Participants
This cross-sectional prospective analytical study was conducted as part of the NeuroCOVID-19 Brazilian Registry (NeuroCovBr, https://www.neurocovbr.com/),36 between October 2020 and May 2021 in Brasilia, Brazil. Participants were recruited with a non-probabilistic sampling strategy from a population of health professionals and patients assisted at the Brasilia University Hospital before the implementation of mass vaccination campaigns. We consecutively contacted a list of 364 patients diagnosed with COVID-19 by real-time quantitative reverse transcription polymerase chain reaction (qRT-PCR) to invite them to the study.
The inclusion criteria for the COVID-19 group (COV+) were: (a) diagnosis of SARS-CoV-2 infection confirmed by detection of viral RNA by qRT-PCR testing of a nasopharyngeal swab, without requiring hospitalization during infection, (b) COVID-19-related persistent subjective hyposmia, and (c) age between 18 and 60 years old. Patients were evaluated at least four weeks after the diagnosis of COVID-19 to ensure that the acute phase had already passed. The control group (COV-) was recruited from the same population (patients or health professionals from Brasilia University Hospital) using convenience sampling, with age, sex, and education level matched to the COV+ group. Subjects in the COV- group had not been previously infected with SARS-CoV-2, had a negative SARS-CoV IgG/IgM test, and had no olfactory dysfunction.
The exclusion criteria for both groups included (a) pre-existing brain structural disorders (e.g., stroke, epilepsy, multiple sclerosis, neoplasia, hydrocephalus, traumatic brain injury, Parkinson's disease, and dementia), (b) severe psychiatric diseases, (c) MoCA global score of less than 15,37 (d) MRI contraindications, and (e) illiteracy.
This study was approved by the local ethics committee at the University of Brasilia and adhered to current regulations, such as the Helsinki Declaration. All participants provided written informed consent and underwent clinical, cognitive, and MRI examinations during the same visit.
Clinical assessment
Demographic and clinical data were collected using electronic forms, including evaluation of neurological, chemosensory, respiratory, and constitutional symptoms. In addition, demographic variables such as age, education, sex, and a list of self-reported comorbidities were obtained.
The Sniffin' Sticks smell identification test (SS-16) was used to evaluate participants' ability to identify odors. This psychophysical test was developed by Burghardt®️ (Wedel, Germany) and previously adapted to Brazilian Portuguese.38,39 The test comprises 16 pens containing common and recognizable odorants. The length of each pen is 14 cm (approximately 5.51 in), with an internal diameter of 1.3 cm (about 0.51 in) and a 4 mL cap containing odorless or odorous liquids dissolved in propylene glycol. The participant had to identify the odor using a four-option forced-choice paradigm.
All participants also responded to a cognitive test, MoCA, to screen for cognitive impairment.37,40 It is a brief 30-point test that assesses attention, executive functions, memory, language, visuoconstructional skills, conceptual thinking, and calculations.
MRI data acquisition
The MRI was performed using a Philips Achieva 3T scanner (Best, Netherlands) equipped with an 8-channel SENSE coil. The following MRI sequences were obtained: (1) Three dimensional (3D) T1-weighted sequence, turbo field echo (TFE), sagittal, with field of view (FOV) = 208 × 240 × 256mm, reconstructed resolution of 1 × 1 × 1 mm, echo time (TE) = min full echo, repetition time (TR) = 2300ms, TI = 900ms, two times accelerated acquisition; (2) Diffusion-weighted sequence, axial, with FOV 232 × 232 × 160 mm, reconstructed resolution of 2 × 2 × 2 mm, TE = 71 ms; TR = 3300 ms, 32 directions (b = 800 s/mm2); (3) Diffusion-weighted sequence, axial, with FOV 232 × 232 × 160 mm, reconstructed resolution of 2 × 2 × 2 mm, TE = 71 ms; TR = 3300 ms (reversed phase encoded b0); (4) 3D-fluid attenuated inversion recovery (FLAIR) sequence, sagittal, with FOV 256 × 256 × 160 mm, reconstructed resolution of 1.2 × 1 × 1 mm, TE = 119 ms, TR = 4800 ms, TI = 1650 ms. (5) T2-weighted sequence, coronal, with FOV 264 x 204 mm, reconstructed resolution of 0.25 × 0.25 × 1.5mm, TR = 2500 ms, TE = 80 ms; flip angle = 90, with coverage of the anterior cranial fossa.
