The axonal connections between the thalamus and the auditory cortex have been investigated in animals using different invasive techniques. These fibers stem from the medial geniculate nucleus of the thalamus (MGN), as first described by von Monakow (1882), and contact a specific area on the posterior part of the Sylvian fissure (Minkowski, 1923; Polyak, 1932), which has been described as a “rudimentary transverse temporal gyrus” (Walker, 1937) in monkeys. In most species, three major divisions of the MGN are identified: ventral (or principal), dorsal (or posterior), and medial (or magnocellular) (Winer et al., 2001; Jones, 2003). Each of these divisions has unique connections to different cortical regions. The ventral division receives input from the central nucleus of the inferior culliculus (Ic) and almost exclusively projects to what is defined as the core region of the auditory cortex (Mesulam and Pandya, 1973; Burton and Jones, 1976; Morel and Kaas, 1992; Morel et al., 1993; Hashikawa et al., 1995; Rauschecker et al., 1997). This region is distinguished by dense immunoreactivity for the calcium-binding protein parvalbumin, as most of its inputs come from the ventral MGN parvalbumin immunoreactive cells (Molinari et al., 1995), even if some connections with the other MGN divisions appear to exist (Luethke et al., 1989; Morel et al., 1993). This region occupies a portion of the caudal superior temporal plane (postero-medial part of the Heschl’s gyrus in humans) and it is characterized by a dense population of small granule cells (e.g., koniocortex) with a well-developed layer IV (Merzenich and Brugge, 1973; Seldon, 1981). Both the ventral MGN and the core region show a tonotopical organization in which the representation of frequencies is spatially organized. This suggests a topographical organization of fibers connecting similar frequency domains in these two structures (Burton and Jones, 1976; Molinari et al., 1995). The core region is surrounded by a secondary narrow belt region and a third, more lateral region that occupies the lateral surface of the superior temporal gyrus (Hackett et al., 1998; Kaas and Hackett, 2000). This latter “para-belt” region is generally considered to be a higher-order auditory region or auditory association cortex that integrates auditory with non-auditory multisensory information (Hackett et al., 1998). These regions are less responsive to pure tone sounds, preferring more complex sounds, and do not show the clear tonotopical organization typical of the core region (Rauschecker et al., 1995; Jones, 2003). The dorsal and medial divisions of the MGN constitute the major inputs to these secondary auditory association regions. These nuclei receive inputs from the external nucleus of the Ic, as well as from lower brainstem relays, and bypass the core region to project to secondary auditory and other cortical regions (Rauschecker et al., 1997; Jones, 2003; Winer and Lee, 2007).

Projections from the ventral MGN to the core region correspond to the most direct classic auditory pathway, also called the lemniscal pathway. These parallel connections are tonotopically organized and their neurons show sharp responses to tones (Morel et al., 1993). Direct projections from the other MGN divisions to secondary auditory cortical regions are part of the extralemniscal (or non-lemniscal) auditory pathway. These fibers are separate from but lie adjacent to those of the lemniscal pathway in the ascending auditory system. In addition, their subdivision continues at the cortical level, where the non-lemniscal pathway has stronger and more diffuse connections with regions surrounding the core region (Lee and Sherman, 2010). These pathways are less tonotopically organized and their neurons demonstrate fewer sharp responses to sounds.

Classical studies in non-human primates focus on the topography of the connectivity between the MGN and the auditory projection cortical territory but provide no information on either the course and extension of the AR tract itself or its relationship with the other WM pathways of the brain. Using Marchi axonal degeneration, Polyak (1932) provided a detailed description of the course of this tract in rhesus macaques (Figure 2). He describes a dense bundle of closely assembled fibers that leaves the MGN, turns laterally and crosses the most ventral portion of the internal capsule (IC) immediately above the lateral geniculate nucleus (LGN) of the thalamus (Figure 2). At this level, these fibers are distinguishable from somato-sensory fibers because of their nearly horizontal orientation. The AR then bends ventrally and reaches the external capsule (EC) by passing through the ventral edge of the posterior putamen. Once there it meets other projection and association bundles before finally reaching the WM of the superior temporal convolution close to the Sylvian fissure (Figure 2). He describes the AR as a regularly arranged projection system, where fibers lie parallel to one another until gradually diverging only when approaching the cortex.

