*4.3.1. Motion detection for a multilayered phantom*

We compared the motion detection by variance calculation in a combined MRI/UWB measurement using appropriate MR-compatible phantoms [29]. The dielectric phantoms were arranged in a sandwich structure to mimic the sequence of biological tissue layers of the human thorax.

**Figure 30. a)** Set-up of the combined MRI/UWB measurement; **b)** Comparison of the reference profile with the data obtained simultaneously from MR and UWB radar measurements. The profiles are offset for clarity.

Such a sandwich was placed in a moveable sledge-like fixture inside a birdcage MR head coil. The motion profile of the sandwich structure was shaped to approximate respiratory motion of the thorax superimposed by cardiac oscillations (Fig. 30). An M-sequence UWB-Radar system (up to 5 GHz) [76] and MR compatible UWB antennas [10], [32] were utilized to detect the motion of the phantom inside a 3-T MR scanner (Bruker MEDSPEC 30/100). A flow-compensated gradient echo CINE (time resolution 50 ms) sequence was used to reduce artifacts generated by the phantom movements.

Additionally, the physiological signatures monitored by UWB-radar were validated by comparison to simultaneously acquired MR measurements on the same subject [13], [29], [77] (cf. Section 4.5.2 and 4.6).

### *4.3.2. Detection of micro motion*

Subject motion appears to be a limiting factor in numerous MR imaging applications especially in the case of high and ultra-high fields, e.g. high-resolution functional MRI (fMRI). For head imaging the subject's ability to maintain the same head position is limiting the total acquisition time. This period typically does not exceed several minutes and may be

### 290 Ultra-Wideband Radio Technologies for Communications, Localization and Sensor Applications

considerably reduced in the case of pathologies. Several navigator techniques have been proposed to circumvent the subject motion problem [73]. MR navigators, however, do not only extend the scan because of the time necessary for acquiring the position information, but also require additional excitation pulses affecting the steady-state magnetization. Furthermore, if the very high spatial resolution offered by ultra-high-field MR scanners shall be exploited, the displacements caused by respiration and cardiac activity have to be considered. Thus, we propose to apply an UWB radar technique to monitor such micro motions.

ultraMEDIS – Ultra-Wideband Sensing in Medicine 291

Especially for the human torso where - due to higher harmonics from the highly nonlinear respiratory cycle - the separation of the cardiac cycle by common signal filtering in the frequency domain is limited, another separation of motion components is necessary. For this reason, the separation of motion components based on blind source separation (BSS) was

The IRF from a single UWB shot is a time series of 511 data points with a dwell time of 112 ps. This defines an IRF time scale of 57 ns but is still instantaneous compared to anatomical motions. These shots are then repeated for instance 2000 times at a rate of 44 Hz covering a total time span of 45 s. For further analysis, only the most interesting regime of the IRF data is considered. These are the 100 data points, i.e. a window of 11.2 ns, right after the IRF maximum due to direct cross-talk between *Tx* and *Rx* antenna. Following the temporal evolution of each selected data point over the 2000 repetitions, 100 virtual channels are obtained and subjected to BSS decomposition (ROI, see Fig. 32.a). By removing the mean values in these virtual channels, the changes of the radar signal on the anatomical time scale become visible, see Fig. 32.b. The motion pattern is dominated by respiration; cardiac

motion is considerably smaller and not immediately visible in the raw data.

**Figure 32. a)** Single IRF and **b)** radargram of one channel with region of interest and **c)** selected 100

The data analysis is based on the BSS and assumes a measured signal *x*(*t*) to be a linear combination of unknown zero-mean source signals *s*(*t*) with an unknown mixing matrix **A**:

> <sup>1</sup> ( ) ( ) ( ,..., )*<sup>T</sup> <sup>m</sup> xt st x x x* **<sup>A</sup>** .

The original sources *s*(*t*) can be estimated by the components *y*(*t*) which can be calculated

In our analysis, a second-order time-domain algorithm (TDSEP, Temporal Decorrelation source SEParation) was applied which is described in detail in [78]. In TDSEP the unknown

.

)()()( \* \* **AAA** *tstxty* (3)

(2)

virtual channels, mean value removed

from the estimation of the de-mixing matrix **A**\* ≈ **A**-1:

developed.