Manual segmentation of the olfactory bulbs
Two independent evaluators blinded to clinical and olfactory data manually segmented the volumes of the olfactory bulbs using ITK-SNAP (version 3.8) (Figure 3).41 The limits of the olfactory bulb in the coronal plane were determined by the surrounding cerebrospinal fluid, while an abrupt diameter change defined the posterior boundary of the olfactory bulb at the transition with the olfactory tract.42 The volumetric measures of the right and left olfactory bulbs were taken independently and then summed. The mean values established by the two evaluators were used in all subsequent analyses. Interobserver agreement was evaluated using Pearson's correlation coefficient and the DSC (Figure 3).
The estimated total intracranial volume (eTIV) was computed using FreeSurfer (version 7.1.1, http://surfer.nmr.mgh.harvard.edu) which normalized the volumes of the olfactory bulbs to eliminate biases caused by unequal head sizes.
Diffusion Magnetic Resonance Imaging Processing
TractoFlow was used to analyze dMRI and T1-weighted images (Figure 4).43 As an automated tool for processing dMRI, it extracts diffusion tensor imaging (DTI) measures. FA, MD, RD, and AD were calculated. Probabilistic whole-brain anatomically constrained particle filtering tractography was performed on a fiber orientation distribution function (fODF) of maximum spherical harmonics order of 6.44 The output of TractoFlow was then further processed via advanced steps to generate structural connectomes using SCILPY library version 1.0.0.45 Then, COMMIT2 with ball & sticks forward model was used to filter the raw tractogram and compute the COMMIT2 weights of each streamline.16,17
Voxel-based diffusion imaging analysis
The TBSS pipeline in FSL (version 6.0)46 was used to compare MRI metric differences between the COV+ and COV- groups (Figure 4). The FA maps were nonlinearly aligned to the FMRIB-58 map in the template space of the Montreal Neuroimaging Institute (MNI). The FA skeleton mean was computed following the deformable registration. The FA maps deformation fields were utilized for FA, MD, RD, and AD. The registered maps were projected onto the FA skeleton.
Network construction
The brain network is composed of nodes and edges. To determine the nodes within the network, we selected 171 grey matter regions of the brain from the AAL3 atlas.47 Each AAL3 region in standard MNI space was back-transformed to the participant’s native diffusion space. The COMMIT2-weighted tractogram and AAL3 parcellations were used to derive COMMIT2-weighted structural connectivity matrices (Figure 4). The COMMIT2 weight of a streamline is a measure that quantifies the contribution to the diffusion MRI signal of each streamline and is proportional to the cross-sectional area of the biological fibers along their path. By its turn, the COMMIT2 weight of a connection corresponds to the sum of the individual weights assigned by COMMIT2 to each streamline connecting two parcels of the matrix and was used as a marker of connectivity strength. Through its ability to take into account the tracking bias related to variations in bundle width, the COMMIT2 weight constitutes a more biological proxy than the frequently used streamline count.15 The possibility to inject priors about brain anatomy and its organization, and not only about microstructural properties, represents a powerful and novel way to tackle the false-positive problem in tractography and brain structural connectivity.16,17 COMMIT2-weighted 171 × 171 whole-brain matrices were computed.
Three-dimensional projections of structural connections and nodes were visualized using BrainNet Viewer (version 1.42),48 for comparison of COMMIT2 weight matrices and graph theory analyses.
Network‐based statistics
NBS was performed following Zalesky's methods with NBS Connectome (version 1.2) to determine the different connections.49,50 NBS is a statistical method based on graph theory and is often used to explore differences in the structural connectivity in the brain WM network. Typically, NBS analysis is conducted to identify subnetworks consisting of pairs of nodes and connections whose structural connectivity strength varies significantly between groups.
