Imaging of Hemorrhagic Stroke in Magnetic Induction Tomography: An In Vitro Study Yasin Mamatjan Electrical-Electronics Engineering Department, Zirve University, Gaziantep, Turkey
Received 3 December 2013; accepted 16 February 2014
ABSTRACT: Magnetic Induction Tomography (MIT) has the potential of providing an inexpensive medical device for regular screening and monitoring of patients. MIT could be used to detect hemorrhagic stroke, if high measurement accuracy and spatial resolution can be reached to allow significant contrast between normal brain tissues and hemorrhaged areas. However, this can be challenging because spatial resolution in MIT is limited due to the small number of independent measurements, the inverse problems are severely ill-posed, and erroneous data cause large artefacts in reconstructed images and lower the detectability threshold of bleeding. The noise components degrading signal and image quality may be caused by thermal drift and noise from acquisition systems, environment or by body movements. The objective of the article is to empirically investigate hemorrhagic stroke in MIT based on in vitro study and to improve stroke detectability and visibility to help monitoring stroke patients. The following approaches were evaluated: (i) level setting, (ii) improved spatial filtering, (iii) averaging of multiple measurements, (iv) the combinations of these three approaches, and (v) wavelet denoising. They were evaluated with an in vitro phantom resembling a cerebral stroke in a pig brain. The results showed that these approaches enhanced stroke visibility, lowered stroke detectability threshold from 15 ml to 5 ml, and improved the localization of phantom hemorrhages such that their combination produced the best results. These methods may make it easy to the estimation of actual stroke volume, and clinical interpretation, and it can be used to longC 2014 Wiley Periodicals, Inc. V term monitoring of stroke progression. Int J Imaging Syst Technol, 24, 161–166, 2014; Published online in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/ima.22090
Key words: magnetic induction tomography; hemorrhagic stroke; stroke detection; image quality; noise reduction
I. INTRODUCTION Magnetic Induction Tomography (MIT) is a cheap, portable, contactless and non-invasive technique for imaging the conductivity distribution within a body (Griffiths, 2001). MIT systems consist of excitation/sensing hardware systems (in order to induce eddy currents within the body and to detect the resulting magnetic field perturbations), a phantom system and an image reconstruction system.
Correspondence to: Yasin Mamatjan; e-mail:
[email protected]
C 2014 Wiley Periodicals, Inc. V
MIT has been proposed for numerous medical applications including edema, stroke classification/monitoring, and lung imaging (Griffiths, 2001; Zolgharni et al., 2009). It can be used to characterize biological tissue properties through spectroscopy (Scharfetter et al., 2003). MIT is a low resolution imaging technique proposed mainly for patient monitoring as a complementary technique to MRI and CT, which produce high resolution images. However, MRI and CT cannot be used for regular screening or continuous monitoring due to ionizing radiation of CT and high cost of MIR scans. The first experimental multi-channel MIT system was designed by Korjenevsky et al. (2000) for biomedical applications with the system frequency of 20 MHz. Watson et al. (2008) designed a lower frequency (10 MHz) 16-channel MIT system for use in low conductivity samples. A similar annular-array 16-channel MIT system was developed by Vauhkonen et al. (2008) using down-conversion methods with the capability of simultaneous parallel readouts for cerebral stroke applications. Multifrequency MIT systems were designed by Scharfetter et al. (2001) using planar gradiometers and a highresolution phase detector (PD), and by Wee et al. (2008) using a new phase stabilization scheme to reduce phase drift of the system. Inverse problems also have been studied including edge-preserving regularization, level-set, artificial neural networks and total variation (Dorn et al., 2000; Korjenevsky, 2001; Casanova et al., 2004; Merwa et al., 2005; Soleimani et al., 2006). In MIT, it is a challenging task to reconstruct images with high data accuracy and high spatial resolution due to a poor signal to noise ratio (SNR), and limited number of independent measurements, since current MIT systems use limited number of coils (mostly 16 excitation/sensor coils). The SNR is not only determined by random background noise, inherent system noise and severe drifts due to temperature dependent behavior of receive channel, but also by a systematic signal degradation by several reasons in long term monitoring applications. Multifrequency MIT may offer advantages over time-difference MIT in terms of reducing movement artifacts and provide more useful diagnostic information for stroke application (Zolgharni et al., 2010), but the frequency dependence of received signal strength in MIT is a limitation. Time-difference imaging has been investigated for monitoring the progression of hemorrhages, but before-lesion reference data could be available in limited scenarios like neurosurgery and for the prescription of tissue plasminogen
