Subjects (N = 100) were acute stroke patients enrolled in two music intervention studies in Helsinki and Turku, Finland. Fifty patients were recruited during 2004–2006 from the Department of Neurology, Helsinki University Central Hospital (HUCH) and 50 patients during 2013–2015 from the Department of Clinical Neurosciences, Turku University Hospital (Tyks). All patients had an MRI-verified acute ischemic stroke or intracerebral hemorrhage in the left (N = 49) or right (N = 51) hemisphere and subsequent cognitive and/or motor deficits, and they were all right-handed. Patients with hearing loss, prior neurological or psychiatric disease, or substance abuse were not included. All subjects gave written informed consent in accordance with the Declaration of Helsinki. The protocol was approved by the Ethics Committees of the HUCH and the Hospital District of Southwest Finland. All patients received standard medical treatment and rehabilitation for stroke. In both studies, all subjects underwent an MRI within 3 weeks of the stroke onset (acute stage) and at 6-month post-stroke stage. Behavioral assessment was performed at the acute and 3-month post-stroke stages. Out of the 100 recruited patients, 90 patients completed the 6-month MRI follow-up (Helsinki N = 47, Turku N = 43) and were included in the present study. The demographic and clinical characteristics of the patients are presented in Table 1.

Following the methodology of our primary study (Sihvonen et al., 2016), music perception was evaluated using a shortened version (Särkämö et al., 2009) of the Montreal Battery of Evaluation of Amusia (MBEA; Peretz et al., 2003), the most widely used, gold standard method for diagnosing amusia. MBEA was assessed at the acute stage and at the 3-month post-stroke stage as a part of a larger neuropsychological testing battery. The average score of the Scale and Rhythm subtests of MBEA was utilized as an index of overall music perception (referred to hereafter as MBEA total score). Following the cut-off values of the original MBEA (Peretz et al., 2003) applied in our previous studies (Särkämö et al., 2009; Sihvonen et al., 2016), patients with the MBEA total score <75% were classified as amusic. Based on the MBEA at the 3-month stage, amusic patients were further divided to those who were tested non-amusic (recovered amusics) and those remaining amusic (non-recovered amusics) based on the MBEA cut-off value.

This classification yielded 19 non-amusics (NAs), 6 recovered amusics (RAs), and 18 non-recovered amusics (NRAs) in the Turku cohort and 37 NAs, 16 RAs, and 37 NRAs in the combined Helsinki-Turku cohort. To evaluate pitch and rhythm amusia separately, similar principle was applied to the Scale and Rhythm subtest scores. Patients with Scale subtest score <73% in the acute stage were defined as pitch-amusic [N = 50, non-pitch-amusic (pNA) N = 40]. At the 3-month stage, 17 patients were classified as recovered pitch-amusics (pRA) and 33 as non-recovered pitch-amusics (pNRA). Rhythm subtest was evaluated with cut-off score <77%: 28 non-rhythm-amusics (rNA), 37 non-recovered rhythm-amusics (rNRA), and 25 recovered rhythm-amusics (rRA). Overall, in both the Turku and the combined cohort, the three groups were relatively well-matched with respect to demographic and clinical variables (Table 1). Education years and acute stage lesion volume showed a group difference, and these variables were therefore included as covariates in the analyses. The RAs and NRAs showed a significant difference in acute stage MBEA total score both in the Turku cohort (p < 0.001) and the combined (p = 0.003) cohort, suggesting that poor recovery of amusia was linked to its initial severity. The patients in both cohorts were originally recruited to a music-based intervention study (for the results of Helsinki study, please see Särkämö et al., 2008, 2010a, 2014; the results of Turku study have not been published yet). To verify that the intervention did not have an effect on the amusia analyses, we calculated a mixed-model ANOVA with Time (Acute/3-month) and Group (3 intervention arms) using the pooled sample. No significant Time × Group interaction effect was found in the MBEA total score (p = 0.248), suggesting that the music intervention did not have any effect on the recovery of amusia and, therefore, does not impact the results of the present study.