**Figure 31.** Motion reconstructed from a measured time interval of 350 s. The right inset at the top displays the four nodding events (~1 mm amplitude, episode [t = 10 s,…, t = 18 s]) to localize the surface of the head. Respiratory displacements are clearly visible (right inset bottom, episode [t = 110 s,…, t = 160 s]) and spontaneous twitches are highlighted.

First *in-vivo* motions reconstructed from a measured time interval of 350 s are shown in Fig. 31. By applying appropriate filters in a selected time interval even the cardiac induced displacements were detected with an amplitude of about 40 m. Thus, we could detect all kinds of involuntary motions (respiratory, cardiac), but also doze-off-events are visible, demonstrating the feasibility of interfacing an MR scanner with an external UWB radar based motion tracking system. Our system is capable of determining the position of interest with sub-millimeter accuracy and an update rate of 44.2 Hz. Using the UWB tracking data of the volunteer's head, the motion artifacts can be compensated for in real time or by postprocessing enhancing the actual resolution of the MR scan [73].

### *4.3.3. Separation of motion components by blind source separation*

Monitoring the motion inside the human body, the correlation functions of transmitted and received signals (i.e. the IRF) contain a mixture of all simultaneously occurring motions. Especially for the human torso where - due to higher harmonics from the highly nonlinear respiratory cycle - the separation of the cardiac cycle by common signal filtering in the frequency domain is limited, another separation of motion components is necessary. For this reason, the separation of motion components based on blind source separation (BSS) was developed.

290 Ultra-Wideband Radio Technologies for Communications, Localization and Sensor Applications

motions.

considerably reduced in the case of pathologies. Several navigator techniques have been proposed to circumvent the subject motion problem [73]. MR navigators, however, do not only extend the scan because of the time necessary for acquiring the position information, but also require additional excitation pulses affecting the steady-state magnetization. Furthermore, if the very high spatial resolution offered by ultra-high-field MR scanners shall be exploited, the displacements caused by respiration and cardiac activity have to be considered. Thus, we propose to apply an UWB radar technique to monitor such micro

**Figure 31.** Motion reconstructed from a measured time interval of 350 s. The right inset at the top displays the four nodding events (~1 mm amplitude, episode [t = 10 s,…, t = 18 s]) to localize the surface

First *in-vivo* motions reconstructed from a measured time interval of 350 s are shown in Fig. 31. By applying appropriate filters in a selected time interval even the cardiac induced displacements were detected with an amplitude of about 40 m. Thus, we could detect all kinds of involuntary motions (respiratory, cardiac), but also doze-off-events are visible, demonstrating the feasibility of interfacing an MR scanner with an external UWB radar based motion tracking system. Our system is capable of determining the position of interest with sub-millimeter accuracy and an update rate of 44.2 Hz. Using the UWB tracking data of the volunteer's head, the motion artifacts can be compensated for in real time or by post-

Monitoring the motion inside the human body, the correlation functions of transmitted and received signals (i.e. the IRF) contain a mixture of all simultaneously occurring motions.

of the head. Respiratory displacements are clearly visible (right inset bottom, episode

[t = 110 s,…, t = 160 s]) and spontaneous twitches are highlighted.

processing enhancing the actual resolution of the MR scan [73].

*4.3.3. Separation of motion components by blind source separation* 

The IRF from a single UWB shot is a time series of 511 data points with a dwell time of 112 ps. This defines an IRF time scale of 57 ns but is still instantaneous compared to anatomical motions. These shots are then repeated for instance 2000 times at a rate of 44 Hz covering a total time span of 45 s. For further analysis, only the most interesting regime of the IRF data is considered. These are the 100 data points, i.e. a window of 11.2 ns, right after the IRF maximum due to direct cross-talk between *Tx* and *Rx* antenna. Following the temporal evolution of each selected data point over the 2000 repetitions, 100 virtual channels are obtained and subjected to BSS decomposition (ROI, see Fig. 32.a). By removing the mean values in these virtual channels, the changes of the radar signal on the anatomical time scale become visible, see Fig. 32.b. The motion pattern is dominated by respiration; cardiac motion is considerably smaller and not immediately visible in the raw data.

