Portal de Periódicos da CAPES

One-second MRI of a three-dimensional vocal tract to measure dynamic articulator modifications

2016; Wiley; Volume: 46; Issue: 1 Linguagem: Inglês

10.1002/jmri.25561

1522-2586

Michael Burdumy, Louisa Traser, Fabian Burk, Bernhard Richter, Matthias Echternach, Jan G. Korvink, Jürgen Hennig, Maxim Zaitsev,

Advanced Neuroimaging Techniques and Applications

Journal of Magnetic Resonance ImagingVolume 46, Issue 1 p. 94-101 Technical DevelopmentFree Access One-second MRI of a three-dimensional vocal tract to measure dynamic articulator modifications Michael Burdumy PhD, Corresponding Author Michael Burdumy PhD [email protected] University Medical Center Freiburg, Department of Radiology, Medical Physics, Freiburg, Germany University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, GermanyAddress reprint requests to: M.B., University Medical Center Freiburg, Department of Radiology, Medical Physics, Breisacher Str. 60a, 79106 Freiburg, Germany. [email protected]Search for more papers by this authorLouisa Traser MD, Louisa Traser MD University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, Germany Department of Oto-Rhino-Laryngology, Head and Neck Surgery, University Medical Center, Freiburg, GermanySearch for more papers by this authorFabian Burk MD, Fabian Burk MD University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, GermanySearch for more papers by this authorBernhard Richter MD, Bernhard Richter MD University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, GermanySearch for more papers by this authorMatthias Echternach MD, Matthias Echternach MD University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, GermanySearch for more papers by this authorJan G. Korvink PhD, Jan G. Korvink PhD Institute of Microstructure Technology, Karlsruhe Institute of Technology, Karlsruhe, GermanySearch for more papers by this authorJürgen Hennig PhD, Jürgen Hennig PhD University Medical Center Freiburg, Department of Radiology, Medical Physics, Freiburg, GermanySearch for more papers by this authorMaxim Zaitsev PhD, Maxim Zaitsev PhD University Medical Center Freiburg, Department of Radiology, Medical Physics, Freiburg, GermanySearch for more papers by this author Michael Burdumy PhD, Corresponding Author Michael Burdumy PhD [email protected] University Medical Center Freiburg, Department of Radiology, Medical Physics, Freiburg, Germany University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, GermanyAddress reprint requests to: M.B., University Medical Center Freiburg, Department of Radiology, Medical Physics, Breisacher Str. 60a, 79106 Freiburg, Germany. [email protected]Search for more papers by this authorLouisa Traser MD, Louisa Traser MD University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, Germany Department of Oto-Rhino-Laryngology, Head and Neck Surgery, University Medical Center, Freiburg, GermanySearch for more papers by this authorFabian Burk MD, Fabian Burk MD University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, GermanySearch for more papers by this authorBernhard Richter MD, Bernhard Richter MD University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, GermanySearch for more papers by this authorMatthias Echternach MD, Matthias Echternach MD University Medical Center Freiburg, Institute of Musicians' Medicine, Freiburg, GermanySearch for more papers by this authorJan G. Korvink PhD, Jan G. Korvink PhD Institute of Microstructure Technology, Karlsruhe Institute of Technology, Karlsruhe, GermanySearch for more papers by this authorJürgen Hennig PhD, Jürgen Hennig PhD University Medical Center Freiburg, Department of Radiology, Medical Physics, Freiburg, GermanySearch for more papers by this authorMaxim Zaitsev PhD, Maxim Zaitsev PhD University Medical Center Freiburg, Department of Radiology, Medical Physics, Freiburg, GermanySearch for more papers by this author First published: 09 December 2016 https://doi.org/10.1002/jmri.25561Citations: 19AboutSectionsPDF ToolsRequest permissionExport citationAdd to favoritesTrack citation ShareShare Give accessShare full text accessShare full-text accessPlease review our Terms and Conditions of Use and check box below to share full-text version of article.I have read and accept the Wiley Online Library Terms and Conditions of UseShareable LinkUse the link below to share a full-text version of this article with your friends and colleagues. Learn more.Copy URL Abstract Purpose To enable three-dimensional (3D) vocal tract imaging of dynamic singing or speech tasks at voxel sizes of 1.6 × 1.6 × 1.3 mm3 at 1.3 s per image. Materials and Methods A Stack-of-Stars method was implemented and enhanced to allow for fast and efficient k-space sampling of the box-shaped vocal tract using a 3 Tesla MRI system. Images were reconstructed using an off-line image reconstruction using compressed sensing theory, leading to the abovementioned spatial and temporal resolutions. To validate spatial resolution, a