Classical topographical descriptions in humans (Figure 3) provide a very similar description (Dejerine and Dejerine-Klumpke, 1895; Flechsig, 1920; Pfeifer, 1920) of the AR, which may be summarized as follows. The AR leaves the MGN and travels in an antero-lateral direction; it then passes through the posterior portion of the IC, proceeds along the corona radiata and curves around the inferior portion of the circular sulcus of the insula before entering the transverse temporal gyrus of Heschl (HG) in a ventral-to-dorsal direction (Pfeifer, 1920). At this macrostructural level, both animal and human studies delineate a bundle with a transverse orientation that lies adjacent to and crosses over other main WM bundles before reaching the auditory cortex. In both animals and humans, the AR intermingles with the fibers of the IC in its most posterior portion. Furthermore, Dejerine and Dejerine-Klumpke (1895) described the close proximity of the AR to the optic radiation (OR) at the stemming point in the thalamus, defining this region as the “carrefour sensitive” (sensory intersection). Polyak (1932) also studied this region, stating that the OR and AR, although lying close together, are completely separate: the AR is located more anterior and crosses at a right angle above the OR, which follows a posterior direction in the sagittal plane. As will be discussed in the following section, this configuration and certain other anatomical features of the AR pose serious challenges to its 3D tractography reconstruction.

The myeloarchitectonic maps from Dejerine and Dejerine-Klumpke (1895) and Flechsig (1920), while being of invaluable historical significance, cannot be used to extract precise anatomical information that can be applied to modern brain atlases or neuroimaging studies. More recently, radiological information about the AR anatomical organization was obtained in human post mortem myelin-stained sections (Rademacher et al., 2002; Bürgel et al., 2006). These studies confirm the classical topographical description of the acoustic fibers and are of great importance as they represent the main reference framework for in vivo imaging studies of this brain region and provide the opportunity to investigate inter-subject variability and hemispheric asymmetry. Previous studies have found that the AR does not enter the lenticular nucleus, but rather runs dorsally to the OR, crossing the temporal isthmus as it ascends to the auditory cortex (Pfeifer, 1920; Polyak, 1932; Bürgel et al., 2006). Fanning of these fibers in HG creates a hat-like structure that covers the posterior end of the lenticular nucleus (Rademacher et al., 2002).

According to Flechsig (1920), these fibers are divisible into two bundles, one of which ascends near the Ec and enters the auditory cortex from the superior-posterior side. The other bundle courses for some distance in the company of the OR before passing behind and below the fossa sylvii where it pierces the bases of the middle and inferior temporal gyri to reach the transverse temporal gyrus or gyri. Similarly, early evidence from animal studies suggests that the AR is subdivided into dorsal and ventral components (von Monakow, 1882), although, Rademacher et al. (2002) found more recently only a single and heavily myelinated bundle.

Overall, the macrostructural description of the AR in humans resembles the description of this tract in non-human primates. The origin and termination of this bundle, together with the extension and relationship to other WM bundles, is maintained. However, the macro- and micro-anatomical correspondence between the different cortical regions in monkeys and humans is not straightforward (Baumann et al., 2013). Compared to non-human primates, the human cortical surface of the auditory regions demonstrates additional gyri and higher inter-subject and interhemispheric variability (Galaburda et al., 1978; Hackett et al., 2001), both of which may affect the AR anatomy. The human auditory cortex also shows higher differentiation into sub-regions as compared to non-human primates, and the core region is larger than the belt region (the opposite is true for monkeys) (Fullerton and Pandya, 2007). Both the differences in cortical anatomy between non-human primates and humans and the major variability of cortical and subcortical structures across subjects and hemispheres in humans (Bürgel et al., 2006) raise interesting questions about the possibility of a relationship between such morphological differences and human-only language abilities.

