Data processing techniques for serial EIT spectroscopy images: a review of some preliminary results

Physiol. Meas. 16 (1995) A129-Al42. Printed in the UK

Data processing techniques for serial EIT spectroscopy images: a review of some preliminary results R H Smallwood and A R Hampshire Department of Medical Physics and Clinical Engineering, University of Sheffield, Royal Hallamshire Hospital, Sheffield SI0 2 F , UK Abstract. Multifrequency EX imaging should allow specific orgam within the body to be identified by their impedance spectrum, and the use of paramelric imaging should lead to a much greater freedom from movement artefacts. This will make ELT more attractive as a monitoring technique, but the data rate will require automated processing of the images. The application of dynamic regions of interest, generated on a frame by frame basis, is described, with examples from the imaging of neonatal lungs and adult stomach. The lung can be objectively identified an a single frame from the f x x . SC and RC i m a m , but the stomach could only be identified on the dynamic images.

1. Introduction The spatial resolution of ELT images is poor. They are only of value if they enable us to localize changes in physiological parameters that are of diagnostic significance. The two areas in which a significant amount of work has been done are the lungs (Harris et a1 1987, 1988, McArdle etal 1988, Smulders and van Oosterom 1992, Brown eral 1994a-c, Nopp etal 1993, Hampshire et al 1995) and the stomach (Avill etal 1987, Mangnall etal 1987, 1988, Baxter et al 1988, M o n t et a1 1988, Evans and Wright 1990, Wright and Evans 1990, Nour etal 1993; Smallwood et al 1993, 1994, Er01 et al 1995, Watson et al 1995). The majority of studies to date have been based on the examination of a relatively small number of dynamic images at a single frequency. In this paper, we will consider some of the key issues in extracting functional data from multifrequency EIT images, illusEate them by examples taken from the monitoring of lung function and gastric function and, finally, discuss some of the implications of these prelimina ;ults. 2. Data collection and preprocessing 2.1. Baric principles A major problem with the use of EIT imaging for patient monitoring is the quantity of image data that is produced. The Sheffield multifrequency EIT system (Brown et a1 1994a) has a hardware speed of 66 frames s-'. At 33 frames s-' (the current maximum acquisition rate), more than 2 Mb of raw data are collected per minute. From each frame of these raw data, an image set consisting of up to 21 images can be produced. This amounts to about 340 Mb of image data per minute, comprising 41 580 images in 1980 frames. Clearly, manual analysis of these quantities of data is not possible, and data reduction and automated analysis is required. The crux of the problem is the characterization of individual pixels. 0967-3334/95/SA0129+14$19.50

0 1995 IOP Publishing Ltd

A129

A130

R H Smallwood and A R Hampshire

The basic principles of handling these large quantities of data are (i) save only the raw data and discard image data after extracting the functional information, (ii) maximize the signal to noise ratio by utilizing all the data collected by the hardware, but sample at the minimum rate to avoid aliasing, (iii) whenever possible, operate in data space rather than image space to minimize computation, (iv) reconstruct only the images or parts of images that are necessary for the measurement and (v) make sure that the processing is robust, because the system is not transparent to the user. We will deal with each of these in turn, after describing the data flow for the Sheffield multifrequency EIT system.

IDENTlPl FEATURES OF INTEREST

IZJ Functlonal

Figure 1. Data Row for the Sheffield multifrequency electrical impedance tomogmphy system

2.2. Dataflow in the Shefield system

The design criteria and hardware for the Sheffield system have been described elsewhere (Brown et al 1994a), so this description will concentrate on the data flow. Figure 1 shows this in schematic form. The raw data contain 64 voltage measurements at each of eight frequencies from 9.6 Wz to 1.2 MHz, together with 64 values of the gain of programmable