phantom with holes of defined sizes was measured. The applicability of the imaging method was validated in an eight-subject study of amateur singers that were required to sustain phonation at a constant pitch, past their comfortable expiratory level. A segmentation of the vocal tract over all phonation time steps was done for one subject. Anatomical distances (larynx position and pharynx width) were calculated and compared for all subjects. Results Analysis of the phantom study revealed that the imaging method could provide at least 1.6 mm isotropic resolution. Visual inspection of the segmented vocal tract during phonation showed modifications of the lips, tongue, and larynx position in all three dimensions. The mean larynx position per subject amounted to 52–85 mm, deviating up to 5% over phonation time. Parameter pharynx width was 32–181 mm2 on average per subject, deviating up to 16% over phonation time. Visual inspection of the parameter course revealed no common compensation strategy for long sustained phonation. Conclusion The results of both phantom and in vivo measurements show the applicability of the fast 3D imaging method for voice research and indicate that modifications in all three dimensions can be observed and quantified. Level of Evidence: 2 Technical Efficacy: Stage 1 J. MAGN. RESON. IMAGING 2017;46:94–101 Over the past years, the imaging of the vocal tract by MRI has become a valuable tool for clinical and scientific research of speech and singing.1, 2 MRI avoids ionizing radiation, provides excellent contrast of the soft tissues of the vocal tract, and has only minimal effects on the singing or speaking process itself. However, there is a trade-off between imaging speed and spatial resolution, which makes MRI inherently slower than other imaging methods, such as computed tomography or ultrasound. In vocal tract imaging, the required temporal resolutions are often defined by the observed motion or articulator state.1 For the imaging of a single mid-sagittal slice of the head, clinical methods have been applied at temporal resolutions of up to 10 frames-per-second (fps).3, 4 Research methods that use non-Cartesian trajectories and parallel imaging have reached higher imaging speeds of 21 fps, 30 fps, and up to 102 fps.5-8 However, imaging only a single slice limits the applications to voice research. If acoustic resonance properties of the vocal tract are to be examined, complete and detailed volume data of the vocal tract are necessary. Furthermore, most speech and singing motions involve modifications of articulators in regions that are not displayed in a mid-sagittal slice and occur independently in spatially separated regions.9, 10 To acquire volume data of performing voice users, such as the singing of scales or pitch glides, the imaging method needs to be as fast as possible to display different configuration states. From the first studies that required more than four minutes measurement time,11 clinical protocols have continuously been proposed to accelerate the imaging of the vocal tract in all three dimensions,12-14 up to an acquisition time of approximately 13 s at a high spatial resolution (1.0 × 1.6 × 1.3 mm3).15 Nonclinically available methods have further accelerated the imaging process. In their study on subjects with sleep apnea, Kim et al acquired the vocal tract at 1.5 × 1.5 × 2 mm3 spatial resolution within 7 s.5 In a follow-up study, they showed the feasibility of a customized receiver coil to further improve the method to an isotropic spatial resolution of 1.25 mm within 9 s.16 Fu et al made use of advances in sparse-sampling theory and reconstructed volumetric data within 6 ms update rate at a coarse spatial resolution of 2.2 × 2.2 × 5 mm,17 requiring their subjects to repeat the phrases multiple times. Regarding the minimum required spatial resolution of vocal tract imaging, for 2D speech imaging 1–2 mm in-plane for measuring velum movement or less than 3.5 mm for tongue movement were recommended.2 During singing, the supra-laryngeal tube can become very narrow (∼2 mm diameter).18 Hence, especially if acoustic properties of the vocal tract are to be analyzed, resolutions less than 2 mm isotropic resolution have been recommended.8, 9, 19 As can be seen from the abovementioned methods on 3D vocal tract imaging, there is still a need to improve imaging, so that high temporal and spatial resolutions can be achieved at the same time. An isotropic voxel resolution of at least 2 mm should be sufficient to represent fine structures and narrow passages of the vocal tract while an imaging speed in the 1-s range would allow for stable phonation of difficult high notes or dynamic singing of scales in one breath. Expanding on earlier studies with radial trajectories,8, 20, 21 we propose and validate an imaging method that generates a volume image of the complete vocal tract at high spatial resolution in 1.3 s and thus enables the research of dynamic singing or sustained speech tasks. Materials