Diffusion MRI (dMRI) tractography allows for the investigation of WM architecture in the human brain non-invasively in vivo. Since its first applications (Mori et al., 1999), most of the well-known WM bundles of the human brain have been reconstructed using diffusion-based tractography methods (Catani and Thiebaut de Schotten, 2008; Lawes et al., 2008). Despite its potentials, dMRI tractography has several important limitations that have been discussed in depth in the literature (e.g., Jones and Cercignani, 2010; Thomas et al., 2014; Maier-Hein et al., 2017). In principle, these limitations affect most tractography reconstructions, but here we focus on how they particularly affect the 3D reconstruction of the AR. Reconstructing the AR three-dimensionally is highly challenging at present due to the anatomical features described in the previous sections: its relatively small size, transversal orientation, and location in a region with a high density of crossing fibers. This largely prevents the inclusion of this particular tract in most tractography investigations.

As shown in the previous section (see Figure 2, 3), in its medio-lateral course from the MGN to the HG, the AR lies in a nearly horizontal position and, for this reason, crosses some of the major fiber systems of the human brain: internal capsule, external capsule, and posterior thalamic radiation (Maffei et al., 2018). Resolving the fiber crossing is a well-known challenge in dMRI (Tuch et al., 2002; Dell’Acqua and Catani, 2012; Jeurissen et al., 2013). The classic tensor model (Basser et al., 1994) is capable of characterizing only one main fiber orientation per voxel and it has been shown to constantly fail in regions where voxels contain complex fiber architectures (Behrens et al., 2007; Jbabdi and Johansen-Berg, 2011). The impact of this limitation is particularly evident for non-dominant tracts, given that the orientation produced by the tensor will be closest to the largest contributing direction in most cases. This effect is amplified at the low resolution of commonly available diffusion protocols due to within-voxel partial volume averaging effects (Tournier et al., 2011). When implemented in tractography studies, the diffusion tensor model has proven to be incapable of detecting the 3D profile of the AR. Streamlines are either truncated when entering voxels containing major inferior-superior orientations or erroneously embedded in the reconstruction of these major projection bundles, with no visible streamlines contacting the HG (Behrens et al., 2007; Crippa et al., 2010; Berman et al., 2013). This has likely been the primary factor preventing the investigation of the auditory system by means of diffusion-based tractography. Some studies have used the diffusion tensor to investigate the WM microstructure of the auditory system, limiting the structural investigation of the auditory pathways to the extraction of mean quantitative diffusion measures [e.g., fractional anisotropy (FA)] from specific regions of interest (ROI) (Chang et al., 2004; Lee et al., 2007; Lin et al., 2008; Wu et al., 2009). However ROI-based analysis can lead to inaccurate results, especially for WM tracts that are extremely variable across subjects, such as the AR (Rademacher et al., 2002). Therefore, it is typically preferable to map the exact anatomy of such WM tracts in individual subjects/patients.

To address the intrinsic limitations of the tensor formalization, more advanced models have been introduced that can better account for fibers crossing, by modeling more than one fiber population per voxel (Tuch, 2004; Tournier et al., 2008; Descoteaux et al., 2009). These models open up the possibility of propagating streamlines through crossing fiber regions, thus allowing the reconstruction of non-dominant WM bundles, such as the AR. However, together with the low-level diffusion model employed, other parameters play a role in the accurate reconstruction of WM bundles, such as the tractography parameters chosen and the strategy to define inclusion ROIs. Here we report studies that reconstruct the AR 3D tractography profile in vivo using multi-fiber-based models (Figure 4), taking into consideration the different combinations of acquisition parameters, diffusion models and ROIs selection strategies that they used (Table 1).

