Echo-Planar Spectroscopic Imaging (EPSI) of the Water Resonance Structure in Human Breast Using Sensitivity Encoding (SENSE)

Abstract

High spectral and spatial resolution MRI, based on echo-planar spectroscopic imaging, has been applied successfully in diagnostic breast imaging, but acquisition times are long. One way of increasing acquisition speed is to apply the sensitivity encoding algorithm for complex high spectral and spatial resolution data. We demonstrate application of a complex sensitivity encoding algorithm to high spectral and spatial resolution MRI data, in a phantom and human breast, with 7-and 16-channel dedicated breast phased-array coils. Very low g factors are obtained using the breast coils, and the signal-to-noise ratio (SNR) penalty for water resonance peak height and water resonance asymmetry images is small at acceleration factors of up to 6 and 4, respectively, as evidenced by high Pearson correlation factors between fully sampled and accelerated data. This is the first application of the sensitivity encoding algorithm to characterize the structure of the water resonance at high spatial resolution.

A number of studies have given evidence that water from different subvoxelar compartments (e.g., extra- versus intracellular, or intravascular versus extravascular water) gives rise to distinct MR signal components and a non-Lorentzian water resonance structure (1–5). High spectral and spatial resolution (HiSS) imaging (5) allows imaging of these different compartments, and this improves image quality and provides access to subvoxelar physiologic and/or anatomic information. This is valuable for diagnosing cancer and other diseases. HiSS has been used to improve anatomic and functional MRI, particularly in the breast (4–6), but also in brain (7,8) and prostate (9,10). The advantage of HiSS lies in being able to image the structure of the water resonance with high spectral resolution, and it is being developed for this purpose (5,6,11,12). Echo-planar spectroscopic imaging (EPSI) is used to produce detailed water spectra for each image voxel. The images proportional to water resonance peak height were shown to provide higher contrast, superior fat suppression, and better anatomic detail than standard clinical images (6), and the use of more detailed structural information in the water resonance is in development. Specifically, the asymmetry of the water resonance has been correlated to tumor vasculature in rodent models (13) and non-Lorentzian spectral features have been correlated with biopsy-proven diagnosis (11).

The main challenge HiSS imaging faces is its long acquisition time. Various methods (e.g., more efficient k-space sampling) could increase its speed, and one promising approach is the application of parallel imaging, specifically through sensitivity encoding (SENSE) (14). This technique is particularly pertinent to breast imaging, where the geometry allows for laterally extended placement of coil elements and thus potentially high acceleration factors. An acceptable SENSE algorithm must preserve the complex spectral information as this is necessary to obtain the desired advantages of HiSS. Thus, a SENSE algorithm that works reliably with large complex datasets is necessary. Here, we demonstrate the application of the SENSE algorithm to complex datasets obtained using a high-resolution EPSI sequence, with 7-and 16-channel dedicated breast phased-array coils. We compare (a) water resonance peak height images and (b) water resonance asymmetry images obtained using the SENSE algorithm to images obtained without SENSE acceleration and find that spectral information is well preserved at acceleration factors of up to 6 and 4, respectively.

The present work builds on earlier work by Lin et al. (15), which demonstrated application of the SENSE algorithm to an EPSI-based chemical shift imaging sequence on an eight-channel head coil. The focus of this previous work was low spatial resolution imaging of metabolites present at low concentration, and the goal was to measure metabolite concentrations using LCModel (16) fitting of the full proton spectrum. In the present work, much higher spatial resolution images were acquired, and the detailed spectral features of the water resonance were studied. Because the water resonance imaged with HiSS is isolated and well resolved, we were able to characterize it over a wide 0.83 parts per million (ppm) interval around the peak and study its structure. Specifically, resonance asymmetry images were constructed and characterized at increasing acceleration factors. This is a novel result: it is the first time SENSE has been used to accelerate high-resolution imaging of water resonance spectral structure.