and Methods After approval of the local institutional review board, all MRI measurements were performed on a 3 Tesla (T) system (Prisma, Siemens Healthcare, Germany), using the manufacturer's 64-channel head/neck coil and an optical microphone for sound recording (OptiMRI, MR Confon GmbH, Germany). The subjects gave informed written consent before the measurements. Sequence Design The custom sequence was accelerated for speed such that dynamic singing tasks, such as scales, could be depicted in short acquisition times. In accordance with previous studies on singers, the required spatial resolution was required to be at least 2 mm isotropic.8, 18, 19 The vocal tract is approximately box-shaped, its dimensions are much smaller in left–right (l-r) direction, approximately 65 mm, than in the directions of the sagittal plane (head–feet, h-f; anterior–posterior, a-p), approximately 150 mm. Thus, the resolution in l-r direction can be increased and acquisition time can be reduced by choosing a slab-selective excitation and phase encoding in this direction, rather than measuring the complete head with frequency gradients (Koosh-Ball trajectory). Consequently, a radial Stack-of-Stars sequence with golden angle rotation sampling was implemented on the scanner (compare with Feng et al and Block et al).20, 21 As basis, we chose a previously reported radio-frequency-spoiled radial GRE sequence that was extended to 3D imaging.8 Gradient timing, maximum amplitude, and rise times were taken from the time-optimized 2D sequence. A phase encoding (slice selection, SS) gradient, as used in regular Cartesian GRE sequences, was added in l-r direction and phase rewinder gradients were added. Figure 1a shows the sequence timing diagram of the sequence: RF-spoiled GRE Stack-of-Stars: TR = 2.9 ms, TE = 1.4 ms, FA = 6 °, FOV 200 × 200 × 62 mm3, voxel resolution 1.6 × 1.6 × 1.3 mm3, BW 1500 Hz/pixel. Figure 1Open in figure viewerPowerPoint Sequence diagram and example of Stack-of-Stars trajectory using 8 partitions, 10 projections in k-space center and a peripheral under-sampling of 4. a: Sequence timing diagram for one repetition time. Depicted are the RF pulse, ADC acquisition window, slice selection gradient GSS and the two (radial) frequency encoding gradients GFE1/2. b: Plot of the trajectory in frequency and phase directions, kFE and kPE, respectively. All projections of one partition are depicted with the same color. c: Plot of the trajectory along phase direction and time of sampling. The crosses mark the occurrence of sampling at the kPE-coordinate at the given time t. The blue and red boxes mark the projections of the k-space center partition and of the most peripheral, respectively. For demonstration purposes, a green line is shown that highlights the linearly decreased number of projections toward the k-space periphery. Also, the figure demonstrates that the projections of the peripheral partitions are acquired in the middle of each time window. The planes (partitions) in k-space along the phase encoding direction, which are defined by the discrete phase encoding steps, are sampled sequentially with one radial projection each. Thus, a complete coverage of all partitions is reached in the shortest possible time. Additional projections in each partition are then rotated by the Golden Angle with respect to the previous projection angle. To maximize information content during one acquisition, none of the projections of one partition had the same angle as the projections of any other partition. The complete measurement was split into temporally separated parts, referred to as time windows. With the given sequence parameters, 9651 projections would be required to fully sample k-space (Nyquist-sampling), leading to an acquisition time of 28 s. By evenly reducing the number of projections in each partition (marked by reduction factor ), measurement time can be reduced. To even further accelerate the acquisition of one time window, in our acquisition scheme the partitions were sampled less densely depending on their distance to the center partition. This additional under-sampling was defined in the protocol by the peripheral under-sampling factor . Hence, the actual under-sampling at the periphery was . Given a defined number of projections at the center partition , the distance from the center along the partition direction in k-space and the maximum distance , the number of projections at any partition calculates as which describes a linear decrease. Regarding the actual acquisition time of a projection during one time window, projections of the outer partitions were sampled closer to the middle of the time window. Figures 1b,c depict the under-sampling scheme in detail, where for viewing purposes the number of partitions and number of center projections were set to 8 and 10, respectively. The under-sampling of high image frequencies can have an effect on the rendering of fine structures of the image and was evaluated in a phantom measurement. Note