Behrens et al. (2007) were able to visualize the course of the AR from the MGN to the cortex using the ball-and-stick model and probabilistic tractography (Behrens et al., 2003) (Figure 4A). The authors demonstrate how this multi-fiber model can reconstruct the AR, overcoming the limitations of the tensor model. Three following studies were then able to demonstrate the reliability of this model for successful in vivo reconstruction of the profile of the auditory tract in both healthy subjects and those with tinnitus (Crippa et al., 2010; Javad et al., 2014; Profant et al., 2014). In these studies, different combinations of inclusion ROIs were used to isolate the AR, including the Ic, the MGN, and both functionally and manually defined HG. Nevertheless, the profile of the 3D tractography reconstruction looks visually similar across the four studies and shows connections between the posterior thalamus and the auditory region WM. However, artifacts are visible along the inferior-superior axis in the middle part of the AR at the level of the crossing with the IC and these false positive reconstructions are likely to be related to the probabilistic nature of the diffusion model and tractography algorithm used. Despite the inclusion of false positive signals in the reconstructions, the ball-and-stick model has the important advantage of being accessible to low b-value (b = 1000 s/mm²) diffusion protocols which makes it suitable for clinical investigations of the AR. Berman et al. (2013) (Figure 4B) used a solid-angle q-ball model (Aganj et al., 2010) and probabilistic tractography to successfully reconstruct the auditory connections between the MGN and HG. This reconstruction looks anatomically very accurate and free of false positive artifacts. However, q-ball methods require higher b-value diffusion data (b ≥ 3000 s/mm²). In these models, the angular resolution of the reconstructed diffusion profiles is increased and the crossing fiber configurations are correctly represented (Tournier et al., 2013), although they pose limitations to its use in clinical populations. Our group (Maffei et al., 2018) used ultra-high b-value Human Connectome Project diffusion data-sets (Fan et al., 2015) and spherical deconvolution (Tournier et al., 2008) to reconstruct AR streamlines using probabilistic tractography. We compared the results to Klinger’s post-mortem blunt micro-dissections (Figure 4C), a method based on a brain freezing technique optimized to reveal WM (Ludwig and Klingler, 1956). This approach has been used in several studies to evaluate tractography accuracy (Fernaìndez-Miranda et al., 2015; De Benedictis et al., 2016; Pascalau et al., 2018). The obtained reconstructions agreed with the AR anatomy revealed in the post-mortem dissections and no additional exclusion regions were needed to isolate the AR profile. However, the ultra-high b-values used in this study (b ≤ 10,000 s/mm²) are very rarely achievable, even in research settings. More recently, Yeh et al. (2018) published a population-averaged atlas of several WM connections, including the AR, using q-space diffeomorphic reconstruction (QSDR) (Yeh and Tseng, 2011) and deterministic tractography for multiple fiber orientations. In their approach, the high angular and spatial resolution of the data (1.25 mm isotropic, and b ≤ 3000 s/mm²) and the large sample (842 subjects) allowed them to reduce the rate of false positive artifacts. However, the AR profile shown in this study, while correctly originating at the posterior thalamus, does not reach the expected auditory cortex on the superior side of the temporal lobe.

Overall, the reconstructed 3D profiles shown in these studies are in accordance with the macrostructural landmarks defined by classic anatomical studies: streamlines originate in the posterior thalamus and course in an antero-lateral direction to terminate in the temporal lobe (Dejerine and Dejerine-Klumpke, 1895). However, AR reconstructions are still highly variable across studies. In particular, while showing similar profiles at the thalamic stemming region, reconstructions differ as they approach the cortex, either falling short of reaching the HG (Crippa et al., 2010; Yeh et al., 2018) or creating false positive artifacts at the intersection with vertically oriented fibers (Javad et al., 2014; Maffei et al., 2018). Moreover, disagreement about the relationship with neighboring tracts exists. Berman et al. (2013) suggest that AR streamlines cross the inferior longitudinal fasciculus (ILF), while Behrens et al. (2007) and Javad et al. (2014) claim that they cross the OR. In a recent work from our group, we did not find that the AR is in close proximity to the ILF (Maffei et al., 2018), supporting older studies that report no crossing between the AR and OR (Pfeifer, 1920; Polyak, 1932), as described above. In addition to variability across studies, low reproducibility across subjects is also reported. Some groups have been able to reconstruct AR tracts on both hemispheres on 100% of subjects (Behrens et al., 2007; Profant et al., 2014; Maffei et al., 2017). In contrast, even when using similar diffusion models (e.g., ball-and-stick), other studies report reconstructions successful in both hemispheres in much lower proportions, such as 35–50% (Crippa et al., 2010) or 71–86% (Javad et al., 2014).