MATERIALS AND METHODS

Data Acquisition

HiSS data were acquired on a 1.5-T clinical Philips Achieva MRI scanner (Philips Healthcare, Cleveland, OH), equipped with 16-channel receiver capability. The HiSS sequence was implemented via a software patch developed in collaboration with Philips Healthcare, enabling a high spatial- and spectral-resolution EPSI-based imaging sequence (17,18). The EPSI sequence was composed of slice selection (with no water suppression), phase encoding, and a series of gradient echo acquisitions produced using trapezoidal gradient pulses with alternating polarity. Thus, a single phase-encode line is acquired at multiple echo times. At the end of the echo train, a “crusher” gradient was applied to eliminate artifacts due to residual transverse magnetization. Shimming was performed using the standard first-order shim protocol. The time domain data associated with each point in k-space were used to obtain spectroscopic information in each voxel, as described in Data Analysis and Synthesis of HiSS images. The sensitivity maps required for SENSE reconstruction were acquired using a turbo spin echo sequence. The detailed imaging parameters for both HiSS and sensitivity scans are given below. Two dedicated breast coils were used: a 7-channel, open-geometry dedicated breast phased-array coil (Invivo, Orlando, FL) and a 16-channel, closed-geometry Mammotrack™ phased-array coil (Invivo, Orlando, FL). Coil element configuration diagrams are given in Fig. 1. Raw data from the scans were exported, and all processing was done off-line, using in-house software written in IDL™ (ITT Corporation). Spatial and spectral resolution was high enough to prevent significant truncation artifacts.

FIG. 1.

Coil element geometry diagrams are shown for the seven-channel Invivo open geometry dedicated breast phased-array coil (a) and for the 16-channel In vivo closed geometry breast phased-array coil ((b), diagram shown upside down for clarity).

Phantom Scans

The phantom consisted of a 1-inch-diameter plastic sphere filled with water and containing a small amount of gadolinium-diethylenetriamine penta-acetic acid (gadolinium-DTPA), 10-mM deuterated trimethysilyl propionic acid, and 0.1% sodium azide (a preservative). The sphere was mounted on a post and embedded in a 2-L leak-proof Nalgene cylindrical container filled with vegetable oil. Two such containers were used, one in each left/right coil compartment. In one of the phantoms, the water-filled sphere also contained 1-mM phosphocholine, which did not affect the present images. A saline bag was placed next to one of the phantoms so that the left/right phantoms, which appear identical, could be distinguished in reconstructed HiSS images.

Coronal two-dimensional HiSS slices were imaged through the two (left and right) water-filled spheres (nonaccelerated matrix 384 × 384, in-plane resolution 1 × 1mm, slice thickness 2mm, pulse repetition time/echo time = 500/45 ms, flip angle 90°, echo train length 63, echo spacing 1.43 ms, spectral bandwidth/resolution 700/11.1 Hz), for a total nonaccelerated acquisition time of 3 min 12 sec. Data were acquired at acceleration factors R = 1, 2, and 3, with acquisition times reduced by the same factor. The readout gradient was applied in the superior-inferior direction to allow SENSE acceleration in the left/right direction. Reference images for coil sensitivity calculations were acquired using a two-dimensional turbo spin echo sequence (matrix/resolution identical to that of HiSS scan, slice thickness 2mm, pulse repetition time/echo time = 3000/40 ms, turbo factor 16, acquisition time 2 min 36 = sec). The sensitivity scans were not speed optimized. The seven-channel In Vivo dedicated breast phased-array coil was used for phantom scans.

Human Volunteer Scans

Four female volunteers were recruited from the university staff and through the University High-Risk Breast Clinic. Subjects were scanned under a protocol approved by the institutional review board after informed consent had been obtained. For three patients, HiSS scans were incorporated into the standard clinical examination and taken after contrast agent administration. One volunteer participated in a separate research-only scan, where no contrast agent was administered.