that due to the irregular sampling of k-space, the time windows were predefined for the acquisition. Thus, sliding window methods8 could not be used in the image reconstruction. For the above sequence with 48 partitions, 21 center projections ( ) and a peripheral under-sampling factor of were chosen, leading to a total of 456 projections per time window (1.3 s), whereas uniform partition sampling would require 1008 projections (2.9 s). Image Reconstruction In a first step, gradient delays in read-out direction were corrected for by a previously described method (8). The imaging process can be described as a discrete forward problem where S is the measured signal, A is the forward encoding operator that includes coil sensitivities, Fourier encoding and projection onto a grid, x is the original signal density distribution and η describes the acquired noise. With the measured signal S the original image can be calculated by minimizing a functional of the form with any arbitrary number of regularizing operators Li. To tap the potential of parallel imaging, the coil sensitivities of each time window were calculated from coil-by-coil images. These were reconstructed by applying a mildly regularized conjugated gradient reconstruction with cut-out center regions of the raw data, as described in Burdumy et al.8 As proposed in Feng et al,20 a pixel-wise total variation (TV) operator was applied that enforces sparsity in time. Additionally, another TV operator was applied that enforces sparsity in the image domain (see Burdumy et al and Block et al).8, 22 The regularization parameters were set to λtemporal=0.05 and λspatial = 0.03. The minimization problem was solved with the method of nonlinear conjugated gradients in Matlab (release 2014a, MathWorks, Natick, MA). The forward operator A was implemented using the Image Reconstruction Toolbox23 and gpuNUFFT.24 Therein, the Stack-of-Stars trajectory could be defined and the precalculated sensitivity maps of each time window could be incorporated. All images were reconstructed on a server with 96 GB memory, using a single CPU (Intel Xeon X5690, 3.5 GHz) and GPU (nVIDIA Tesla C2075) per dataset. Experimental Setup The spatial resolution defined in the protocol might not be reached by the image reconstruction method, due to under-sampling, regularization, and further peripheral under-sampling in phase encoding direction. Hence, the spatial resolution was evaluated in a phantom measurement. To this end, a previously described water-filled bottle phantom was used.8 The PVC board in the bottle comprises drilled holes of defined diameters that show up in images as bright spots. The bottle was positioned in the iso-center of the magnet, inside the head/neck coil. The bottle was measured with a transversal slice orientation, such that the phase encoding resolution could be evaluated. As the field of view (FOV) is limited in this direction, only the part of the bottle was measured that contains the narrowest (1.6 mm) and widest holes (6 mm) (see Fig. 2 for details). The spatial resolution in frequency-encoding directions alone was already validated in a previous study.8 Figure 2Open in figure viewerPowerPoint Sketch of the PVC board inside the water-filled phantom, also displaying hole diameters and distances between holes. Special care was taken to position the bottle phantom and the slice orientation so as to have all holes aligned on a straight line parallel to the image border, to prevent partial volume effects and interpolation errors. The reconstructed images were imported into Matlab. In each volume data, a partition was found that slices the mid-center of the board. In these 2D slice images, the pixel values along lines that pass through all holes were evaluated. As there are two rows of holes on the PVC board, one line goes through the larger, while the other goes through the narrower holes. The contrast of the holes was evaluated by inspecting the line profiles for distinguishable paired local minima and maxima. To this end, first, pixel intensities were normalized. All pixel values were divided by the mean intensity of a manually set rectangular, fluid-filled region, named foreground (FG) in the following. The FG was set to similar regions for all images of the respective slice orientations. Local minima and maxima were searched for using Matlab's extrema function, with empirical thresholds, so as to omit noisy locations. The minima and maxima were paired for each respective hole and grouped for each hole diameter. For each of these groups, the mean peak-to-peak values from minimum to maximum intensities were calculated. To decide if a hole (and thus the corresponding spatial resolution) could be resolved, the normalized peak-to-peak intensity value should be . To prove the applicability of the method for dynamic voice tasks, 8 subjects (4m, 4f) were measured. To prevent head tilting, the heads of the subjects were loosely fixated by foam pads. The male subjects were required