activator (t-PA) for stroke patients to burst a blood clot, where the hemorrhages are expected in around 9% of patients. The dynamic changes caused by system instabilities and unspecific physiological changes i.e. body movement in the subject make it challenging to detect the stroke and to post-process the data properly (G€ursoy et. al., 2009a). Since inverse problems are severely ill-posed in MIT, most reconstructed images contain artifacts due to the corrupted data. Effectively utilizing advanced hardware and modeling errors (G€ursoy et. al., 2009b, 2011), signal/image processing techniques to improve stroke detectability and to provide dynamic adaptation to environmental alterations and system instability could provide the performance required for clinical applications. The objective of the article is to experimentally examine hemorrhagic stroke in MIT based on in vitro study and to improve stroke detectability and visibility to help monitoring stroke patients for a long period of time using practical signal and image processing techniques. The following five noise reduction approaches were proposed to enhance stroke visibility, improve stroke detectability and have localized stroke detection compared to conventional reconstruction. These approaches are (i) level setting, (ii) improved spatial filtering, (iii) averaging of multiple measurements, (iv) the combinations of the previous approaches, and (v) wavelet denoising. These approaches were experimentally evaluated using the MIT system and an in vitro phantom resembling a cerebral stroke in a pig brain phantom. In the result and conclusion sections, reconstructed stroke images are compared and discussed based on original image versus the proposed methods. II. METHODOLOGY A. MIT System Overview. The MIT system comprises electronic front end components, phantom system and a computer based data processing system. The former is used to excite the sample space to be imaged with an alternating magnetic field and then sense changes in the signal resulting from the sample presence (i.e. stroke). The computer system is used to process the sensed data and to generate an image of the sample space representing its conductivity distribution. The MIT measurement system consists of an annular-array 16 channel MIT system and used down-conversion methods with a capability of simultaneous parallel readouts of 16 receiver channels. The MIT system operated at 10 MHz (Vauhkonen et al., 2008). The detailed formulations of the MIT forward model can be found in Maimaitijiang, et al. (2008). If a conductive volume is placed between excitation and detection coils as shown in Figure 1, alternating current at the excitation coil
creates an alternating primary magnetic field B, which induces alternating eddy currents within the sample. The induced eddy currents produce their own secondary magnetic field DB. This causes a secondary signal DV, which is measured at the detection coil (Griffiths, 2001). The eddy currents induced from the second magnetic fields are negligible compared to the currents from the primary induction. The general procedures of creating the secondary signal can be seen from Figure 1 and described as follows: i. An excitation coil with sinusoidal alternating current creates the magnetically-induced field B (Ampere’s law), ii. Magnetically-induced field creates eddy currents in an object with certain conductivity (Faraday’s law), iii. Eddy currents generate the secondary magnetic field DB (Biot-Savart’s law), iv. Receiver coils pick up the secondary magnetic fields in the form of voltages and they are measured.
B. Inverse Problems. A commonly used Tikhonov based regularization method is applied in order to solve the inverse problems as below: Dr5ðJT J1kIÞ21 JT DV
(1)
where Dr is the change of the conductivity, k is the regularization parameter and I is the identity matrix. DV is the normalized differrÞ ence (DV5 ðViV2V ) obtained from the reference data (Vr ) and subser quent "imaging data" (Vi ) obtained based on the bleeding test. J sensitivity matrix (Jacobian), JT is the transpose of J. The regularization parameter k determines the trade-off between minimizing the residual sum of squares and the norm of the estimate. C. MIT Measurement Protocols. In vitro tissue resembling peripheral hemorrhagic stroke was used where 5 ml, 10 ml, and 15 ml of blood in a biomembrane package (pig intestine) to avoid the blood leaked into the brain tissue were placed in a swine brain of 100 ml (inside a bottle with 9.4 cm diameter and 16 cm height) as shown in Figure 2. Heparinized swine blood was used to avoid blood coagulation. The swine brain has the conductivity of 0.32 S/m, while the blood has conductivity of 0.97 S/m. The conductivity of materials was measured using an Agilent 4294A impedance analyzer. The time interval between successive measurements was set to 16 s. A series
Figure 1. A diagram for the eddy currents induced in a conducting volume by B, and the secondary magnetic field, which induces a secondary signal in the receiver coil. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
162
Vol. 24, 161–166 (2014)
let denoising algorithm filters out distortions (high-frequency) due to noise caused by electronics hardware and environmental noise. This article applied the hierarchical denoising approach proposed by Maimaitijiang et al. (2010) for both the measured signals and reconstructed images in order to show very small bleeding (i.e. 5 ml and 10 ml of bleeding volume). Wavelet based de-noising involves selection of suitable mother wavelets, proper threshold for each different application, since it involves procedures influenced by or dependent on the shape of the signals to be extracted from noisy data. The selected wavelet families include Meyer (dmey), Daubeachies (db3) based on the comparison given by Maimaitijiang et al. (2010).