Data processing for serin1 i m g e s

A131

gain voltage measuring amplifiers (PGAs). In addition, there is a calibration matrix which contains gain and offset calibrations for each .voltage measuring channel. The calibration matrix is calculated from measurements made on a resistive phantom. The first step is to calibrate the raw data i.e. to correct for gain, offset and PGA setting for each channel. This gives the calibrated data, which are ,stored. Two sets of images can be reconstructed from the raw data. The first set is the dynamic images. which are the same as the images produced by the single-frequency Sheffield systems (Barber and Brown 1984, Smith et al 1990), apart from the use of interleaved rather than adjacent drive and receive. The data for the reconstruction of dynamic images are normalized with respect to reference data, which can be taken from any frame in the data set, but which are usually derived by averaging the calibrated data for a number of frames at the beginning of the data collection period. The second set is the frequency imges. The frequency images contain data from only a single frame, so that they can be considered to be static images. The data for the reconstruction of the frequency images are normalized with respect to data collected at any one of the eight frequencies. The reference frequency in this paper is the lowest frequency i.e. 9.6 kHz, but this may not be the optimum choice. The frequency images are slices of the impedance spectrum of each pixel, and therefore should give information about the characteristics of the tissue corresponding to each pixel. The impedance spectrum of each pixel can be modelled by a simple R, S, C model due originally to Cole (Cole and Cole 1941, Brown etal 1994a, Lu etal 1995). As the absolute resistivity is not known, only two of the calculated variables are independent. The resulting parametric i m g e s are displayed as images of RIS, f R s c , RS and CS, together with an image showing the rms error on the fit of the model to each pixel.

2.3. Preprocessing of the raw and calibrated data In many cases, the sampling rate required to satisfy the Shannomyquist criteria will be less than the maximum data collection rate. The signal to noise ratio should be maximized by averaging the raw data before sampling. For a 1 s sample rate with 33 frames s-' hardware, 33 frames would be averaged and stored. The origin of the gain and offset errors (which are removed in the calibration process) should be investigated before averaging of the raw data is implemented, to determine whether calibration should be performed before averaging. The Sheffield reconstruction algorithm (Barber and Brown 199Oa) is linear so that, in principle, any operation in image space has a counterpart in data space. The advantage of operating in data space is that the computational effort is greatly reduced. It should be noted that some operations do not commute-filtering reconstructed dynamic images is not identical to reconstructing filtered data. 2.4. Reconstruction of images

Harris et a1 (1988) described a fast reconstruction technique in which only the pixels within a region of interest were reconstructed. In general, we believe that fixed regions of interest should be replaced by regions that are calculated on a frame by frame basis, to take account of movement and of changes in shape and size of the organ that we wish to image. Nevertheless, as part of the preprocessing of the data, it is often possible to define a fixed part of the image which contains the dynamic region of interest, and then reconstruct only that part of the image. As the aim is to extract dynamic functional information from the images, the images themselves are only of interest during the validation process, and can be discarded as soon

A132

R H Smallwood and A R Hampshire

as the functional information has been obtained. Indeed, if the appropriate operation in data space can be found, there is no need to reconstruct any images.

3. EIT measurement of lung function 3.1. Introduction The majority of lung imaging has been performed using single-frequency systems, with a 16-electrode array placed around the chest, usually at the level of the sixth intercostal space. The change in impedance of the lungs is linearly proportional to lung volume (Harris et a1 1987, 1988), but the rate of changc of impedance varies from individual to individual. The lung impedance is also proportional to the volume of fluid in the lungs, as has been demonstrated in patients undergoing both dialysis and lung lavage, and in volunteers (McArdle et nl 1988). It is not possible, with a single-frequency system using Barber's algorithm, to determine the absolute resistivity of the lungs. It has been suggested that, by using a priori information on the conductivity distribution within the thorax (i.e. a simple model of the thorax), and knowledge of the electrode position and thorax shape, the absolute resistivity of the lungs can be determined (Smulders and van Oosterom 1992). A method of correcting the data set for non-circularity and non-uniform spacing of the electrodes has been described (Barber and Brown 1990b. Kiber eta1 1990), but not applied to the determination of absolute resistivity. Changes in the impedance spectrum of the in vivo lung at different lung volumes was first demonstrated using a four-electrode technique (Brown et a1 1994b. c). The extension of this to multifrequency impedance imaging has shown, in animal experiments (Brown 1994), that the change in lung water correlates with parameters that describe the impedance spectrum, so it should be possible to determine the volume of water in the lung. Measurements on a limited number of patients with pulmonary oedema indicate that individual patient measurements fall outside the normal range (Brown 1994). The electric model of tissue (Cole and Cole 1941) which has been used as the basis for the parametric modelling of the multifrequency data, has been related by Brown et a1 (1994~)to a simple model of lung tissue. This gives testable hypotheses relating the measured parameters to changes in the lung physiology. More detailed modelling of the lung has been performed by Nopp et a! (1993), and in vitro measurements of the impedance spectrum of lung tissue have been performed by a number of groups (Gersing and Osypka 1995, Lu et a1 1995). 3.2. Datu collection