to hold the pitch C4 (f0 = 262 Hz) for as long as possible, whereas the female subjects sang the pitch C5 (f0 = 523 Hz). The reconstructed images present snapshots of the vocal tract configuration, i.e., the average states during one time window of 1.3 s. Modifications faster than this window would be suppressed. Only when the subjects started phonation, was the scan process of the MRI system started. The correct pitch was verified by the examiner using the audio feedback system. The subjects were expected to sustain phonation at different lengths, depending on the air drawn and their individual singing capability. In a first step, the end of phonation was identified for each subject from the images. For one subject and each recorded time step of phonation (in this case n = 16), the vocal tract was segmented manually using ITKSnap's (3.0.4, www.itksnap.org) Snake-ROI method, with a threshold-based presegmentation. The filled binary voxel representations of these segmentations were added onto each other. Hence, the summarized volume contained voxel intensities from 0 to 16, depending on how many times the segmented volume was present at a certain location. Using this representation, regions of vocal tract modifications could be highlighted by a change in voxel intensity. Second, for each subject, the midsagittal slice was identified. In accordance with previous studies, the larynx position (LP), which corresponds roughly to the distance of the glottis to the velum, was measured for each time frame and all subjects (refer to Burdumy et al8 and Fig. 3 for a more detailed description). Figure 3Open in figure viewerPowerPoint Location of landmarks for morphometric measurements in the vocal tract image. Depicted is a mid-sagittal slice of a random subject. The crosses depict the anterior commissure of the larynx, anterior tubercle of the atlas and highest frontal point of the sixth cervical vertebra. Parameter LP is then calculated as depicted. A simplified depiction of parameter PA is given, that requires 3D image data to segment the pharynx area and find the narrowest slice. Third, the narrowest transversal pharynx area (PA) was evaluated for each subject and time step. A fast region-growing segmentation in a manually defined bounding box was applied using MevisLab (2.8, MeVis Medical Solutions AG, Germany). In these surface models, the area of each slice in transversal direction was calculated and the minimum found using Matlab. The location of both parameters is provided in Figure 3, depicting a mid-sagittal image slice taken with the proposed method. The mean and deviation of parameters LP and PA were calculated across subjects. As statistical analysis for the limited number of subjects is difficult, the course of parameters LP and PA was assessed qualitatively by visual inspection of two experienced vocal tract anatomy experts (L.T., otolaryngologist; M.E., otolaryngologist and phoniatrist). Results Figure 4 displays the image slice of the phantom measurement that was used to analyze the spatial resolution in the phase encode direction, wherein the evaluated regions are marked by rectangular boxes. To the right, the normalized line profiles Inorm along the upper and the lower line of holes are depicted, along with the peak-to-peak values Ipp. Although the value steadily declines with lower hole diameters, it remains above , and even the smallest structures (Ø = 1.6 mm) could be resolved. Hence, the actual spatial resolution lies within the magnitude of the nominal resolution ( . Figure 4Open in figure viewerPowerPoint Results of the phantom measurement on spatial resolution in phase direction. Left: Image of the hole phantom. The colored boxes mark the area where line profiles were measured. The mean of area FG was used for normalization of all pixel values. Right: Plots of normalized intensity values Inorm. The background color corresponds to the box edge color of the left image. Calculated extrema are marked by crosses. The peak-to-peak values Ipp (averaged over holes of the same diameter) are displayed by a yellow line. The minimum contrast difference Ithresh = 0.1 is displayed by the dotted red line. The holes of the lower row are 6 and 5 mm in diameter. The holes of the lower row are 2.2, 2.0, 1.8, and 1.6 mm in diameter (from left to right). The results of the first analysis method of one subject are displayed in Figure 5. The varying voxel counts of the summary of segmented air-filled cavities are displayed as a heat map in three orthogonal slices that are located in regions of the vocal tract that are of resonatory interest. Both values of 0 (no voxel occurrence) and 16 (constant voxel occurrence) are displayed opaque, as only changing articulators were of interest. The figure highlights regions of strong articulator modifications, for example the position of the tongue in the sagittal slice, or narrowing of the pharynx and changing of larynx height in the two other orientations. It becomes evident that