The central auditory system starts at the auditory nerve (AN) which conveys action potentials in response to neurotransmitters released by the hair cells in the cochlea. The cochlear nucleus (CN) is the first nucleus of the auditory system receiving information from the ipsilateral cochlea. A next step in the pathway is the superior olivary complex (SOC), in which input from the two cochlear nuclei converge. Neurons from the CN and SOC project to the inferior colliculus (IC) through axons that form the lateral lemniscus (LL). The next step is formed by connections between the inferior colliculus and the medial geniculate body (MGB) of the thalamus, a relay station of several types of information of which the auditory pathway is only one. Fibers leaving the MGB project to the primary auditory cortex (AC). For more detailed information we refer to Ehret and Romand [1]. In addition to this classical auditory pathway there are connections between the auditory system and the limbic system [2]. The limbic system is involved in motivation, mood and emotion [3], and consists of many subsystems, including the hippocampus, the amygloid complex, the cingulate gyrus and the prefrontal cortex [4]. Typical complaints attached to tinnitus such as anxiety, depression, and emotions such as fear indicate the association of the limbic system with tinnitus [5]. Cognitive therapies focus on the reduction of alteration of the emotional content of the percept of tinnitus by habituation [6, 7]. Changes in regional cerebral blood flow (rCBF) and blood oxygenation (BOLD) signal have been reported in the limbic system by several studies [8]. The connection between the auditory system and the limbic system may thus be of importance in the pathology of tinnitus.

The central auditory system starts at the auditory nerve (AN) which conveys action potentials in response to neurotransmitters released by the hair cells in the cochlea. The cochlear nucleus (CN) is the first nucleus of the auditory system receiving information from the ipsilateral cochlea. A next step in the pathway is the superior olivary complex (SOC), in which input from the two cochlear nuclei converge. Neurons from the CN and SOC project to the inferior colliculus (IC) through axons that form the lateral lemniscus (LL). The next step is formed by connections between the inferior colliculus and the medial geniculate body (MGB) of the thalamus, a relay station of several types of information of which the auditory pathway is only one. Fibers leaving the MGB project to the primary auditory cortex (AC). For more detailed information we refer to Ehret and Romand [1]. In addition to this classical auditory pathway there are connections between the auditory system and the limbic system [2]. The limbic system is involved in motivation, mood and emotion [3], and consists of many subsystems, including the hippocampus, the amygloid complex, the cingulate gyrus and the prefrontal cortex [4]. Typical complaints attached to tinnitus such as anxiety, depression, and emotions such as fear indicate the association of the limbic system with tinnitus [5]. Cognitive therapies focus on the reduction of alteration of the emotional content of the percept of tinnitus by habituation ;[6, 7]. Changes in regional cerebral blood flow (rCBF) and blood oxygenation (BOLD) signal have been reported in the limbic system by several studies [8]. The connection between the auditory system and the limbic system may thus be of importance in the pathology of tinnitus.

In this paper we study the characteristics of white matter fiber tracts defining the (classical) auditory pathway, especially the pathways from the IC to the MGB up to the primary auditory cortex. We also investigate the connections between the auditory system and the limbic system, especially the amygdala (AM). Both pathways were studied using diffusion tensor imaging (DTI) methods. Additionally, we investigated possible differences in structural brain connectivity between subjects affected by tinnitus and healthy subjects. Since hemispheric differences have been reported in patients [9-11], also lateralization of DTI findings was investigated.

Diffusion tensor imaging (DTI) is a recently developed MR acquisition modality that enables the measurement of structural organization of tissues [12, 13]. As a powerful and non-invasive technique for in vivo exploration of human tissues, DTI is widely used in various medical fields, especially in brain imaging [14, 15].

DTI is based on the diffusion properties of water molecules in white matter of the brain. The diffusion is limited by the fibrous nature of white matter: the well-organized axon structure, axon membranes, neurofilaments and overall the myelin coating surrounding the neurons induce the displacement of water molecules to occur preferentially along the axon fibers rather than perpendicularly to them [16]. This anisotropic diffusion of water molecules can be measured by an MR scanner, allowing us to infer information on white matter connectivity. Two interesting features of DTI techniques are the ability to derive local information, such as the amount of anisotropy and the principal water diffusion direction in a single brain voxel [17], and the possibility to track fiber bundles from a selected brain area [18].