Patients from the Helsinki study were scanned with a 1.5T Siemens Vision scanner (Siemens Medical Solutions, Erlangen, Germany) of the HUCH Department of Radiology to obtain high-resolution T1 images (flip angle = 15°, TR = 1,900 ms, TE = 3.68 ms, voxel size = 1.0 × 1.0 × 1.0 mm). Patients from the Turku study were scanned using a 3T Siemens Verio scanner (Siemens Medical Solutions, Erlangen, Germany) of the Medical Imaging Centre of Southwest Finland and T1-weighted MPRAGE were obtained (flip angle = 9°, TR = 2,300 ms, TE = 2.98 ms, voxel size = 1.0 × 1.0 × 1.0 mm).

Preprocessing steps equivalent to the primary study (Sihvonen et al., 2016) were carried out. First, to achieve optimal normalization of MRI images containing stroke lesions, cost function masking (CFM) was applied (Brett et al., 2001). Using CFM prevents post-registration lesion shrinkage and out-of-brain distortion (Ripollés et al., 2012). To define the CFMs, A.J.S. and T.S. created binary masks of the stroke lesions by manually depicting the precise lesion boundaries on a slice by slice basis using T1 images of individual patients. Lesion tracing was carried out by using MRIcron software package (http://people.cas.sc.edu/rorden/mricron/index.html; Rorden and Brett, 2000). A sum image of all patients' lesions (from both cohorts) is shown in Figure 1.

T1 images and the created lesion masks were processed using the Statistical Parametric Mapping software (SPM8, Wellcome Department of Cognitive Neurology, UCL) under MATLAB 8.4.0 (The MathWorks Inc., Natick, MA, USA, version R2014b). T1 images were segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid probability maps using CFM and unified segmentation (Ashburner and Friston, 2005) with medium regularization. The created probability maps were then normalized into the MNI space (Montreal Neurological Institution). This technique corresponds to our primary study (Sihvonen et al., 2016) and has been widely used in stroke patients (Crinion et al., 2007; Andersen et al., 2010; Ripollés et al., 2012). The GM and WM images were modulated to preserve the original signal strength. Residual inter-individual variability was reduced by smoothing the GM and WM probability maps using an isotropic spatial filter (FWHM = 6 mm). Lastly, the binary lesion masks created in native space were also registered to MNI space.

Using the normalized acute stage lesion maps, VLSM was carried out with the Non-Parametric Mapping software (Chris Rorden's NPM, version 6 June 2013) in the combined Helsinki-Turku cohort (N = 90) adding 13 patients to the original VLSM study (Sihvonen et al., 2016) for more statistical power. Continuous VLSM analyses were carried out using the acute stage MBEA total score, Rhythm score, and Scale score. The following binary VLSM analyses were performed: NRA vs. NA, NRA vs. RA, and RA vs. NA. As the acute stage Rhythm and Scale subtest scores correlated strongly (r = 0.63), only the continuous analyses in rhythm or pitch amusia were carried out (note that the resulting maps show a great overlap and binary analyses would have yielded similar results). All voxels damaged at least in 10% of the patients were included in the statistical analysis (Dovern et al., 2011; Mirman et al., 2015; Timpert et al., 2015; Sihvonen et al., 2016). Multiple comparisons were accounted for with False Discovery Rate (FDR) correction with p < 0.05 threshold.

Voxel-based morphometric analysis was carried out using SPM8 software. Our original VBM study was based on the Helsinki cohort (N = 47), which is included in the present study. Therefore, in the current study with a higher statistical power, VBM was carried out using both the replication cohort (Turku; N = 43) and the combined Helsinki-Turku cohort (N = 90). The preprocessed GM and WM images were entered into a second-level analysis using a Group (NA/RA/NRA) × Time (Acute/6 months) mixed between-within subjects analysis of variance (ANOVA). Three different Group (NA > NRA, NA > RA, and RA > NRA) × Time (6 months > Acute) contrasts were calculated. Rhythm and pitch amusia were evaluated using preceding contrasts but with pNRA, pRA, and pNA, and rNRA, rRA, and rNA groups. Results were thresholded at a whole-brain uncorrected p < 0.005 at the voxel level (extent threshold: k > 100 voxels). Only clusters surviving an FWE-corrected p < 0.05 threshold are reported. Neuroanatomical areas were identified using the Automated Anatomical Labeling Atlas (Tzourio-Mazoyer et al., 2002) included in the xjView toolbox (http://www.alivelearn.net/xjview/). In addition to the previously determined covariates (education and lesion size), a covariate for the scanner was added in the gray and white matter VBM analyses, since the Turku and Helsinki patients were scanned with two different MRI scanners.