Axial two-dimensional HiSS slices through both breasts were acquired from human volunteers (non-accelerated matrix 256 × 384, in-plane resolution 1 × 1mm, slice thickness 2mm, pulse repetition time/echo time = 500/45 ms, flip angle 60 or 90°, echo train length 63, echo spacing 1.43 ms, spectral bandwidth/resolution 700/11.1 Hz), for a total nonaccelerated acquisition time of 3 min 12 sec. Data were acquired at acceleration factors R = 1, 2, and 3, with acquisition times reduced by the same factor. Readout gradient was applied in the anterior/posterior direction to allow SENSE acceleration in the left/right direction. Reference images for coil sensitivity calculations were acquired using a two-dimensional turbo spin echo sequence (matrix/resolution identical to that of HiSS scan, slice thickness 2 or 5mm, pulse repetition time/echo time = 3000/40 ms, turbo factor 16, acquisition time 2 min 36 sec). The sensitivity scans were not speed optimized. Both the seven-channel In Vivo and the 16-channel Mammotrack In Vivo dedicated breast phased-array coils were used for volunteer scans, and each volunteer received one scan only, using one of the coils.

SENSE Reconstruction

SENSE reconstruction, following the work of Pruessmann et al. (14), was applied to each “gradient echo image” obtained along the EPSI echo train (see “Data Acquisition”). No noise cross-correlation measurements were performed, resulting in a small signal-to-noise penalty (14) but a largely simplified experimental setup. The sensitivity values outside of the object were set to zero, and a small area of up to five voxels was extrapolated around the object or patient tissue boundary. Data were acquired at acceleration factors R of 1, 2, and 3. Decimation of the full k-space data was also implemented, achieving R of up to 6. The SENSE reconstruction resulted in a set of 63 full-matrix size images, acquired at different echo times; a time-domain Fourier transform was applied to obtain proton spectra in each voxel, as described below. The geometric factor g was also calculated.

Data Analysis and Synthesis of HiSS Images

A two-dimensional Fourier transform in the two k-space directions was applied to the raw HiSS data. This resulted in a series of high-resolution images acquired at different echo times, to which the complex SENSE algorithm was applied. A third Fourier transform in the temporal direction resulted in a high-resolution proton spectrum assigned to every voxel in the image, after a phase correction in the temporal dimension was introduced to correct for spectral Nyquist ghosting, intrinsic to EPSI data. This ghosting is due to the fact that the k-space/temporal data are not acquired on a rectangular Cartesian grid as there is phase evolution during the echo acquisition. The above resulted in proton spectra in which only the water and fat resonances were present and were well resolved. Due to the high spatial resolution, the signal-to-noise ratio (SNR) of HiSS acquisitions is too low for detection of metabolite peaks. The water and fat resonances were identified in the following manner. The highest-intensity spectral component (peak) was identified in each voxel, and the highest-signal voxel was identified. Based on the frequency offset from this seed pixel, the highest peaks in neighboring voxels are classified as either the same or different (water/fat) chemical species as that in the seed voxel, and this process is repeated using a region-growing program until all the voxels have been classified. When the first voxel bearing a highest peak of the species different from the seed voxel is identified, the species are finally classified as either water or fat, based on the frequency shift between the two. Frequency wraparound due to the periodic nature of the Fourier transform was accounted for. This process produced a fat and water spectral frequency map. This robust algorithm relies on the fact that there are no sharp macroscopic gradients within the breast and is described in detail in an earlier publication (19).

Following water/fat peak identification, the proton spectrum in each voxel was fitted to a dual-Lorentzian functional form that included a constant baseline. The baseline and the fit to the fat spectral line were subtracted from the acquired spectrum, and thus the pure water resonance signal was obtained. Since it is difficult to phase water and fat resonances simultaneously, phasing was applied to the water resonance only.