Network measures
The Brain Connectivity Toolbox (BCT) computed network measures for each subject.51 For global networks, betweenness centrality (corresponding to the fraction of all shortest paths in the network), modularity (reflecting the segregation of the network), assortativity (reflecting whether nodes tend to be connected to other nodes with similar strengths), participation (measure of diversity of intermodular connections), clustering coefficient (fraction of connected triangles around a node), mean strength (corresponding to the average of all the nodal strengths, where the nodal strength is the sum of the weights of links connected to the node), global efficiency (corresponding to the average inverse shortest path length in the network and inversely related to the characteristic path length), density (corresponding to the fraction of present connections to possible connections), characteristic path length (average of the shortest path length across all nodes), edge count, and small-worldness (ratio of average clustering coefficient to characteristic path length) were analyzed.
We analyzed regional network measures, calculated for each node, including betweenness centrality (number of shortest paths that pass through a node), clustering (fraction of connected triangles around a node), edge count, local efficiency (average of the inverse shortest path length in the neighborhood a node), nodal strength (sum of weights of links connected to the node), path length (shortest path length across the (average of the shortest path length across all nodes), and participation (a measure of the diversity of intermodular connections of a node).
Local network measures were calculated for the olfactory-related brain regions (olfactory cortex, gyrus rectus, medial orbital gyrus, anterior orbital gyrus, posterior orbital gyrus, lateral orbital gyrus, insula, hippocampus, parahippocampal gyrus, amygdala, caudate nucleus, putamen, pallidum, thalamus [mediodorsal medial nucleus and mediodorsal lateral nucleus], anterior cingulate cortex [subgenual, pregenual and supracallosal], and nucleus accumbens).52,53
MRI quality control
The MRI images were inspected for significant gross geometric distortion, mass movement, and signal drop artifacts to ensure their quality. For T1-weighted and dMRI images, a Nextflow pipeline for dMRI quality control (Dmriqc-flow) was also utilized.54
Statistical analysis
Demographic and clinical assessments
The demographic and clinical characteristics of the groups were compared using independent-sample t-tests for normally distributed continuous variables, Mann-Whitney tests for nonnormally distributed continuous variables, and χ2 for categorical variables. Fulfillment of the normality assumption was inspected through visual examination of variable distributions and the Shapiro-Wilk test. The significance level was set at p < 0.05. All statistical analyses were performed in R, version 4.1.0 (R Foundation for Statistical Computing, Vienna, Austria).
Segmentation of the olfactory bulbs
The olfactory bulb volumes were compared using a t-test for independent samples. The eTIV corrected the volumes obtained with the formula: (volume of the olfactory bulb/ eTIV) x 100.
The level of interobserver agreement for the segmentation of the olfactory bulbs was assessed by the Pearson's correlation coefficient and the DSC. The DSC is an overlap similarity index that reflects agreement in size and location. It ranges from 0 (no overlap) to 1 (complete overlap) (Figure 3). A satisfactory overlap exists when DSC > 0.70.55
Voxel-based diffusion imaging analysis
To test for group differences, a general linear model (GLM) with contrast was performed on VBA data. The TBSS framework46 includes nonparametric permutation testing (5,000 permutations) to correct multiple comparisons and threshold-free cluster enhancement (TFCE). Age and sex were used as nuisance covariates. Results were considered significant at p < 0.05, TFCE corrected for multiple comparisons. WM regions were named according to the Johns Hopkins University white-matter tractography atlas.
Network‐based statistics
Between-group differences (COV- > COV+ and COV- < COV+ contrasts) were tested on structural connectivity matrices for a range of primary thresholds (from t = 2.5 to t = 3.5), with age and sex as nuisance variables. Five thousand permutations were used, with intensity as the measure of network size and a statistical significance threshold set at p < 0.05.
Network metrics
Between-group differences were tested with either Mann-Whitney (modularity, clustering, and nodal strength) or independent-sample t-tests (other global network metrics). A GLM was used to analyze the local network metrics differences in olfactory-related brain areas between the control and COVID-19 groups, using age and sex as covariates. All results were corrected using the false discovery rate (FDR) method.56
We performed a partial correlation analysis between global network measures, SS-16 test score, and olfactory bulb volumes, adjusting for age, sex, and education. Data were analyzed using Spearman's coefficient. Statistical significance was defined as a two-tailed p < 0.05.