Longitudinal VBM analyses using the replication (Turku) cohort revealed similar results as our previously published study utilizing the Helsinki-cohort: Non-recovered amusia was associated with greater GMV decreases in the right STG and MTG, pitch amusia with GMV decreases in the right MTG, and rhythm amusia with anterior temporal GMV decreases. Additional GMV and WMV findings were observed and the results for the replication cohort are presented in the Supplementary Material.

Across the VLSM and VBM analyses, several lateral and medial temporal regions in the right hemisphere were found to be related to amusia and its recovery. In the acute stage VLSM, converging results from both continuous analyses of MBEA total score, Scale and Rhythm subtest scores and binary analyses comparing NRAs and NAs showed that severe non-recovered amusia was caused by lesions in the right STG, MTG, and insula. This is well in line with findings from neuroimaging studies of healthy subjects implicating superior temporal (Griffiths et al., 1998; Gutschalk et al., 2002; Patterson et al., 2002; Tramo et al., 2002; Hyde et al., 2008; Kumar et al., 2016) regions in pitch and/or melodic processing. Similarly, studies of both acquired amusia (Liegeois-Chauvel et al., 1998; Ayotte et al., 2000; Kohlmetz et al., 2003; Terao et al., 2006; Särkämö et al., 2010b; Hochman and Abrams, 2014; Sihvonen et al., 2016) and congenital amusia (Hyde et al., 2007; Albouy et al., 2013) have reported lesions/structural gray matter abnormalities specifically in right superior temporal and/or insular areas. Importantly, our novel VLSM and VBM results showed that the right STG/MTG has a crucial role also in amusia recovery as the NRAs had more lesions at the acute stage and also more GMV and WMV decrease from acute to 6-month stage compared to the RAs, suggesting that initial damage and further atrophy of these regions is a strong indicator for severe amusia that has a poor prognosis.

Separate VBM analyses for pitch and rhythm amusia showed that recovered pitch amusia was related to smaller GMV decrease in the right posterior STG/MTG compared to non-recovered pitch amusia. In contrast, recovered rhythm-amusics showed less GMV decrease in the anterior MTG, ITG, and fusiform gyrus as well as less WMV in the ITG and fusiform gyrus than the rNRAs. These results provide crucial information for pinpointing the functional organization of the right temporal lobe for pitch and rhythm processing, and moreover, they are in line with previous observations where anterior temporal lesions have been associated with rhythm processing deficits (Kester et al., 1991; Liegeois-Chauvel et al., 1998) and posterior temporal structures to spectral processing (Warren et al., 2005; Jamison et al., 2006; for a meta-analysis, see Samson et al., 2011).

Compared to our previous results, in addition to temporal areas, frontal, and parietal regions showed an association with amusia in the present study with higher statistical power. In the VLSM and VBM analyses, the NRAs had more lesions and more GMV decrease in the right IFG compared to both the NAs and the RAs. This pattern in NRAs extended also to the right MFG and PreCG when compared to the NAs. These results suggest that, in addition to the right STG/MTG, the right IFG-MFG-PreCG appears to be another crucial hub in acquired amusia as its lesions and atrophy are linked to initial severity and poor recovery. Previously, these areas have been implicated in the sequencing of auditory information and structural (syntactic) analysis of music (Koelsch, 2005; Tillmann et al., 2006; Bianco et al., 2016) in healthy subjects.

Reduced activation and connectivity and gray and white matter abnormalities in the right IFG have also been reported in congenital amusia (Hyde et al., 2006, 2007, 2011; Albouy et al., 2013). Interestingly, the left IFG emerged as the only lesion site in the VLSM contrast between the RAs and NAs. Along with its right hemisphere homolog, the left IFG (or Broca's area) has been found to process syntactic information in both language and music (Maess et al., 2001; Kunert et al., 2015) and, conversely, that its damage impairs both linguistic and musical syntactic processing (Patel et al., 2008; Sammler et al., 2011). Our results provide support for the role of the left IFG in amusia, but suggest that its acute damage results in less severe and transient form of amusia (Kumar et al., 2016). VBM results also showed that compared to the NAs, both the NRAs and the RAs had more GMV/WMV decrease also in right parietal areas, especially in right IPL/SPL, suggesting that this region is generally associated with amusia, regardless of its recovery. Right parietal areas have been implicated in neuroimaging studies to be involved in the processing of more higher-level melodic features, such as tonality (Foster et al., 2013; Royal et al., 2016), as well as in tonal working memory and recognition (Jerde et al., 2011; Schulze et al., 2011; Albouy et al., 2017). Importantly, in the pitch domain, recovered pitch amusics showed more GMV in the right IPL than the pNRAs.