Figure 2. A phantom for in-vitro testing of bleeding in the brain (2D top view). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
of consecutive measurement sequences were taken while a single reference measurement was used. D. Averaging of Measurements. Measured raw data are averaged over 5 to 10 data sequences to improve measurement precision by reducing random electronic noise and improving measurement SNR. This in turn may produce more stable reconstructed images. E. Spatial Filtering. Spatial filtering is used to smoothen reconstructed images by penalizing strong artifacts which often exhibit high spatial frequency while attempting to preserve image information content. This is achieved by applying low-pass masks to each pixel in the image as described in Figure 3. F. Level Setting. Level setting method is used to eliminate random artifacts and adjust the energy level threshold of reconstructed image value to be displayed as an output image. The pixel values are capped using a threshold value that is a ratio of the image’s maximum value and the rest was eliminated by setting them to zero. The ratio value was set to 75% in this article. This value was empirically found to keep the most of the wanted image while reducing the most of the unwanted noise and artifacts. However, it can be tuned depending on applications and noise levels in the MIT hardware. G. Wavelet Denoising. Wavelet approach decomposes signals into approximation (high-scale low-frequency components) and detail coefficients (low-scale high-frequency components), where the approximation gives the signal their identities as the low frequency contents are the most important parts (Daubechies, 1990). The wave-
Figure 3. Pseudo-code for low-pass spatial filtering. f is the filter level and n is the total number of image pixels.
III. RESULTS A. In Vitro Tissue for Stroke Detection. Figure 4 shows the reconstructed images on 15 ml bleeding within a bio-membrane package placed in a 100 ml of pig brain as a detectable bleeding case. Four consecutive measurement sequences were evaluated. The following cases were compared (i) original reconstruction, (ii) level setting, and (iii) level setting plus spatial filtering. It can be seen from Figure 4 that large artifacts appeared in original reconstructed images. Level setting approach filtered out most of the artifacts. A further improvement was achieved by combining level setting with spatial low-pass filtering. This cancelled the remaining small image artifacts that level setting alone could not remove and produced clear reconstructed images. Figure 5 shows the reconstructed images on 15 ml bleeding within a bio-membrane package placed in a 100 ml of pig brain. It shows (i) original reconstruction, (ii) averaging of five measurements, (ii) averaging of 10 measurements, and (iv) averaging of five measurements plus level setting and spatial filtering. All reconstructions were performed with a fixed regularization parameter (k 5 1029). The three original image reconstructions in Figure 5 showed that drift and noise have a great effect on the image quality. The averaging of five measurements and subsequent ten measurements suppressed random artifacts due to the hardware noise and produced more stable reconstructed images as shown in Figure 4. The level setting and spatial low-pass filtering approach further improved the image quality by clearing the artifacts below the defined threshold. Figure 6 shows the reconstructed images based on the wavelet denoising approach for the bleeding volumes of 5 ml and 10 ml in a bio-membrane package placed in the brain, which represent the scenarios below the bleeding detectability threshold of the system. It can be seen from Figure 6 that without de-noising, images were not reconstructed for both 5 ml and 10 ml cases (in biomembrane package) and some large artefacts appeared in wrong position. The wavelet denoising algorithm based on db3 and dmey helped for better image reconstruction by showing both bleeding volumes of 5 ml and 10 ml. IV. DISCUSSION AND CONCLUSIONS This article empirically examined hemorrhagic stroke in MIT based on in-vitro study and evaluated five suitable noise reduction approaches to enhance visibility of stroke images and lower the detectability threshold of stroke volume based on (i) level setting, (ii) improved spatial filtering, (iii) averaging of multiple measurements, (iv) the combinations of the previous approaches, and (v) wavelet denoising. All algorithms of this work were implemented in Matlab. For this article, the random noise from the MIT electronic components was considered to be the source of noise, as the rest of noise are ignored. Although for a typical clinical situation such as
Vol. 24, 161–166 (2014)
163
Figure 4. Reconstructed images based on: (i) original reconstruction (ii) level setting, and (iii) level setting and spatial filtering. (Note: All the reconstruction were performed with the fixed regularization of k 5 1029). [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