Although several imaging studies have examined the auditory system and tinnitus [10, 19, 20], only a few of them applied DTI techniques [21-23]. The main reason for the infrequent application of DTI in studies on the auditory system is the poor spatial resolution of DTI images. Partial volume effects lead to underestimation of diffusion anisotropy, and multiple nerve fiber tracts crossing a single brain voxel may disturb the fiber tracking algorithm. According to tracing studies on the macaque [24] using radioactive tracers, we know that the fibers of the auditory pathway intersect motor bundles (in the CN-SOC-IC path) and are close (within millimeter range) to the corticospinal tract (in the IC-MGB path) and do cross the internal capsule [25] when connecting the MGB with the AC.

Standard deterministic fiber tracking techniques only consider the main diffusion direction in each voxel, which is provided by the eigenvector corresponding to the largest eigenvalue of the diffusion tensor in the voxel. This technique is incapable of resolving voxels with multiple fiber directions that occur when two fibers cross each other within a voxel, but also has problems in the presence of gray matter voxels. For this reason, we used probabilistic tractography with multiple fiber orientations [26]; this technique has been shown to provide significant advantages in sensitivity when tracking non-dominant fiber populations and allowed us to track the auditory paths, which are impossible to detect with deterministic tractography. Although probabilistic tractography does not allow the visualization of the actual fiber bundles, it outputs a whole-brain probabilistic connectivity map for localizing white matter tracts. Using this method, we investigated the connections of the auditory pathways and the connections between the auditory system and the limbic system in a focused per-subject analysis. We also studied differences in tractography results of these paths in subjects with and without subjective tinnitus.

The auditory system and tinnitus have been studied using functional imaging techniques [10, 19 , 20]. DTI methods have been used in a few studies on the auditory system, but in most cases only scalar values derived from DTI analysis, such as the fractional anisotropy (FA) index, describing the amount of anisotropy per voxel, were considered. For instance, Lee et al. [23] performed a study where the FA index in several areas of the brain was determined. Reduced FA values in tinnitus patients were found in the left frontal Arcuate Fasciculus and the right parietal Arcuate Fasciculus. However, the classical auditory pathway was not studied. Lutz et al. [22] were able to detect changes in FA maps in cortical and subcortical auditory regions, in relation with the age of the subjects (although the study did not concern tinnitus): elder subjects showed bigger FA values in IC and lower FA values in the auditory radiation and in the temporal gyri. Moreover, Lee et al. [27] showed differences in FA values between subjects with conductive hearing loss and subjects with profound sensorineural hearing loss, reporting that DTI findings of subjects with profound sensorineural hearing loss revealed neural damages in the auditory pathway. De Groof et al. [28] used DTI to study the auditory and vocal system in birds. They discerned a number of song control and auditory nuclei, and discriminated the tracts running from and to these nuclei. Using DTI fiber tracking techniques, Upadhyay et al. [29] studied the connectivity patterns within the primary auditory cortex.

Different methods are available for exploring differences in brain structure between two groups. Using voxel-based morphometry (VBM), brain data from different subjects are aligned to a reference volume so that voxel-wise statistics can be computed [30]. This method has been used for comparing diffusion tensor data [31, 32] and provides a whole-brain picture of structural differences; the major drawback of this technique is that the results could be strongly influenced by the quality of the alignment among the brain data and by the spatial smoothing usually applied to the data [33]. Tract-based spatial statistics [33] is another tool for comparing DTI data among subjects. This technique seems to be very robust to misalignment of brain data; however, it only allows comparisons of high-FA white matter voxels and is not able to analyze a specific fiber bundle per se.