We suggest that the present variability and low reproducibility in the reconstructed AR profile are related to a combination of some of the specific characteristics of the AR that make its tractographic reconstruction quite challenging, even for state of the art tractography techniques: its anatomical location, small size, and inter-individual anatomical variability.

As alluded to in the previous section, the AR constitutes a compact but relatively short and small bundle that lies horizontally in a region with a high density of vertical fibers. Even if multi-fiber models proved capable of representing this crossing, the degree to which this crossing can be accurately resolved in the final 3D reconstruction also depends on the tractography algorithm used and the intrinsic angular resolution of the dMRI data (Tournier et al., 2013). For example, using a higher b-value (Berman et al., 2013; Maffei et al., 2018) might help improve accuracy of results relative to those obtained with lower b-value data (Crippa et al., 2010). However, in these studies several exclusion ROI have been employed to either constrain tractography (Behrens et al., 2007) or clean the results (Crippa et al., 2010). This renders the accuracy of the final reconstructions sensitive to the selection of these ROI, complicating comparisons across studies.

The use of different ROI selection strategies can strongly affect the resulting tractography reconstructions. The HG is a complex structure that shows large variability in sulcal landmarks across subjects and hemispheres (Rademacher et al., 2001), whereas the MGN is a very small structure, varying from 74 to 183 mm³ (Kitajima et al., 2015), making it difficult to locate in neuroimaging data and also highly variable across individuals and hemispheres (Rademacher et al., 2002). Therefore, this variability poses difficulties in the selection of the ROIs used to initiate the tractography reconstruction or to perform the virtual dissections. For example, while reliable automatic segmentation tools for the entire thalamus are available in different public software packages (e.g., FSL¹, Freesurfer²), it is far more challenging to automatically segment smaller structures, such as the MGN, due to both their size and their lower MRI contrast with neighboring WM. Alternative solutions exist but are also challenging. For example, subject-specific manual segmentations of MGN can lead to high anatomical accuracy, but they are very time intensive and, therefore, costly to do in large subject cohorts. Brain atlases also can be used to define the MGN, however given the small size and inter-subject variability, atlas-driven segmentations are unlikely to provide a good anatomical match for all subjects. The development of more accurate automatic parcellation techniques for the thalamic nuclei is expected to improve the accuracy of seed-to-target definition and, thus, of the resulting tractography reconstructions. Recently, a new and promising probabilistic atlas of the thalamic nuclei has been proposed based on a combination of ex vivo MRI and histology (Iglesias et al., 2018).

Future research in the field should focus on how to improve the sensitivity and reproducibility of the tractography reconstruction of this bundle. This could be achieved by inputting prior anatomical knowledge in the tractography process, as it is implemented in global-tractography-based frameworks (Yendiki et al., 2011) or more recently developed bundle-specific algorithms (Rheault et al., 2019). Parallel to these advances at the tractography level, there have been efforts in validating tractography results at a micro-anatomical scale. As this validation process progresses, we expect to expand our knowledge of the exact boundaries of the human AR and, consequently, better inform its tractography reconstruction and improve accuracy of tractography results. A more accurate tractography investigation of the AR could expand our structural and functional knowledge of the auditory system, as proposed in the next section.