**Figure 32. a)** Single IRF and **b)** radargram of one channel with region of interest and **c)** selected 100 virtual channels, mean value removed

The data analysis is based on the BSS and assumes a measured signal *x*(*t*) to be a linear combination of unknown zero-mean source signals *s*(*t*) with an unknown mixing matrix **A**:

$$\mathbf{x}(t) = \mathbf{A}\mathbf{s}(t) \quad \mathbf{x} = \begin{pmatrix} \boldsymbol{\chi}\_1 \dots \boldsymbol{\chi}\_m \end{pmatrix}^T \cdot \tag{2}$$

The original sources *s*(*t*) can be estimated by the components *y*(*t*) which can be calculated from the estimation of the de-mixing matrix **A**\* ≈ **A**-1:

$$\mathbf{y}(t) = \mathbf{A}^\* \mathbf{x}(t) = \mathbf{A}^\* \mathbf{A} \mathbf{s}(t) \tag{3}$$

In our analysis, a second-order time-domain algorithm (TDSEP, Temporal Decorrelation source SEParation) was applied which is described in detail in [78]. In TDSEP the unknown

### 292 Ultra-Wideband Radio Technologies for Communications, Localization and Sensor Applications

mixing matrix **A** is calculated by simultaneous diagonalization of a set of correlation matrices **R**(*x*) for different choices of .

$$\begin{aligned} \mathbf{R}\_{\tau(x)} &= \left< \mathbf{x}(t)\mathbf{x}^T(t-\tau) \right> \\ \mathbf{R}\_{\tau(x)} &= \left< \mathbf{A}\mathbf{s}(t)(\mathbf{A}\mathbf{s}(t-\tau))^T \right> = \mathbf{A}\mathbf{R}\_{\tau(s)}\mathbf{A}^T \end{aligned} \tag{4}$$

ultraMEDIS – Ultra-Wideband Sensing in Medicine 293

*4.4.2. Multi-channel UWB radar applying two receiver channels for cardiac trigger events* 

We started our multi-channel UWB radar development with the single module device enabling us to add a second receiver (*Rx*) antenna, oriented towards the left-anterior oblique direction [68] (Fig. 33.a), to the existing *Rx* and *Tx* antennas facing the antero-posterior direction. The UWB data were recorded at 44.2 Hz. Corresponding to the data selection in Section 4.3.3, we obtained 200 virtual data channels from the IRFs of two UWB

**Figure 33. a)** Scheme of the UWB radar with one transmitter (*Tx*) and two receiver (*Rx*) antennas and measurement set-up; **b)** Cardiac UWB signal applying both *Rx* channels and the calculated trigger

In the cardiac UWB signal, we chose the points of maximum myocardial contraction during the heart cycle. These points are related to the minima of the UWB signal (Fig. 33.b: squares). To increase the robustness of this detection scheme, we combined it with a simple difference calculation at the trailing edge of the minima. Additional consistency checks on the

By employing two *Rx* channels (Fig. 33.a) the UWB radar detection of the cardiac cycle worked reliably, even in the free breathing mode. In simple cases, e.g. under breath-holding conditions, it is possible to detect cardiac motion with just one *Rx* channel. However, this

By using two *Rx* channels, it was still necessary to align the antennas properly towards the heart. This becomes more critical for measurements during cardiac MRI where the MR coil is placed on the chest of the subject, partly blocking the free line of sight between radar antenna and the heart, see Fig. 38. With our development of a multiple-antenna set-up it is much easier to handle this adjustment by just choosing the 'good channels' in a pool of

events in the signal by combination of low peak and slew rate calculation.

oscillation amplitude were used to suppress double triggering.

will not work in general, more complicated situations.

*4.4.3. Application of up to 32 receiver channels* 

available channels.

measurement channels for the decomposition by blind source separation (BSS).

where the angular brackets denote time averaging. The quality of signal separation depends strongly on the choice of . However, solving **R**(*x*) = **AR**(*s*) **A**T for several by simultaneous diagonalization eliminates this obstacle. It is recommended by biomagnetic research to choose the number of time shifts larger than 40 and to include the time constant of those components which are known a priori, e.g. the range of possible cardiac frequencies 1/cardiac [79]. Additionally, Principal-Component Analysis (PCA) compression was applied to reduce the number of channels used for generating the correlation matrices **R**(*x*) and reduce computation time for the BSS. The components of the resulting sources are calculated using eq. (3). Automatic identification of the cardiac component was provided by a frequency-domain selection criterion because for non-pathological conditions the main spectral power density of the heart motion falls in a frequency range of 0.5 Hz to 7 Hz. The algorithm searches for the highest ratio between a single narrowband signal (fundamental mode and first harmonic) within this frequency range and the maximum signal outside this range. A high-order zero-phase digital band pass filter of 0.5–5 Hz was applied to the cardiac component of the UWB signal. In a similar way, respiration can be identified by the BSS component with the maximum L2 norm in the frequency range of 0.05 Hz to 0.5 Hz.