After the pure water resonance signal was obtained in each voxel, “water resonance peak height” images were generated, proportional to the peak intensity of the water resonance in each voxel. These images exhibit a mixture of T₁ and T₂ weighing, excellent fat suppression, and high dynamic range and can be used to improve breast cancer diagnosis (6). Asymmetry of the water peak has been shown to correlate with tumor vascularity because of the effect of paramagnetic deoxyhemoglobin on water resonance line shape (13) and is calculated as the absolute difference of the left- and right-hand-side integrals around the peak of the resonance, relative to the integral of the entire spectral line. A spectral interval of 55.5 Hz (0.83 ppm; five spectral bins) centered around the water peak is included in the integral. As small phasing errors will result in large errors in asymmetry calculation, magnitude spectra were used to generate asymmetry images. Correlation coefficients of R = 1 and R = 2, 3, 4, and 6 data were calculated using the Pearson correlation coefficient (r) to evaluate fidelity of peak height and asymmetry values obtained from accelerated data. Ideally, different datasets, acquired using different SENSE acceleration factors, would be compared to evaluate the SENSE algorithm. However, the correlation between different datasets is very sensitive to small changes in the spectra due to motion. To avoid problems caused by motion, as well as noise effects arising from nonsimultaneous acquisition, in correlation calculations, data acquired at SENSE factors R > 1 were simulated by decimating the full R = 1 dataset. The full k-space dataset was decimated at one-half, one-third, one-fourth, and one-sixth rate to simulate undersampling that occurs in SENSE acquisitions. Only voxels with a water signal integral larger than 0.025 in and around an invasive ductal carcinoma (IDC) lesion were used in the asymmetry calculations. (The peak height values are measured on an arbitrary scale, and this number is given for comparison to values in Fig. 6.)

FIG. 6.

Correlation plots of water resonance peak height values obtained from fully sampled HiSS data and those obtained at acceleration factors R = 2, 3, 4, and 6 are shown in (a–d). A region of interest surrounding an IDC lesion (shown in Fig. 5) is used. The Pearson correlation coefficients (r) are noted on the plots. Data were obtained by decimating the nonaccelerated HiSS dataset so as to not include patient motion and noise contributions. There is virtually no degradation in peak height values at higher acceleration factors, as evidenced by high Pearson correlation coefficients.

RESULTS

Figure 2 shows the image produced from the first gradient echo (in the echo train) image (echo time ~1.5 ms) of the water/oil phantom, obtained using the seven-channel breast coil, reconstructed using the standard sum-of-squares reconstruction, and using the SENSE algorithm for acceleration factors R = 1, 2, and 3. The absolute and phase values of SENSE-reconstructed images are shown separately, showing some SNR degradation with increasing R and demonstrating that the phase information is preserved at higher acceleration factors. This is true for all echo times in the echo train, and thus the complex time-domain data and the resulting spectral information are preserved with R > 1. The proton spectra in the phantom show single, narrow lines for water and oil signal, with the width, position, and intensity preserved at higher acceleration factors. The fourth column shows the absolute value of the difference of absolute images between R = 2 and 3, and R = 1, multiplied by 2, on the same as column 2. The geometric factor g was very close to 1 (1.00 < g < 1.25) throughout the phantom volume, indicating minimal SNR degradation due to coil geometry.

FIG. 2.

T₁-weighted gradient echo (first echo) images of a water/oil phantom, obtained at acceleration factors of R = 1, 2, and 3, in a seven-channel dedicated breast coil are shown. The standard sumof-squares (SOS) reconstruction is shown in the first column. The second and third columns show the absolute and phase images of the complex-SENSE reconstructed data, demonstrating some SNR degradation and preservation of complex phase at R > 1. The fourth column shows the absolute value of the difference of absolute images between R = 2 and 3, and R = 1, multiplied by 2, on the same scale as column 2. The g factor, shown in the last column is close to 1, indicating minimal SNR degradation due to geometric factors. The grayscale inset describes values of the g factor.

In Fig. 3, the same information as in Fig. 2 is shown for a single axial slice taken from a healthy volunteer, using the seven-channel phased-array coil. The geometric factor g is almost identical to 1 for R = 2 and, at R = 3, has low (<2) values throughout the majority of breast area. The spectra obtained from a human volunteer show both water and fat resonances, sometimes in the same voxel, and occasionally structure in the water resonance can be observed. This information, and thus HiSS-derived images, is preserved at higher acceleration factors.

FIG. 3.

The same information as show in Fig. 2 is shown, obtained for a healthy human volunteer, in a seven-channel dedicated breast coil. The complex phase is preserved at higher acceleration factors, and the g factor, although higher than in the phantom images, is still very low in the majority of the breast tissue.

