of a segmentation and 3D reformation software. Computers in Biology and Medicine, 30: 329-340, 2000^b.

Uotila J, Dastidar P, Heinonen T, Ryymin P, Punnonen R, Laasonen E. Magnetic resonance imaging compared to ultrasonography in fetal weight and volume estimation in diabetic and normal pregnancy. Acta Obstetricia et Gynecologica Scandinavica, vol 79: 255-259, 2000.

Laarne PH, Tenhunen-Eskelinen ML, Hyttinen JK, Eskola HJ. Effect of EEG electrode density on dipole localization accuracy using two realistically shaped skull resistivity models.  Brain topography, 12(4): 249-54, 2000.