modifications occur in all three dimensions. Figure 5Open in figure viewerPowerPoint Heat map of anatomical changes during sustained phonation of one subject. The gray anatomical images depict three orthogonal slices of the vocal tract during the first time step, the exact slice locations are marked by dotted lines. The colored regions show the result of the binary addition of the segmented vocal tract from all time steps. Low values mean that this region was rarely occupied, large numbers mark more frequent occupations. Constant occupation and zero occupation are not color-coded, as only modifications were analyzed. Regarding the second analysis method, Figure 6a displays the measured parameter LP for all subjects over time. The plots are arranged to end at the same time point. Thus, the last segment, from 11.9 s to 13.2 s reflects the maximally forced phonation when there was almost no air left. Subject 1 (displayed in Fig. 5) was able to sustain phonation longer than the other subjects. For sake of clarity, the earlier measurements of this subject are omitted from Figure 6. Figure 6Open in figure viewerPowerPoint Plot of parameters larynx position and pharynx area. Note that subjects 1–4 were female, and 5–8 were male. a: Detailed plot of larynx positions of each subject over time. The lines are color coded by subject (see legend in c). As the length of phonation differed between subjects, the plots were arranged to end at the same point. b: Larynx position over subjects. Plotted are the mean and standard deviation of the full phonation for each subject. c: Detailed plot of pharynx area of each subject over time. d: Mean pharynx area over subjects, including standard deviation. No clear course of the graphs can be observed. Some subjects raised their larynx, while others lowered it or kept it at approximately the same height. Figure 6b shows the mean LP over time for each subject. In general, the mean LP is lower for the female singers (52–60 mm) and higher for the male singers (80–85 mm), both due to the given gender-specific anatomies and the different pitch. The magnitude of LP deviation per subject over all phonation times was in the range of 0.2–3.1 mm (0.2%–5%). Three subjects showed a deviation that was greater than the nominal spatial resolution of 1.6 mm. Figures 6c,d display the results of the third analysis method that evaluated PA over time, similarly arranged as the top figures. In contrast to parameter LP, parameter PA is not related to gender and deviates more between subjects. The mean pharynx width per subject lies between 32 and 181 mm2. As for parameter LP, some subjects increased pharynx width continuously over time, others decreased or kept it constant. The LP deviation of each subject during phonation was between 2.8 and 14 mm2 (3%–16%). Discussion Using the proposed imaging method, which includes Golden-Angle Stack-of-Stars acquisition, peripheral under-sampling, and regularized iterative reconstruction, volume images of both a phantom and eight subjects were successfully acquired. The phantom measurement validates that a resolution in phase encoding direction better than 1.6 mm was achieved using only projections within a time window of 1.3 s, corresponding to an average Cartesian under-sampling factor of 14 (radial 21). Because the phantom is unmoving, under-sampling artifacts are removed and spatial resolution is restored by temporal regularization. The small holes within a large homogeneous area, however, present a worst-case scenario for spatial regularization. The results of the phantom measurement can only provide the upper limit of the achievable spatial resolution of in-vivo measurements. Regularization factors should always be chosen as small as possible to reduce spatial and temporal blurring. The measurement of eight nonprofessional singer subjects yielded volume data that could be used to either segment the vocal tract or take morphometric measurements in the anatomy. If untrained singers pass their resting expiratory level (REL) during singing, most often a constriction of the pharynx and upward movement of the larynx can be observed.25 In our data, however, both analyzed parameters failed to show a common strategy of the singers to deal with the difficult sustained phonation, especially toward the end of the strained phonation. This becomes evident in the time course of the two measured parameters LP and PA that differed strongly between subjects. The observed parameters changed in both directions and occurred at all times, either sporadically or continuously over time. The different parameter courses might be related to different compensation strategies or inconsistent behavior related to the untrained state of the measured singers. Yet, the results of the segmentation and measured parameters clearly show that deviations exist and could be resolved by the imaging method, both in in all three dimensions and at any time. This underlines the necessity to acquire volume images of the vocal