This is illustrated in Fig. 4, where water peak height images, derived from HiSS data obtained from a volunteer with a large IDC lesion, using the 16-channel phased array coil, are shown for acceleration factors R = 1, 2, and 3 (Figs. 4a–c). There are no significant differences between these water peak height images. In Figs. 4d–g, sample phased water resonance lines for four voxels with predominantly water-bearing tissue (white arrowheads, anterior to posterior location, respectively) are shown for R = 1 (solid line), R = 2 (dashed line), and R = 3 (dotted line). The variability of the water resonance structure reflects the variability in the underlying subvoxelar structure of the water compartments, e.g., the volume ratio of intra- versus extravascular compartments, or a cluster of blood vessels with increased deoxyhemoglobin. The acquired spectra retained their width, position, intensity, and structure across acceleration factors. This was the case for both phantom and human volunteer spectra. The small differences between different acceleration factors shown in Fig. 3 are due to noise or patient motion between scans. The spectra obtained by reconstructing k-space data decimated from the full HiSS dataset are virtually identical to those obtained from full k-space data (not shown).

FIG. 4.

Water resonance peak height images obtained from a patient with an IDC lesion, in a 16-channel breast coil, are shown for acceleration factors of R = 1, 2, and 3 ((a), (b), and (c), respectively). Sample spectra in four voxels ((d–g); arrowheads; anterior to posterior, respectively) are shown, as obtained at R = 1 (solid lines), 2 (dashed lines), and 3 (dotted lines). Variation in water resonance shape is evident and preserved over different acceleration factors.

Water resonance asymmetry values calculated for acceleration factors of R = 1, 2, 3, 4, and 6 for a patients with an IDC lesion are shown in Fig. 5. The 16-channel breast coil was used to acquire these data. In Fig. 5a, the water peak height image (acquired at R = 1) is shown, with the arrowhead pointing to the lesion. In Figs. 5b–f, the same water peak height image, magnified by a factor of 2, is shown in an area around the lesion, and a color-coded map of water resonance asymmetry values is over-laid, with the color-coded scale shown on the right. Only voxels for which average water signal was larger than 0.025 are displayed. (The peak height values are measured on an arbitrary scale, and this number is given for comparison to values in Fig. 6.) There is very high agreement between the asymmetry values calculated at different acceleration factors. Due to high sensitivity to patient motion, the asymmetry values are calculated from k-space data obtained from decimated full k-space datasets, rather than accelerated data acquired at different time points.

FIG. 5.

Water peak height image of a patient with an IDC lesion (arrowhead) is shown in (a), and the same image, enlarged by a factor of 2, is shown in (b–f). Overlaid are water resonance asymmetry maps calculated at R = 1, 2, 3, 4, and 6, shown on a color-coded scale (right). Due to very high sensitivity of asymmetry images to patient motion, these images were reconstructed from data obtained by decimation of the fully sampled k-space data. There is very little change in the asymmetry images after decimation.

We calculated Pearson correlation coefficients (r) of water peak height and asymmetry values obtained at different acceleration factors, in a region of interest surrounding the lesion shown in Fig. 5. In order to exclude noise and patient motion from the analysis, decimated datasets obtained from full k-space datasets were used for R > 1. Figure 6 shows the correlation plots of water peak height values, calculated at R = 1 and at R = 2, 3, 4, and 6, respectively. The relationship is linear, with Pearson correlation coefficients of 1.00, 1.00, 1.00, and 0.99, respectively. Figure 7 shows similar correlation plots of water resonance asymmetry values. The Pearson correlation coefficients are 0.96, 0.95, 0.92, and 0.78, respectively. The Pearson coefficients are high, despite the noise amplification intrinsic to k-space data undersampling. There is virtually no degradation of peak height data and little degradation of asymmetry data for R = 2–4.

FIG. 7.