The data collection method for multifrequency imaging of neonates has been described by Hampshire et al (1995). Sixteen Medicost Blue Sensor neonatal electrodes were equally spaced in a transverse plane at the xiphisternal level. Data were collected at the maximum rate of 33 frames s-'. A total of 2000 frames was collected. 3.3. Image reconstruction

Averaged dynamic, frequency and parametric images were reconstructed, as described by Hampshire et nl (1995), for preliminary visual analysis. We show here only parametric images from the average of frames at maximum inspiration on ten successive breaths. Figures 2 and 3 show the SC and RC images respectively. Superimposed on the images are contours at 0.6, 0.8, 1.0 and 1.2 ps. R J S is not a good indicator of lung position, but

Data processingfor serial inuages

A133

oftmsl!.xtdw

) of data vftn then analysed as fotlows: for ea&

frame,the pixels for which SC =- 1.2 fis were idedk&&d uoad:w Q w g b &Wet&

on the dynamic images. The average resistivity change within the region of interest was calculated for each of the eight frequencies.The reference for reconshuction80f?hdynarniC

A134

R H Smallwood and A R Hampshire

of

a0 sacc#sire

A135

Data processing for serial images J4A [50l:lOOO] roi-mean, SC>1.2 us, 19.2 kHz image

0

100

200

300

500

400

Frame number Figure 4. Mean resistivity change against frame number for pixels for which SC

5

1.2 11s.

applications in the gastroesophageal reflux (Erol et a1 1995), gastric stasis and dumping, and duodenogastric reflux. King efal (1987) used real time ultrasonography to measure the period and frequency of occurrence of antroduodenal motility, and Evans er a1 (1993) have used magnetic resonance imaging to demonstrate gastric contractions. EIT measurements of gastric emptying have involved pharmacological blockade of gastric secretion with e.g. cimetidine, because the acid secreted into the stomach is a major determinant of the conductivity of the stomach contents (Watson er a1 1995). Both saliva and duodenogastric reflux will alter the conductivity of the stomach contents, particularly during basal secretion (Gardham and Hobsley 1970, Hobsley 1974). The fixed regions of interest used in gastric emptying studies are a relatively crude method of identifying the stomach contents. If the stomach size decreases. or the relative position of electrodes and stomach changes as a result of postural change, the fixed region of interest must necessarily include non-stomach. We will present an alternative. 4.2. Data collection

The data collection protocol was similar to that described by Smallwood et a1 (1994) for single-frequency studies. A planar (rosette) array of electrodes, 17 or 21 cm in diameter, was placed on the abdomen. Images of the stomach were obtained from normal volunteers after a liquid followed by a solid meal using the Sheffield multifrequency imaging system

A136

R H Smallwood and A R Hampshire

(Brown et a1 1994a). Data were collected at a 1000 ms frame rate (average of 33 frames of raw data). The subjects fasted for at least 10 h prior to the test. A 2 min recording in the fasted state was used as the reference period. A drink of 200 ml of beef consomm6, conductivity 15 mS cm-’, was followed by 2 min of data collection to enable the stomach area to be identified. A cheese sandwich was then eaten to elicit gastric motility, and the overall length of the recording was 60 min. In this paper we will consider only the liquid component. 4.3. Filtering and averaging of raw &tu

