This prospective study was approved by the ethics committee of Konkuk University Medical Center. Sixteen patients (mean age, 43.2 ± 10.2 years; 11 males) with traumatic anosmia were included in the study between November 2012 and May 2014. The inclusion criteria were as follows: 1) presentation within 2 years of head injury; 2) fulfilment of criteria for anosmia based on the Korean Version of the Sniffin' stick (KVSS) II test; and 3) age between 18 and 65 years. The exclusion criteria were as follows: 1) history of olfactory impairment; 2) history of sinonasal disease; or 3) current nasal symptoms. All patients completed psychophysical olfactory testing, including the KVSS I and II, and underwent nasal endoscopy to exclude the possibility of obstructive olfactory loss.

Twelve healthy controls (mean age, 26.8 ± 8.4 years; 8 males) were also recruited from the local community. The inclusion criteria for healthy controls were as follows: 1) normal olfactory function; 2) no brain lesions or prior head trauma; and 3) no history of psychiatric or neurologic diseases. All study participants were right-handed. All participants provided written informed consent prior to study participation.

An otorhinolaryngologist with 17 years of clinical experience performed the endoscopic examination of the nasal cavity and the clinical olfactory performance testing in all patients to ensure that patients met the criteria for anosmia. Clinical olfactory performance measures included the KVSS I and II. The KVSS I test is a modified Sniffin' stick test optimized for Korean patients, which includes odors familiar to Korean individuals, and the KVSS II is the Korean equivalent of the Sniffin' stick test (1617). Clinical diagnosis of anosmia was based on the KVSS II scores; total scores between 0–20 were classified as anosmia, scores of 20.25–27 were classified as hyposmia, and scores of 27.25–48 were classified as normosmia (1617).

All participants underwent MRI using a 3T MR scanner (Signa HDxT; GE Healthcare, Milwaukee, WI, USA) with an 8-channel head coil. High spatial resolution T1-weighted three-dimensional (3D) anatomic images were obtained in the axial plane using a fast-spoiled gradient echo sequence (repetition time [TR]: 7.8 ms; echo time [TE]: 3.0 ms; matrix: 256 × 256; flip angle: 13°; number of sections: 134; field of view: 240 × 240 mm²; section thickness: 1.3 mm).

rs-fMRI was performed in the axial plane using a gradient-echo echo-planar imaging (EPI) sequence (TR: 2000 ms; TE: 30 ms; matrix: 64 × 64; flip angle: 90; number of sections: 40; field of view: 240 × 240 mm; section thickness: 3.5 mm; scan of volume: 800) with a voxel size of 3.75 × 3.75 × 3.5 mm.

fMRI data pre-processing was performed using SPM12 (Wellcome Trust Centre for Neuroimaging, London, UK) and in-house codes of MATLAB R2016b (MathWorks, Natick, MA, USA). The first three volumes of each dataset were discarded to allow for equilibration effects. We performed slice timing correction to realign fMRI slices of each scan to the first slice that was the reference slice based on its relative timing within the entire scan. Then, spatial realignment was performed to remove movement artifacts from the fMRI data. Scans were spatially realigned to a reference scan in order to revise any spatial mismatches caused by movement (18). We chose the first scan as the reference scan and used the default option of SPM12 (quality: 0.9; separation: 4; smoothing: 5; interpolation: 2nd degree). fMRI scans from each subject were then co-registered to their T1-weighted 3D anatomic images (19). The default option of SPM12 (objective function: normalized mutual information; histogram smoothing: 7) was used here as well. Next, steps were taken to reduce the noise signals. All scans were temporally band-pass filtered (0.01–0.08 Hz) to remove slow signal drifts (2021). Brain tissues were segmented using the co-registered T1-weighted MRI scans, and white matter and cerebrospinal fluid signals were extracted. We linearly regressed out the white matter and cerebrospinal fluid signals, while the linear temporal trends and motion parameters acquired from the realignment step (six parameters) were also considered as nuisance parameters. Finally, we performed spatial normalization to the EPI template with a 2-mm slice thickness of the MNI space (Montreal Neurological Institute: International Consortium for Brain Mapping) and spatial smoothing with the 4-mm full-width at half-maximum Gaussian kernel.

In this study, we used a total of 279 functional areas for brain network analysis. We further added the olfactory network to the pre-defined whole brain network.

First, the pre-defined whole brain network suggested by Jonathan Power and colleagues was constructed with 264 functional areas (22). Those areas were then divided into 12 predefined well-known networks, involved in different cognitive processes. Among the 264 functional areas, 33 areas with uncertain network classifications were also included (22). Next, the olfactory network was defined by statistical localization (23). The olfactory network was comprised of 21 functional areas including the piriform cortex, insular cortex, prefrontal cortex, and anterior cingulate/frontal cortex (Supplementary Table 1). All regions-of-interest (ROIs) were near-sphere in shape (3 × 3 × 3 cube with an additional voxel on the center of each face, total 33 voxels) and located at their center. Six ROIs from the pre-defined whole brain network were discarded as they overlapped with the ROIs from the olfactory network. A final check was performed to ensure there were no overlapping voxels between the ROIs.