Finally, the pattern of acute lesions and the longitudinal GMV/WMV decreases observed for NRAs vs. NAs also encompassed right subcortical areas, both in the striatum (caudate, putamen) and in limbic areas (amygdala, hippocampus, parahippocampal gyrus). Together with the GMV decrease in PreCG, the striatal changes are most likely related to rhythmic deficits, as these regions have been strongly linked to rhythm processing in neuroimaging studies (Penhune et al., 1998; Grahn and Brett, 2009; Grahn and Rowe, 2009; Alluri et al., 2012). In line with the key role of rhythm in mediating the emotional valence and arousal induced by music, musical pulse, or rhythm has also been shown to engage amygdala and hippocampus (Alluri et al., 2012; Toiviainen et al., 2014), the latter playing a role also in auditory working memory (Burunat et al., 2014). Musical training has also been linked to increased hippocampal and amygdala volume (Oechslin et al., 2013; Dohn et al., 2015; Vaquero et al., 2016), providing converging support for our finding of decreased volume of these structures in non-recovered amusia.

Converging evidence from neuroimaging studies in healthy subjects utilizing natural music stimuli suggest that music perception and analysis is a highly wide-spread process in the brain, engaging a large-scale network of bilateral temporal, frontal, parietal, and subcortical regions (Schmithorst, 2005; Brattico et al., 2011; Alluri et al., 2012; Burunat et al., 2014; Toiviainen et al., 2014). In contrast, structural deficits, indicated by volumetric and cortical thickness measures that have been reported in congenital amusia have thus far been limited to right superior temporal and inferior frontal areas (Hyde et al., 2006, 2007; Albouy et al., 2013). Our results support the strong right hemispheric basis for amusia, but suggest that in acquired amusia the lesion and atrophy pattern underlying the severity and persistence of the deficit might be more extensive and wide-spread. While the statistical power in the VLSM analyses in the current study was higher in the right hemisphere than in the left hemisphere, the observed longitudinal results do not appear to be due to a simple mass effect since the overall volume of the lesions was controlled for in the VBM analyses. However, future studies investigating acquired amusia in patients with left hemisphere damage would be of great interest. As stroke may have impact on other modalities, the effect of other cognitive deficits on MBEA performance should be also taken into consideration.

On the other hand, given that the large-scale atrophy pattern extended from right prefrontal (IFG/MFG) all the way to right posterior (parietal/occipital) regions, which are not considered to be part of the music perception network, it is possible that this may reflect damage to the long-range white matter pathways, such as the inferior fronto-occipital fasciculus (IFOF). The IFOF is a ventral pathway that runs from its posterior terminations (inferior and middle occipital gyri, parietal lobe) through the external capsule, between the insula and putamen, and connects to multiple temporal (STG/MTG) and frontal (IFG/MFG) areas (Catani et al., 2002; Hau et al., 2016), thereby covering most of the lesion/atrophy areas observed in our NRA patients. Thus, far using diffusion weighted MRI (DW-MRI), amusia has been linked only to the right arcuate fasciculus (Loui et al., 2009), a dorsal pathway connecting the IFG and STG. However, this finding has been recently challenged using a larger sample of congenital amusics and multiple algorithms (Chen et al., 2015) and the crucial white matter connections affected in congenital amusia remain unclear (for a review, see Peretz, 2016).

In future, more research on the roles of the ventral and dorsal pathways in amusia is clearly needed. Furthermore, although the data presented here sheds more light on the recovery mechanisms in acquired amusia, an interesting idea would be to use machine learning methods to predict, by means of lesion data, not only which patients will become amusics, but also which one will recover (Rondina et al., 2016). Further research is needed, especially studies using multimodal MRI data.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.