In total, 3600 frames were collected in each study. A high frame rate is essential to avoid aliasing in the images. The frame rates used were adequate to avoid aliasing of impedance changes associated with respiration, but not for cardiac-related changes. In the present series of studies, the impedance changes in the stomach were of the order of 50%, compared to typical cardiac-related changes of about 1% (Leathard et a1 1994), so cardiac-related changes would be at the noise level of the system, and can be ignored. The image files were analysed off lie. The raw multifrequency data were calibrated, and then pairs of points were averaged and resampled to give 1800 frames at 1 frame/2 s. This is the frame rate used in previous single-frequency studies, and was used to facilitate comparison of singlefrequency and multifrequency data. It would be possible to reconstruct and save all 1800 frames at one frequency, but the memory requirement for reconstructing and storing all the parametric images is excessive. The data. were therefore processed in blocks of 100 frames, and the image data were discarded after the required information had been extracted from each frame. 4.4. Identifying the stomach

One a single-frequency dynamic image, the only parameter is the impedance change relative to the reference frames. The stomach can be identified by changing the conductivity of the contents, as well as by identifying the characteristic motility at 0.05 Hz. This can be done by swallowing a highly conductive liquid, hence the test meal of soup with a conductivity of 15 mS cm-’. This is a much higher conductivity than the surrounding tissue, and is only exceeded by the stomach contents when pH< 2 (Watson et al 1995). The areas of the images that show a large increase in conductivity will thus correspond to the stomach. The contractions of the stomach wall, with a 20 s period, are complex, and no single frame will show the maximum extent of the stomach. However, as the stomach position is identified by high conductivity, the maximum extent of the stomach can be shown by forming an image of the maximum conductivity for each pixel in the image over the time span of the investigation. This image for the soup period is shown in figure 5. Two regions of high conductivity (negative resistivity) change can be distinguished. The extent of the stomach can be defined by the contour corresponding to a fixed change in conductivity. Contours corresponding to a resistivity change of -0.1 and -0.2 are superimposed on the image. 4.5’. Area measurement and liquid meals

Given a definition of the stomach area (e.g. > 10% change in conductivity), it should be possible to follow gastric emptying by calculating the stomach area on successive frames. Figure 6 shows the contour area (normalized by division by the area of the contour on figure 5 ) for the liquid meal. As the image in figure 5 clearly shows two areas, the contour size for the (true) left and right areas have also been calculated separately. The area on the left has previously been shown to be stomach. Comparison of figure 5 with an anatomical

Daw processingfor serial images

isquired. Rgun emplyingai* lquid mlai Ram tlte stom& volumcs with low noisc, which is osse 4.6. Paratnem'c imaging

A137

A138

R H Smllwood and A R Hampshire S19F2 [100:160] total

100

110

120

130

140

150

160

S19F2 [100:160] left roi

"l"".

100

110

120

130

140

150

160

150

160

S19F2 [100:160] right roi

100

110

'

120

130

140

frame number Figure 6. Change in area of contours at a resistivity change of -0.1 on the 19.2 kHz dynamic image for the liquid meal against frame number: top, total area;middle, hue left (stomach) area;. bottom, hue right area.

Contrary to the prediction that the stomach contents (the liquid meal) should appear as an area of high fRsc, there is no apparent correlation between any of the parametric images and the area of stomach shown on the dynamic images. 5. Discussion

The expectation, at the beginning of this study, was that the paramehc images could be used to generate a dynamic region of interest for monitoring changes in both lungs and stomach. We have shown that, even in the difficult case of natural neonatal respiration, the lung area can be clearly identified on fRsc, SC and RC images. The small cross-sectional

Data processing for serial images

A139

A140

R H Smallwood and A R Hampshire

Data processing for serial images

A141

the short data acquisition time (15 ms), are the key to removing the effects of movement artefacts, and thus making monitoring possible in unrestrained infants and ambulant subjects. Multifrequency imaging should give a much greater immunity from movement artefact. In the stomach, the situation is much less clear. With a liquid meal, it is certainly possible to apply dynamic regions of interest to dynamic images of the stomach, and extract sensitive measures of stomach size. Measurement of gastroduodenal transport may be possible, but we have been unable to identify the stomach on parametric images. It is not clear why this should be the case. An improved understanding of the relationship between the impedance distribution and impedance spectrum of tissues, and the resulting images, is urgently required, as is rigorous comparison of four-electrode in virro and in vivo measurements with images.