Functional connectivity was measured as the Pearson correlation coefficient between the ROI time-series. The whole brain network with 279 × 279 connectivity was constructed. We used binary network type for calculating the network properties, where the connectivity of p value < 0.05 (positive and significant correlation) takes the value of 1, otherwise the connectivity takes the value of 0 (24).

Two network properties were measured to explore functional integration and segregation (2526). Global efficiency is a measure of how efficiently a network exchanges information. Global efficiency has been suggested as a measure of functional integration and is calculated as the inverse of the harmonic mean of the shortest path length in the network:

Where d_ij is the shortest path length connecting nodes i and j, and G represents the graph.

Modularity is a statistical unit that quantifies the degree of network subdivision into clearly delineated modules. The Louvain community detection algorithm was applied to construct modules in the network (27). The Louvain algorithm optimizes and constructs the modules to maximize the modularity of the network.

Where n is the number of all edges in the graph, K_ij is the binary value of the edge between nodes i, and j. k_i and k_j are the number of the edges attached to nodes i and j. c_i and c_j represent the modules of the nodes and δ is a simple delta function.

A neuroradiologist with 20 years of experience who was blinded to the clinical data assessed the images from the T1-weighted fast-spoiled gradient echo sequence for structural brain abnormalities. For assessment, multi-planar reformatted images as well as original axial images were used. When tissue loss was present, the maximal dimension of the tissue loss was measured on a picture archiving and communication system (Centricity Enterprise Imaging, GE Healthcare, Barrington, IL, USA).

To compare the demographic variables between patients with anosmia and healthy controls, the nonparametric Mann–Whitney test and Fisher's exact test were used due to our small sample size using the Statistical Package for the Social Sciences (SPSS, version 23.0, IBM Corp., Armonk, NY, USA). P values less than 0.05 were considered to be statistically significant.

Statistical analysis for the rs-fMRI data was performed in MATLAB using in-house codes. All group comparisons were tested using a two sample t test after controlling for age difference between the groups. The age effect unaffected by the disease was estimated by linear regression for the healthy controls and was controlled for both groups (28). Individual functional connectivity was tested using Student's t distribution and was transformed into p values to determine whether the functional connectivity was significant in the brain networks. Functional connectivity comparison within networks was further tested in consideration for multiple comparisons. False discovery rate (FDR) was controlled in the olfactory network comparison (2930). For the other comparisons, we adopted a cluster-wise inference method, a degree-based statistic to find clusters significantly influenced by the disease (31). Finally, multiple linear regression was used to evaluate the explanatory power of the network properties in predicting the behavioral scores of patients.

The mean age of the patients with traumatic anosmia was significantly greater than that of healthy controls (p < 0.001), with no significant difference in the sex ratio (p = 1.000). The mean time between head injury and clinical consultation was 9.6 months (range, 3–24 months) in the anosmia group. Four of the 16 patients exhibited focal encephalomalacia on the structural MRI, with the maximum dimension ranging from < 1.0 cm in 3 patients to approximately 3.0 cm in one patient; the lesions were located at both gyri recti (n = 1), the left gyrus rectus (n = 2), or the right gyrus rectus (n = 1). The patient with the largest lesion showed involvement of the left medial orbital gyrus as well. There were no additional structural abnormalities in the brains of all subjects. In patients with anosmia, the mean KVSS II score was 9.44 ± 3.7 (range, 3–15), indicating that all patients met the diagnostic criteria for anosmia.

Compared with healthy controls, patients with anosmia exhibited different patterns of connectivity in the olfactory network (Fig. 1). Intra-cluster connectivity was decreased in patients, specifically in the insular cortex and anterior cingulate/frontal cortex (FDR-corrected p < 0.05). Inter-anatomical-cluster connectivity (across piriform cortex, insular cortex, prefrontal cortex, anterior cingulate/frontal cortex) was increased in the patients compared with that in the healthy controls (FDR-corrected p < 0.05).

Patients with traumatic anosmia showed increased connectivity in most brain regions (Figs. 2, 3) compared with that in healthy controls. Furthermore, the functional connectivity in the patients was primarily increased among the thalamic and sensory networks (visual and somatomotor hand) compared with the healthy controls (Table 1).

Compared with healthy controls, patients with traumatic anosmia showed significantly increased global efficiency (0.753 ± 0.056 vs. 0.704 ± 0.042) (p = 0.019) and decreased modularity (0.165 ± 0.044 vs. 0.245 ± 0.054) (p < 0.001) in the whole brain network.

In patients with traumatic anosmia, the global efficiency of the whole brain network was significantly correlated with the KVSS II score (r² = 0.71, p = 0.002) and remained significant after controlling for disease duration (r² = 0.75, p < 0.001) (Fig. 4). Modularity of the whole brain network was also significantly associated with the KVSS II score (r² = 0.79, p < 0.001) and remained significant after controlling for disease duration (r² = 0.82, p < 0.001).