hemorrhagic stroke detection, the reference measurement is not usually available, the time-difference measurement can be useful to monitor if the aim is to detect bleeding after brain surgery or after a prescribed clot-busting drugs is administered—typically TPA to the patients with ischemic stroke which may also cause bleeding in around 9% of patients. The MIT systems suffered from the measurement noise and drift. For the first 1.5 h, temperature drift coefficient was measured at 11.7 (millidegree/ C) and averaged phase standard deviation was 26.3 (millidegree) (Wee et al., 2008). The level setting, improved spatial filtering and averaging approaches as shown in Figures 4 and 5 enhanced image quality and reduced noise by preserving the edges and reducing artifacts. Image quality improved with the averaging of multiple measurements on both the reference measurements and the real measurements. The improved spatial filtering approach further enhanced the reconstructed image quality by getting rid of small image artifacts from the reconstructed images while preserving majority image information content. A combination of measurement averaging followed by the level setting and the improved spatial filtering appeared to produce the best results by effectively reducing image artefacts and preserving the edges. All these approaches complement each other and can be integrated into the MIT image reconstruction system. However, these methods were more applicable to lower levels of noise
164
Vol. 24, 161–166 (2014)
and it may not be possible to apply them in the presence of high levels of noise i.e. 5 ml bleeding scenario. However, in-vitro results in Figure 6 demonstrated that the wavelet denoising algorithm helps to reconstruct and visualize very small bleeding (i.e. 5 ml bleeding can be visualized) and also reduced artefacts in large bleeding (i.e. 15 ml bleeding—the results not shown here). Although wavelet has a lot of advantages compared to other methods like simpler filtering approaches, it is much more complex to select suitable mother wavelets and proper threshold values for different applications and systems, since it involves procedures under the influence of or dependent on the shape of the signals to be extracted from noisy data. If the bleeding is strong and above the detectability threshold of the system, most of artifacts can be removed by the improved spatial filtering and level set approaches. The average of multiple measurements increased the total computation time as it takes 16 s to obtain one frame of MIT measurement data with the combination of 16 excitation 3 16 detection channels. However, a fast data acquisition system was achieved by Maimaitijiang et al. (2011), therefore exploiting advanced hardware and parallel algorithms could be an option to achieve a fast MIT system. Wavelet based denoising algorithms with adaptive regularization (Maimaitijiang et al., 2010) were found to be an effective means to lower the detectability threshold and augment image quality
Figure 5. Comparison of reconstructed images based on: (i) original reconstruction, (ii) averaging of five measurements, (ii) averaging of 10 measurements, and (iv) averaging of five measurements plus level setting and spatial filtering. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
while hardware improvement (Wee et al., 2008) and combined multimodality approaches (G€ursoy et al, 2011) were proposed for improving measurement accuracy and image quality respectively. After a certain period of time (about 8 min), the blood was coagulated and was found to have a lower conductivity in in vitro case studies which affected the detectability of MIT. To avoid tissue degeneration in our test for long time monitoring, a phantom tank was filled with a liter of half saline and half agar of the same conductivity (0.2 S/m) and
cooked cubic agar phantom (5 ml, 0.6 S/m) was prepared and placed in the central plane of the tank near the edge. A series of phantom measurements were carried out for a long period of time and reconstructions were performed at certain time intervals as shown in Figure 7. It can be seen from the Figure 7 that the level setting and spatial low-pass filtering approaches helped improve image quality for over 68 min while original reconstruction showed different image artifacts (the flame images) outside the region of interest. The phantom
Figure 6. Reconstructed images of 5 ml (row one) and 10 ml (row two) of blood in a biomembrane package placed in the brain based on (i) without denoising, (ii) db3 wavelet denoising, and (iii) dmey wavelet denoising. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
Vol. 24, 161–166 (2014)
165
Figure 7. Level set and improved spatial filtering assisted monitoring for the phantom measurement. [Color figure can be viewed in the online issue, which is available at wileyonlinelibrary.com.]