Acknowledgments Liqin Lu provided, with great efficiency, the C routines for parametric modelling of multiple frames. Rosie Er01 and Jo Watson assisted, with unfailing good humour, in the collection of gastric data, and Rob Primhak allowed us to take measurements on his neonatal patients.

References Avill R, Mmgnall Y F, Bird N C, Brown B H, B a r k D C. Seagar A D, Johnson A G and Read N W 1987 Applied potential tomography: a new noninvasive technique for measuring gastric emptying Gartroenterology 92 1019-26 Barber D C and Brawn B H 1984 Applied potential tomography 3, Phys. E; Sci. Insuum 17 7 2 S 3 3 -199Oa Progress in electrical impedance tomography Inverse Problem in Partial Differentid Equations ed D Colton, R Ewing and W Rundell (Philadelphia, PA SIAM) pp 1 5 1 6 4 - 1990b Shape comction i n AFT image rewnstruction Proe. Copenhgen Meeting on Electrical Impedance Tomography (Copenhogen, 1990) (Sheffield Univerrity of Sheffield).pp U 5 1 Bxber D C, Johnson A G and Read N W 1988 Comparison of applied potential tomography and impedance epigastrography as methods of measuring gastric emptying Clin Phys. PhysioL M e a . 9 249-54 Baxter A 1, Mangnall Y F, Loj E H, Brown B, Barber D C, Johnson A G and Read N W 1988 Evaluation of applied potential tomography as a new non-invasive gastric secretion test. Gut 29 1730-5 Brown B H 1994 Personal communication Brown B H, Barber D C, Leathard A D, Lu L, Wang W and Smallwood R H 1994a High frequency Err data collection and parametric imaging Innov. Technol.B i d . Med. 15 1-10 Brown B H,Barber D C, Morice A H and Leathard A D 1994b Cardiac and mpiratory related changes in the human thorax IEEE T m . Biomed. Eng. BME-41729-34 I, A R, Mackay R and Brown B H, Barber D C, Wang W. Lu L. Leathard A D, Smallwood R €Hampshire Hatzigalanis K 1994c Multi-frequency imaging and modelling of respiratory related electrical impedance changes Phys. M e a . 15 (Supplement A) A1-A12 Cole K S and Cole R H 1941 Dispersion and absorption in dielectrics J. Chem Phys. 9 341-51 Er01 R A, Smallwood R H. Brown B H, Cherian P and Bardhan K D 1995 Detecting oesophagd-related changes using electrical impedance tomography Physiol. Meas. 16 (Supplement A) A143-Al52 Evans D F, h m o n t G, Stehling M K, Blamire A M, Gibbs P. Coxon R, Hardcastle J D and Mansfield P 1993 Prolonged monitoring of the upper gastrointestinal Vact using echo planar magnetic resonance imaging Gut 34 848-52 Evans D F md Wright I W 1990 Is acid suppression necessary when measuring gastric emptying using applied potential tomography? Pmc. CopenhagenMeeting on Electrical Impedance Tomography (Copenhagen 1990) (Sheffield University of Sheffield) pp 249-55 Gardham J R C and Hobsley M 1970 The electrolytes of human alkaline gastric juice Clin. Sci. 39 7 7 4 7 Hampshire A R, Smallwood R H, Brown B H and Primhak R A 1995 Multifrequency parametric EIT images of the neonatal lungs Physiol. Meas. 16 (Supplement A) A175-AI89