DTI data were acquired from 25 subjects: 15 healthy subjects and 10 subjects with tinnitus. The group of healthy volunteers consisted of 12 men and 3 women; the age of these varied from 26 to 75, with a mean of 46 and a standard deviation of 16. The subjects affected by tinnitus (9 men and 1 woman) had ages varying from 30 to 70, with an average of 49 and a standard deviation of 12. All the subjects were right-handed. Six of the patients suffered from bilateral tinnitus. Four suffered from tinnitus in the left ear only. Fig. (1A) shows the hearing levels of the patients at different octave frequencies; Fig. (1B) shows the severity of tinnitus as perceived by the patients during rest, according to a visual analogue scale (VAS) ranging from 1 to 10. The frequency and loudness level of tinnitus were determined by a matching procedure. Frequency matching was performed with an external tone presented at the non-tinnitus ear (unilateral tinnitus) or at both ears (bilateral tinnitus) at a comfortable level. The loudness level was then determined by adjusting the level of this tone to match the tinnitus loudness. All subjects gave written informed consent for their participation in the study.

Fig. (1). A: Hearing levels (in dB HL) of the tinnitus patients at several frequencies. Circles represent the average levels for the right ear, crosses represent average levels for the left ear. The whiskers show standard deviations. B: Severity (VAS), frequency and loudness level (indB SL) of tinnitus perceived by the patients. An “X” is used when data were not available.

All the imaging experiments were performed on a 3T Philips Intera MRI scanner. DTI was performed using a diffusion weighted spin-echo, echo-planar imaging technique. The DTI parameters were as follows: field of view = 240 × 240 mm; matrix size = 128 ×128 no. of slices = 50; imaging resolution = 1.85 × 1.85 × 2 mm; TR = 5485 ms; TE = 74 ms. In total, 61 volumes were acquired per subject, one without diffusion weighting (b = 0 s/mm²) and 60 volumes with diffusion weighting (b = 800 s/mm²) along 60 noncollinear directions. To correct for susceptibility artefacts, two acquisitions were used: one with fat-shift direction in the posterior direction (APP) and one in the anterior direction (APA).

An anatomical scan was acquired to serve as reference (T1-weighed fast-field echo scan, TR= 25 ms, TE = 4.5 ms, flip angle = 30°, imaging resolution = 0.94 × 0.94× 1 mm³, matrix = 256 × 256 ×160 slices).

For the region of interest (ROI) selection, functional MRI data were acquired. These consisted of 2179 ms single shot T2^* -sensitive echo planar imaging (EPI) sequences, 41 slices, 2 mm thickness, TR = 10 s, TE = 22 ms, flip angle 30^°, matrix 128×128, field of view 224 mm, SENSE reduction factor 2.7, 51 acquisition per subject. The influence of acoustic scanner noise on fMRI data was reduced by using a sparse sampling strategy [34]. A TE of 22 ms was chosen for optimal SNR at areas close to air-tissue boundaries, like the temporal lobes, at the cost of a small SNR decrease in the BOLD signal (see e.g., the introduction of [35]).

Auditory stimuli were delivered by a MR-compatible electrodynamic system (MR Confon GmbH [36]). This system was driven by a PC equipped with a digital-analogue card (National Instruments 6052E) controlled by Labview 6.1 (National Instruments Corporation, Austin, TX). The auditory stimuli were generated off-line using Matlab^© and consisted of temporally and spectrally modulated broadband “rippled” noise [37]. The stimuli had a frequency-range of 125 -8000 Hz with a spectral modulation density of 1 cycle per octave, a temporal modulation frequency of 2 cycles per second and a modulation-amplitude of 80%.

Each of the three functional runs consisted of the acquisition of 61 volumes under 3 experimental conditions: (i) a condition in which bilateral rippled noise noise was presented at a level of 70 dB (SPL), (ii) a condition in which subjects were visually instructed to protrude their jaw, and (iii) a combination of these two conditions (jaw protrusion + sound stimuli). Each condition was presented 15 times per functional run.

Before DTI analysis, preprocessing of raw data is necessary. This preprocessing step includes susceptibility correction, as well as eddy current and gradient table correction, depending on the scanner used for the acquisition. Susceptibility correction, which requires two DTI acquisitions with different fat-shift directions, was performed using MatLab, combined with the Statistical Parametric Mapping toolbox [40], the FieldMap toolbox [41], and the MatLab routines created by Andersson et al. [42]. Correction for eddy currents was performed with the software package FSL [39], a library of analysis tools for fMRI, MRI and DTI data. The gradient table modification was achieved using a MatLab toolbox created by Farrell et al. [43]. The last step before tracking is skull stripping (i.e., disregarding all non-brain data) and it was performed using BET [44].