Correlation plots of water resonance asymmetry values obtained from fully sampled HiSS data, and those obtained at acceleration factors R = 2, 3, 4, and 6 are shown in (a–d). A region of interest surrounding an IDC lesion (shown in Fig. 5) is used. The Pearson correlation coefficients (r) are noted on the plots. Data were obtained by decimating the nonaccelerated HiSS dataset so as to not include patient motion and noise contributions. There is only a low level of degradation in asymmetry values at R = 2, 3, and 4, as evidenced by high Pearson correlation coefficients.

DISCUSSION

We demonstrate the application of the SENSE algorithm to complex HiSS datasets and demonstrate that two quantitative measures used to describe the water resonance, the water peak height and the asymmetry of the resonance peak, are preserved at acceleration factors up to R = 6 and 4, respectively. The phase of the HiSS signal along the acquired echo train, and hence the structure of the water resonance and the information on subvoxelar anatomy and physiology that it carries, is preserved. We were able to acquire single-average HiSS images with adequate SNR even at high SENSE acceleration factors. The high Pearson correlation coefficients between R = 1 and higher R for water peak height and asymmetry values are encouraging. They show that the effect of noise amplification, intrinsic to k-space undersampling, is not strong, except for asymmetry values at R = 6. These calculations were done in breast tissue and not in axillary regions that have a lower SNR due to the coil geometry.

Water peak height HiSS images have been shown to detect breast lesions prior to contrast agent administration, even at relatively low spectral resolutions of ~10 Hz (19). Thus, full-breast precontrast HiSS scans could be used to guide subsequent fast dynamic contrast-enhanced MRI scans by identifying suspicious lesions so that higher spatial and/or temporal resolution can be prescribed in these areas. In addition, HiSS could be useful for patients for whom contrast agent administration is contraindicated (20). The application of the complex SENSE algorithm will make full breast bilateral HiSS coverage possible in clinically reasonable times, i.e., under 10 min, although interleaved k-space sampling may be necessary to preserve the long pulse repetition time and thus higher SNR. In the experiments described here, we were not able to perform a full-breast bilateral scan as the large amount of data is a technical issue in its own right (e.g., disk space requirements, data export and transfer). Future hardware upgrades should resolve this problem.

HiSS acquisition at spectral resolution high enough to resolve individual components of the water resonance (~ 2.5–5 Hz) would be prohibitively long in full-breast coverage applications but would be useful in limited coverage, e.g., for lesion characterization. SENSE application would allow coverage of a wider region. Several slices through a lesion can be acquired at such high spectral and spatial resolution, and water resonance characteristics, e.g., asymmetry—shown to be correlated with tumor vascularization in rodent studies (13)—can be calculated. Other water resonance characteristics, e.g., local amplitude of static field shifts, peak splitting, and differences in spectral component contrast, can be calculated but were not included in this work.

Earlier work by Lin et al. (15) demonstrated an application of SENSE to an EPSI-based spectroscopic sequence, but several differences arise in treating HiSS data. The earlier work concerned high-resolution water-and fat-suppressed spectra of metabolites in large (7 × 7mm in-plane resolution; slice thickness not given) voxels, while HiSS images nonsuppressed water and fat spectra in small (1 × 1 × 2mm) voxels. While water peak height images are similar in concept to the metabolite concentration maps shown in the earlier work, the asymmetry calculations are unique to HiSS imaging: metabolite peaks are typically treated as simple Lorentzian lines with no internal structure. The asymmetry values showed a stronger SNR dependence with increased acceleration factors, as compared to the water peak height values, thus underscoring the need for a separate evaluation of performance of this important quantitative measure under the SENSE algorithm.

CONCLUSIONS

In conclusion, we demonstrate for the first time an application of SENSE algorithm to characterization of water resonance intensity and structure, and this was achieved at high spectral and spatial resolution. Phantom studies were used to validate the algorithm. In vivo imaging demonstrated that accelerated HiSS imaging can provide water resonance peak height and water resonance asymmetry measures that are in good agreement with fully sampled data, up to the acceleration factors of R = 4. This algorithm can allow fast bilateral water peak height imaging of human breast, notably prior to contrast administration, for detection of breast lesions. Alternatively, fast high-spectral-resolution scans can be used to characterize known lesions, using water resonance asymmetry or other quantitative measures.