measurement with the same methodology (Fig. 7) produced more stable results than the in-vitro tissues resembling cerebral stroke in Figure 4, since the biological tissue property may be affected by a lot of other parameters i.e. tissue degeneration, temperature and tissue coagulation. Measurement noise and image artifacts are a major problem in many patient monitoring especially stroke monitoring because dynamic changes caused by system instabilities, environmental changes and unspecific physiological changes in the subject make it challenging to acquire accurate measurements and to post-process the data properly. The proposed software approaches can be easily combined and integrated to the MIT image reconstruction system. These methods improved image quality and detectability threshold of stroke, and thus may make it easy to estimate actual stroke volume, monitor the progression of stroke and clinical interpretation. ACKNOWLEDGMENTS The author thank to Dr. J. Kahlert from Philips Medical in Germany for his support and fruitful discussions. He had agreed to use the measurement data covered partially in Maimaitijiang et al., 2010. REFERENCES A. Korjenevsky, V. Cherepenin, and S. Sapetsky, Magnetic induction tomography: Experimental realization, Physiol Meas 21 (2000), 89–94. A. V. Korjenevsky, Application of artificial neural networks for static image reconstruction in "Soft Field" tomography, Proc. 3rd EPSRC Engineering Network Meeting on Boimedical Applications of EIT, 4 – 6 April 2001, UCL, London, 2001. D. G€ursoy and H. Scharfetter, Reconstruction artefacts in magnetic induction tomography due to patient’s movement during data acquisition, Physiol Meas 30 (2009), 165–174. D. G€ursoy and H. Scharfetter, The effect of receiver coil orientations on the imaging performance of magnetic induction tomography, Meas Sci Technol 20 (2009), 105505. D. G€ursoy and H. Scharfetter, Imaging artifacts in magnetic induction tomography caused by the structural incorrectness of the sensor model, Meas Sci Technol 22 (2011), 015502. D. G€ursoy, Y. Mamatjan, A. Adler, and H. Scharfetter, Enhancing impedance imaging through multi-modal tomography, IEEE Trans Biomed Eng 58 (2011), 3215–3224. H. Griffiths, Magnetic induction tomography, Meas Sci Technol 12 (2001), 1126–1131. H. Scharfetter, H. K. Lackner, and J. Rosell, MIT: Hardware for multifrequency measurements in biological tissues, Physiol Meas 22 (2001) 131–146.
166
Vol. 24, 161–166 (2014)
H. Scharfetter, R. Casa~ nas, and J. Rosell, Biological tissue characterization by MIT spectroscopy: Requirements and limitations, IEEE TBME 50 (2003), 870–880. H.C. Wee, S. Watson, R. Patz, H. Griffiths, and R.J. Williams, “A Magnetic Induction Tomography system with sub-millidegree phase noise and high long-term stability,” In IFMBE Proceedings of the 4th European Conference, Antwerp, Belgium, 2008, pp. 744–747. I. Daubechies, The wavelet transform, time-frequency localization and signal analysis, IEEE Trans Inform Theory 36 (1990), 961–1005. M. Soleimani, W. R. B. Lionheart, A. J. Peyton, M. Xiandong, and S. R. Higson, A three-dimensional inverse finite-element method applied to experimental eddy-current imaging data, IEEE Trans Magnet 42 (2006), 1560–1567. M. Vauhkonen, M. Hamsch, and C.H. Igney, A measurement system and image reconstruction in magnetic induction tomography, Physiol Meas 29 (2008), 445–454. M. Zolgharni, H. Griffiths, and P. D. Ledger, Frequency-difference MIT imaging of cerebral hemorrhage with a hemispherical coil array: Numerical modeling, Physiol Meas 31 (2010), 111–125. M. Zolgharni, P. D. Ledger, D. W. Armitage, D. S. Holder, and H. Griffiths, Imaging cerebral haemorrhage with magnetic induction tomography: Numerical modelling, Physiol Meas 30 (2009), 187–200. O. Dorn, E. Miller, and C.M. Rappaport, A shape reconstruction method for electromagnetic tomography using adjoint fields and level sets, Inverse Problem 16 (2000), 11–19. R. Casanova, A. Silva, and A. R. Borges, MIT image reconstruction based on edge-preserving regularization, Physio Meas 25 (2004), 195–207. R. Merwa, K. Hollaus, P. Brunner, and H. Scharfetter, Solution of the inverse problem of magnetic induction tomography (MIT), Physiol Meas 26 (2005), 241–250. S. Watson, R. J. Williams, W. Gough, and H. Griffiths, A magnetic induction tomography system for samples with conductivities below 10 Sm21, Meas Sci Technol 19 (2008), 1–11. Y. Maimaitijiang, M.A. Roula, and J. Kahlert, Approaches for improving image quality in magnetic induction tomography, Physiol Meas 31 (2010), 147–156. Y. Maimaitijiang, M.A. Roula, S. Watson, G. Meriadec, K. Sobaihi, and R. J. Williams, Evaluation of parallel accelerators for high performance image reconstruction for magnetic induction tomography, J Selected Areas Software Eng (JSSE) 02 (2011), 1–7. Y. Maimaitijiang, M.A. Roula, S. Watson, R. Patz, R.J. Williams, and H. Griffiths, Parallelization methods for implementation of magnetic induction tomography forward models in symmetric multiprocessor systems, J Parallel Comput 34 (2008), 497–507.