### **4.4. Analyses of cardiac mechanics by multi-channel UWB radar**

### *4.4.1. Compatibility of MRI and UWB radar*

Compatibility is the most challenging issue when combining MRI with other modalities. Therefore, the UWB antennas employed are important parts. Eddy currents due to the switching magnetic gradient fields as well as the interference with the powerful RF pulses from the MRI scanner, see Section 2.3.3, were minimized by proper antenna design. The cutoff frequency of the MR-compatible double ridged horn antennas at 1.5 GHz [32] marks the lower limit of the transmitted and received signal frequencies. Coupling to the narrowband MRI frequencies (300 MHz at 7 T, 125MHz at 3T) is thus minimized in both directions. Additionally, the inputs of both our UWB radar systems (MEODAT GmbH, Ilmenau, Germany), one single-module: 1*Tx*-2*Rx*-device and one four-module multi-input-multi-output: 4*Tx*-8*Rx*-MIMO device, were protected by 1.2 GHz high pass filters. In both UWB systems, the transmitted radar signals were generated by a pseudo-random M-sequence. With *m* = 9 it has a length of 2*m*-1 = 511 clock signals at *f0* = 8.95 GHz [76]. The equivalent UWB power spectrum extends up to *f0*/2. The impulse response function (IRF) is obtained as mentioned before by correlation of the received signal of the investigated object with the M-sequence [76]. By means of this technique, the signal-to-noise ratio is improved due to the removal of the uncorrelated noise by the correlation of the received signals with the transmitted signal pattern. In this way, even smallest parts of the RF pulses of the MRI were avoided.

### *4.4.2. Multi-channel UWB radar applying two receiver channels for cardiac trigger events*

292 Ultra-Wideband Radio Technologies for Communications, Localization and Sensor Applications

.

*T*

)()(

*txtx*

. However, solving **R**

**4.4. Analyses of cardiac mechanics by multi-channel UWB radar** 

even smallest parts of the RF pulses of the MRI were avoided.

*tsts*

matrices **R**

1/ strongly on the choice of

choose the number of time shifts

(*x*) for different choices of

*x*

*x*

)(

**R**

*4.4.1. Compatibility of MRI and UWB radar* 

mixing matrix **A** is calculated by simultaneous diagonalization of a set of correlation

**AAR AAR**

))()((

)( )(

where the angular brackets denote time averaging. The quality of signal separation depends

diagonalization eliminates this obstacle. It is recommended by biomagnetic research to

components which are known a priori, e.g. the range of possible cardiac frequencies

reduce computation time for the BSS. The components of the resulting sources are calculated using eq. (3). Automatic identification of the cardiac component was provided by a frequency-domain selection criterion because for non-pathological conditions the main spectral power density of the heart motion falls in a frequency range of 0.5 Hz to 7 Hz. The algorithm searches for the highest ratio between a single narrowband signal (fundamental mode and first harmonic) within this frequency range and the maximum signal outside this range. A high-order zero-phase digital band pass filter of 0.5–5 Hz was applied to the cardiac component of the UWB signal. In a similar way, respiration can be identified by the BSS component with the maximum L2 norm in the frequency range of 0.05 Hz to 0.5 Hz.