The ability to communicate through speech is quintessentially human. However, the anatomical organization and the functional mechanisms underlying speech comprehension in the brain are still not understood completely. The acoustic information that reaches the primary auditory cortex via the AR fibers is processed within neural networks that depend on cortico-cortical short- and long-range connections involving temporal, parietal and frontal regions, as schematized in the dual-stream model (Hickok and Poeppel, 2007; Saur et al., 2008; Friederici, 2009). Within this processing network, it is unclear where language-specific processing starts and whether the auditory cortex is involved in speech-specific analysis. Some theories suggest that the left auditory cortex is specialized in processing temporal cues that are fundamental for speech comprehension (Zatorre et al., 2002; Poeppel, 2003) and that this language-specific encoding might actually start at the subcortical level (Hornickel et al., 2009). The diffusion-based reconstruction of the AR, and of the auditory pathways at large, could help address the structural-functional relationship of speech perception. At the structural level, it would be interesting to understand whether the AR exhibits a degree of leftward lateralization in its volume, as demonstrated for some of the other WM bundles implicated in language processing (Catani et al., 2007). Reports on the macroscopic volumetric asymmetry of cortical auditory regions have been known for some time (Von Economo and Horn, 1930; Galaburda et al., 1978; Geschwind and Galaburda, 1985; Penhune et al., 1996), but only one study specifically investigated the hemispheric lateralization of the AR (Bürgel et al., 2006). At the functional level, the recent association of tractography and neurophysiological techniques [such as magnetoencephalography (MEG) and electroencephalography (EEG)] opens interesting possibilities for investigating these topics. EEG metrics have been recently correlated with diffusion metrics in the investigation of the OR (Renauld et al., 2016). Similarly, EEG/MEG- and tractography-derived measures could be combined to investigate the relationship between temporal cortical regions and auditory function in both healthy subjects and patients.

On a finer scale, AR streamline terminations could be combined with functional MRI to provide critical insights into the subdivision of the auditory cortex, the borders of which are still not clearly defined using in vivo neuroimaging methods (Baumann et al., 2013). In this sense tractography could be used to investigate the topographical organization of the AR with respect to the different subdivisions of the auditory cortex. The different auditory cortical regions show a hierarchical organization in information processing (Tardif and Clarke, 2001), from highly specialized core regions to more integrated tertiary para-belt regions. This is confirmed by functional MRI studies, which suggest a gradient of increasingly more complex and abstracted processing from primary to higher-order auditory regions (Rauschecker et al., 1995; Humphries et al., 2014). Direct connections from MGN to secondary regions have been shown in monkeys (Rauschecker et al., 1997), suggesting that this functional organization might be maintained in the topographical organization of the AR fibers. At its present stage, tractography can locate and delineate the profile of major WM bundles with some accuracy, but it is very difficult to achieve precise site-to-site connectivity analysis with it; this limits the in vivo investigation of the WM topographical organization of the human brain. However, some studies use diffusion tractography to parcellate functionally different cortical regions (Rushworth et al., 2006; Anwander et al., 2007) and investigate the topographical organization of major bundles (Lee et al., 2016), and new methods have been proposed to advance the use of tractography for this purpose (Aydogan and Shi, 2016). In this scenario, it would be interesting to investigate whether some language-specific connections exist inside the AR and whether these project to higher-order language-specific cortical regions. Moreover, this would also allow for the investigation of whether the tonotopical organization of the primary auditory cortex is reflected in its thalamo-cortical connections, as was recently shown in the mouse brain (Hackett et al., 2011). As dMRI acquisition (in particular, spatial resolution), diffusion modeling and tractography techniques improve, we will be able to bridge the gap between the micro-anatomical knowledge we have of the thalamo-cortical connections in animals and the macro-anatomical description in humans.

Today, X-ray therapy (XRT) is the standard of care for most brain tumors. However, XRT can damage normal brain tissue, causing neurocognitive deficits in different cognitive domains (Makale et al., 2016). The consideration of neuroimaging techniques for treatment planning is gaining importance as it helps avoid such complications by minimizing the unnecessary absorption of radiation in sensitive regions outside the tumor. Nevertheless, further improvement of radiation treatment will require tailored radiotherapy based on intra-treatment response (Wong et al., 2017).

Additionally, studies have shown XRT side effects in both gray (Bahrami et al., 2017) and white matter, although it is still largely unclear how variable sensitivity to radiation injury is across various regions of the brain (Connor et al., 2017). Kawasaki et al. (2017) show that tractography can help evaluate how radiation from XRT differentially affects WM regions and pathways. This knowledge, together with an understanding of how damage to such regions and pathways affects cognitive processes, could be used in the future to further optimize radiation treatment planning.