A142

R H Smallwood and A R Hampshire

Harris N D. Suggett A J, Barber D C and Brown B H I987 Application of applied potential tomography (AFT) in respiratory medicine Clin. Phys. Physiol. Mens.8 (Supplement A) A155-66 -1988 Applied potential tomography; a new technique for monitoring pulmonary function Clin. Phys. Physiol. Mem. 9 (Supplement A) A79-86 Hobsley M 1974 Pyloric reflux: a modification of the two-component hypothesis of gastric secretion Clin. Sci. Mol. Med 47 131-141 Houghton L A, Mangnall Y F and Read N W 1990 Etfect of incorporaling fat into a liquid test meal on the relation betwen inagastric distribution and gastric emptying in human volunteers Gut 31 1226-9 Kiber M A. Barber D C and Brown B H 1990 Estimation of object baundary shape from the voltage gradient measurements Proc. Copenhagen Meeting on Electricnl I m p e h c e Tomography(Copenhagen, 1990) (Sheffield University of Sheffield) pp 52-9 King P M, Plyde A and Heading R C 1987 Transpyloric fluid movement and antroduodenal motility in patients with gastro-oesophageal reflux Gut 28 545-8 Lamont G L. Wright J W, Evans D F and Kapila L A 1988 An evaluation of Applied Potential Tomography in the diagnosis of infantile hypemophic pyloric stenosis Clin. Phys. Physiol. Mear. 9 (Supplement A) A65-9 Leathard A D, Brown B H, Campbell 1, zhang F, Morice A H and Taylor D 1994 A comparison of ventilatory and cardiac related c h a n w in EIT images of normal human lungs - and of lungs - with pulmonary emboli Phvslol. Mcm, 15 (Supplement A) A137-&46 Lu L. Brown B H, Barber D C and Leathard A D 1995 A fast parameter modelling.algorithm wifi the Powell . method Physiol Mens 16 (Supplement A) A39-A47 Mangnall Y F, Barber C and Brown B H, Barber D C, Johnson A G and Read N W 1988 Comparison of applied potential tomography and impedance epigastrography as methods of measuring gamic emptying Clin.Phys. Pkysiol. Mem. 9 %9-54 Mangnall Y F, Baxter A J, Avill R, Bud N C. Brown B H, Barber D C, Seagar A D , Johnson A G and Read N W I987 Applied potential tomography: a new non-invasive technique for assessing gastric function Clin. Phys. Physiol. Mem. 8 (Supplement A) A119-29 McArdle F J, Suggeu A J, Brown B Hand Barber D C 1988 An assessment of dynamic images by applied potential tomography for monitoring pulmonary perfusion Clin Phys. Physiol. Mem. 9 (Supplement A) A 8 7 4 2 Nopp P, Rapp E, Pfutzuer H. Nakesch H and Ruhsam Ch 1993 Dielectic content of lung tissue as a function of air content Pkys. Med B i d . 38 699-716 Now S, Mangnall Y,Dickson J A S, Pearse R and Johnson A G 1993 Measurement of gastric emptying in infants with pyloric stenosis using applied potential tomography Arch Dis. Childhood 68 4 8 M I Mens, Osypka M and Gersing E 1995 Xssue impedance spectra and the appropriate frequencies for E ~ Pkysiol. 16 (Supplement A) A 4 9 4 5 5 Rigaud B, Hamzaoui L, Chauveau N, Granii M, Scotto Di Rinaldo J-P and Morucci J P 1994 Tisue characterization by impedance: a multi-frequency approach Physiol. Mens. 15 (Supplement A) A13-AZO Smallwood R H. Manmall Y F and h t h a r d A D 1994 T m s ~ o r tof eastric contents Phvsiol. Mens. 15 (Supplement A) Ai75-Al88 Smallwood R H. Now S, Manmall Y and Smvthe A 1993 Impedance lmaeine and Gastric Motilitv Clinical and Phyrwlogicol Applicorioi of Elcetdcnl Impedance Tomogioplty ed D tiolder (Landon: UCL P& pp 145-53 Smith R W M. Brown B H. Freerton I L and McArdle F J 1990 Pme. CopenhagenMeering on Eleerrtrol/mpe&cc Tomogrophy (Copenhogen. 1990) (Sheffield University of Sheffield) pp 212-6 Smulderr L A Wand van Oostemm A 1992 Appliwtion of electrid impedance tomogmphy to the determinniion of lung volume Clin P h p Phyriol. Mens. 13 (Supplement A) A167-70 Warson S I. Smallwood R H, Brown B H. Chcrim P md Bydhan K D 1995 Determination of intragastric pH using ELT Proc. Concened Action on Impedance Tomgrophy Workshop (Ahnro. 1995) Wright I W and Evans D F 1990 Applied Potential Tomography (AFT): a non-inv?siuve method of detecting the m i w n g motor complex (MMC) Pmc. Copenhagen Meering on Elecrrical lmpedonce Tomogrophy (Copenhagen, 1990) (Shcffiicld Univmity of Sheffield) pp 270-5

-