We studied a number of brain areas involved in auditory processing and investigated how they are connected to each other. In particular, we determined the connectivity profiles between auditory cortex and inferior colliculus (AC-IC), auditory cortex and amygdala (AC-AM), inferior colliculus and amygdala (IC-AM) and all reverse connections. We disregarded the connections between inferior colliculus and cochlear nucleus (IC-CN), because a preliminary analysis showed that the motor fibers in the brain stem, running next to the acoustic fibers, have a strong influence on the results due to the poor spatial resolution of DTI. The selection of these ROIs (AC, IC, and AM) was done manually and for each individual patient, using the anatomical T1 scan as a primary reference. When possible (in 6 datasets), the location of the AC was checked by a comparison with BOLD activation maps (contrast: sound - baseline) provided by fMRI analysis. Location and size of the ROIs were checked by overlapping the drawn ROIs, normalized to the Montreal Neurological Institute (MNI-152) template brain [45], to the correspondent areas defined by the Juelich Histological Atlas, thresholded at 5% [46, 47]; ACs showed an average overlap of 70%, AMs showed an average overlap of 77%.

Following Anwander et al. [48], each ROI selection included both the gray matter and a part of the white matter directly beneath it. This was done to avoid dispersion of samples already at the early stages of the tracking procedure, due to the low anisotropy of gray matter areas. In this way the first part of the axon bundles leaving from the concerned gray matter area was included in the ROI.

To prevent the tractography from finding connections passing through cortical regions or the cerebrospinal fluid (CSF), masks were defined where tracts could not enter. These were positioned in the CSF above the AC and along Reil's Insula, and served as barriers to avoid tracts passing through these regions. Nevertheless, the usage of these masks was not sufficient and after the probabilistic tracking an additional filtering was necessary, i.e., fiber tracts leading to the cerebellum or to the motor cortex were manually removed.

The BEDPOST tool [49], which runs a Markov Chain Monte Carlo sampler, was used for building distributions of parameters describing the diffusion direction in each voxel. Probabilistic tractography was accomplished using the FMRIB Diffusion Toolkit [49]. For each voxel in a ROI, 5000 samples were taken from the distribution.

The probabilistic tractography outputs a connectivity map depicting the voxel-wise probability to reach any given voxel starting from a user-defined ROI, and it is defined as the percentage of samples leaving from the starting ROI that pass through that voxel. The connectivity map was then filtered so that only those paths connecting two different ROIs, a seed and a target ROI, were considered. An intrinsic problem of the sampling is that regions nearby the starting ROI are reached by a large number of samples, whereas the further a region is from the starting ROI, the fewer samples are able to reach it. This is a well-known and still unsolved problem. We addressed it by computing the connections between two ROIs in both directions: first starting from a region and reaching the other one, and then the other way around. Statistics on the paths, as described in the next section, were computed independently on each single path.

Although the voxel-wise FA value is the conventional statistical measure used in DTI analysis, we considered various novel statistics to compare the two groups. For each path between a seed ROI and a target ROI (and vice versa) the following three statistical values were computed. First, the mean FA value of a path was determined as the mean of the FA values of all voxels in the path. Second, the weighted mean FA value (wFA) of a path was determined, where the weighting was provided by the probability value of each voxel (i.e., the chance that a voxel is reached as determined by the probabilistic tracking). Voxels with a larger weight are more likely to represent the actual anatomical pathway. The third statistical value was the strength (S) of a path, defined as the percentage of the samples leaving from the starting ROI that were able to reach the target ROI. We consider the strength of a path as a key feature that describes the relevance of that path with respect to any other path connecting the starting ROI to other brain areas. Additionally, hemispheric differences were determined for all three statistical values (features) for each subject and path, using a lateralization index L defined as:

These statistical values were used to compare the two groups, resulting in a distribution of statistical values for each group and path. Since a chi-square test revealed that the distributions were not normally distributed, we used the non-parametric Kruskal-Wallis test in which the equality of the medians of the distributions is assessed to test for differences between control subjects and tinnitus patients.