Compatibility is the most challenging issue when combining MRI with other modalities. Therefore, the UWB antennas employed are important parts. Eddy currents due to the switching magnetic gradient fields as well as the interference with the powerful RF pulses from the MRI scanner, see Section 2.3.3, were minimized by proper antenna design. The cutoff frequency of the MR-compatible double ridged horn antennas at 1.5 GHz [32] marks the lower limit of the transmitted and received signal frequencies. Coupling to the narrowband MRI frequencies (300 MHz at 7 T, 125MHz at 3T) is thus minimized in both directions. Additionally, the inputs of both our UWB radar systems (MEODAT GmbH, Ilmenau, Germany), one single-module: 1*Tx*-2*Rx*-device and one four-module multi-input-multi-output: 4*Tx*-8*Rx*-MIMO device, were protected by 1.2 GHz high pass filters. In both UWB systems, the transmitted radar signals were generated by a pseudo-random M-sequence. With *m* = 9 it has a length of 2*m*-1 = 511 clock signals at *f0* = 8.95 GHz [76]. The equivalent UWB power spectrum extends up to *f0*/2. The impulse response function (IRF) is obtained as mentioned before by correlation of the received signal of the investigated object with the M-sequence [76]. By means of this technique, the signal-to-noise ratio is improved due to the removal of the uncorrelated noise by the correlation of the received signals with the transmitted signal pattern. In this way,

to reduce the number of channels used for generating the correlation matrices **R**

cardiac [79]. Additionally, Principal-Component Analysis (PCA) compression was applied

(*x*) = **AR**

*T*

*T s*

(*s*) **A**T for several

larger than 40 and to include the time constant of those

by simultaneous

(*x*) and

(4)

We started our multi-channel UWB radar development with the single module device enabling us to add a second receiver (*Rx*) antenna, oriented towards the left-anterior oblique direction [68] (Fig. 33.a), to the existing *Rx* and *Tx* antennas facing the antero-posterior direction. The UWB data were recorded at 44.2 Hz. Corresponding to the data selection in Section 4.3.3, we obtained 200 virtual data channels from the IRFs of two UWB measurement channels for the decomposition by blind source separation (BSS).

**Figure 33. a)** Scheme of the UWB radar with one transmitter (*Tx*) and two receiver (*Rx*) antennas and measurement set-up; **b)** Cardiac UWB signal applying both *Rx* channels and the calculated trigger events in the signal by combination of low peak and slew rate calculation.

In the cardiac UWB signal, we chose the points of maximum myocardial contraction during the heart cycle. These points are related to the minima of the UWB signal (Fig. 33.b: squares). To increase the robustness of this detection scheme, we combined it with a simple difference calculation at the trailing edge of the minima. Additional consistency checks on the oscillation amplitude were used to suppress double triggering.

By employing two *Rx* channels (Fig. 33.a) the UWB radar detection of the cardiac cycle worked reliably, even in the free breathing mode. In simple cases, e.g. under breath-holding conditions, it is possible to detect cardiac motion with just one *Rx* channel. However, this will not work in general, more complicated situations.

### *4.4.3. Application of up to 32 receiver channels*

By using two *Rx* channels, it was still necessary to align the antennas properly towards the heart. This becomes more critical for measurements during cardiac MRI where the MR coil is placed on the chest of the subject, partly blocking the free line of sight between radar antenna and the heart, see Fig. 38. With our development of a multiple-antenna set-up it is much easier to handle this adjustment by just choosing the 'good channels' in a pool of available channels.

By integration of a MIMO UWB device (MEODAT GmbH, Ilmenau, Germany) containing four modules, each with one *Tx* and two *Rx* channels [76], up to 32 channels became available. In a 1 *Tx* \* 8 *Rx* configuration a sampling frequency up to 530.4 Hz can be realized. Using sequentially activated transmitters the set-up can be extended to 32 channels (4 *Tx* \* 8 *Rx*) at a reduced sampling frequency of up to 132.6 Hz. For cardiac motion detection, the four *Tx* and eight *RX* antennas are placed over the chest as depicted in Fig. 34 and adjusted to aim for one central point at a distance of 100 cm.

ultraMEDIS – Ultra-Wideband Sensing in Medicine 295

jitter of their trigger contributions and smearing out the sharpness in the combined signal determined by the BSS. By the preparatory check those channels with the highest variation in their 'cardiac' signals were excluded as they were likely contaminated with noise or other motion components. By rejecting the eight noisiest channels and recalculating the BSS with the remaining channels, a cardiac signal is obtained with sharp trailing slopes and well-

The primary goal of this development was to simplify the system handling during cardiac navigation for high-resolution MRI. In addition, the capability of monitoring non-invasively the cardiac activity of a person in an unknown position, e.g. in a patient bed, can be important for a variety of novel medical applications in clinical medicine and biomedical research. As multi-channel UWB radar is unimpeded by bedding or clothing, it is applicable not only in conjunction with MRI. It would also be a valuable stand-alone modality for intensive care monitoring of patient groups not permitting the use of skin contact sensors. Neonates, children

Stand-alone UWB radar enables the detection of cardiac activities by different illumination conditions as shown in [68] for the radiographic standard position. The illumination of the heart from only one side at a time, like the frontal direction for motion detection as depicted in Fig. 33, was extended to the simultaneous illumination of two sides. No averaging was performed to enable the comparison of single heart beats [30]. This approach can open the field for new diagnostic applications by detecting differences and disturbances in comparative measurements of the left and right ventricle, thus recognizing potentially pathological patterns [69]. Two groups of four *Rx* and one *Tx* antennas were applied for this purpose. The first was placed in the left lateral and the second in the right anterior oblique

Each antenna group consisted of a single *Tx* antenna surrounded by four *Rx* antennas. All antennas were directed towards the estimated center position of the heart. The challenge was to measure the cardiac motion even from the lateral position, where the attenuation of the reflected signals from the heart is much higher due to the prolonged propagation path in tissue. The data analysis by BSS was applied for both antenna groups separately. For

For lateral position, the UWB signal from the cardiac motion is considerably weaker and much more affected by noise. However, by increasing the number of *Rx*- channels, the signal quality improved substantially, effectively compensating the strong signal attenuation (s. Fig. 36.b). Only healthy volunteers were examined in this particular study but even among them characteristic peculiarities can be found. In both ventricles, the contraction velocity (trailing edge of the UWB motion curve) is higher than the velocity of ventricle dilatation. The duration of the dilation period, on the other hand, is longer for the right ventricle compared to its counterpart on the left. More characteristic features are

comparison, the data of only two or all four *Rx* channel per group were analyzed.

expected to be visible in patients with cardiac diseases or malfunctions.

at risk of sudden infant death syndrome or burn victims are just a few examples.

*4.4.4. Illumination of human thorax by multiple antenna groups* 

defined trigger events (s. Fig. 35.c).

position.

**Figure 34.** a) Scheme of the UWB radar set-up with 8 *Rx* and 4 *Tx* antennas b) MR compatible measurement set-up.

**Figure 35.** Cardiac signal and detected trigger events for **a)** two hand-picked best channels, **b)** all 32 channels, **c)** the 24 "good channels".

The procedure to identify the most useful channels for triggering starts with a short preparatory measurement, where each channel is analyzed by the BSS to decompose the complex UWB signals [80], extracts the relevant cardiac component and calculates the trigger events as described in Section 4.4.2. The quality of each measurement channel is assessed by calculating the variation of the time interval between trigger events. For comparison, Fig. 35.a depicts the result of the BSS analysis by the best two channels, manually selected for the smallest variation between the trigger events. The cardiac signal based on these two hand-picked channels represents the best achievable result for a set-up like in the section before. By utilizing all 32 channels for the BSS, a smoother cardiac signal is detected, and the motion amplitude shows less variation over the time. However, the sharpness of the trailing slopes is also reduced. Due to this fact, the third trigger event escaped detection (s. Fig. 35.b). Some of the 32 channels contained much noise resulting in a jitter of their trigger contributions and smearing out the sharpness in the combined signal determined by the BSS. By the preparatory check those channels with the highest variation in their 'cardiac' signals were excluded as they were likely contaminated with noise or other motion components. By rejecting the eight noisiest channels and recalculating the BSS with the remaining channels, a cardiac signal is obtained with sharp trailing slopes and welldefined trigger events (s. Fig. 35.c).

The primary goal of this development was to simplify the system handling during cardiac navigation for high-resolution MRI. In addition, the capability of monitoring non-invasively the cardiac activity of a person in an unknown position, e.g. in a patient bed, can be important for a variety of novel medical applications in clinical medicine and biomedical research. As multi-channel UWB radar is unimpeded by bedding or clothing, it is applicable not only in conjunction with MRI. It would also be a valuable stand-alone modality for intensive care monitoring of patient groups not permitting the use of skin contact sensors. Neonates, children at risk of sudden infant death syndrome or burn victims are just a few examples.
