Studies into cortical thickness in psychiatric diseases based on T1-weighted MRI frequently report on aberrations in the cerebral cortex. Due to limitations in image resolution for studies conducted at conventional MRI field strengths (e.g. 3 Tesla (T)) this information cannot be used to establish which of the cortical layers may be implicated. Here we propose a new analysis method that computes one high-resolution average cortical profile per brain region extracting myeloarchitectural information from T1- weighted MRI scans that are routinely acquired at a conventional field strength. To assess this new method, we acquired standard T1-weighted scans at 3 T and compared them with state-of-the-art ultra-high resolution T1-weighted scans optimised for intracortical myelin contrast acquired at 7 T. Average cortical profiles were computed for seven different brain regions. Besides a qualitative comparison between the 3 T scans, 7 T scans, and results from literature, we tested if the results from dynamic time warping-based clustering are similar for the cortical profiles computed from 7 T and 3 T data. In addition, we quantitatively compared cortical profiles computed for V1, V2 and V7 for both 7 Tand 3 T data using a priori information on their relative myelin concentration. Although qualitative comparisons show that at an individual level average profiles computed for 7 T have more pronounced features than 3 T profiles the results from the quantitative analyses suggest that average cortical profiles computed from T1-weighted scans acquired at 3 T indeed contain myeloarchitectural information similar to profiles computed from the scans acquired at 7 T. The proposed method therefore provides a step forward to study cortical myeloarchitecture in vivo at conventional magnetic field strength both in health and disease. . . . . . Keywords 3 Tesla 7 Tesla Automatic segmentation Laminar structure Cortical profiles Myeloarchitectonics Introduction Electronic supplementary material The online version of this article (https://doi.org/10.1007/s12021-018-9356-2) contains supplementary The cerebral cortex has a distinct laminar structure and the material, which is available to authorized users. different layers are associated with different neuronal func- * Bart Ferguson tions. To better understand how neurological and psychiatric firstname.lastname@example.org diseases affect the brain it is not only important to determine which of the cortical brain regions are implicated but also to determine which of the layers may be involved in these re- Brain Center Rudolf Magnus, Department of Psychiatry, Brain Division, University Medical Center Utrecht, Utrecht University, gions (Charney et al. 2013). The layers of the cortex can be HPNR A01.126, Heidelberglaan 100, 3584, CG defined on the basis of the size, shape, and arrangement of Utrecht, The Netherlands neuronal cell bodies, known as cytoarchitecture (Brodmann Radiology Department, Imaging Division, University Medical 1909; von Economo and Koskinas 1925) or by the presence of Center Utrecht, Heidelberglaan 100, 3584, CG myelinated nerve fibers, referred to as myeloarchitecture Utrecht, The Netherlands (Vogt and Vogt 1903). The cerebral cortex can be parcellated Experimental Psychology, Helmholtz Institute, Utrecht University, into regions with the same cyto- or myeloarchitecture Heidelberglaan 1, 3584, CS Utrecht, The Netherlands (Nieuwenhuys 2013). Studying the cytoarchitecture, however, CNSR, Psykiatrisk Center Glostrup, Ndr. Ringvej 29-67, usually requires invasive staining methods and can therefore DK-2600 Copenhagen, Glostrup, Denmark only be applied post-mortem in humans (Zilles and Amunts 182 Neuroinform (2018) 16:181–196 2015). In contrast, the myeloarchitecture can be non- certain layers by comparing the thickness of adjacent gyri and invasively studied in vivo using magnetic resonance imaging sulci in a particular brain region (Wagstyl et al. 2015). (MRI), rendering MRI well-suited to study large groups of Here we propose a new fully automatic post-processing patients and/or healthy subjects. This is particularly important method to extract information on myeloarchitecture from a when studying psychiatric diseases, where the enrolment of T1-w scan that is routinely acquired at 3 T. This method (de- large groups is necessary because of the expected small effect scribedin detail inthe BMethods^ section) combines sizes associated with this type of research. deconvolution and profile realignment to compute one de- Different types of MRI contrasts can be used to study the tailed average T1-w based signal profile per brain region, as- myeloarchitecture including T1 contrast, for which the signal suming equal laminar organization within that region of variation in grey matter is for a large part explained by myelin interest. concentration (Eickhoff et al. 2005; Lorio et al. 2016; Stüber We start out with a basic simulation to demonstrate that et al. 2014). Other possible contrasts that can reveal laminar application of this method to a simulated T1-w volume results structure include T2*-weighted magnitude or phase images, in a detailed average profile that contains information on sig- which are putatively sensitive to iron or ferritin (Duyn et al. nal characteristics and to assess the sensitivity of the method 2007; Duyn and Schenck 2016; Fukunaga et al. 2010). Much to potential artefacts such as deconvolution-related Gibbs effort has been put into high-field high-resolution MRI myelin ringing. mapping, especially in the occipital and primary visual area For further assessment of the method we acquired whole (Abdollahi et al. 2014; Clark et al. 1992;Serenoetal. 2013; brain T1-w scans from three healthy volunteers that were Trampel et al. 2011), the auditory cortex (Wasserthal et al. scanned both at a 3 T and a 7 T scanner. For 3 T a standard 2014) and sensorimotor cortex (Weiss et al. 2011), and recent- T1-w acquisition protocol (3T-T1) was used while for 7 T a ly, whole brain (Fracasso et al. 2016). In addition, sensorimo- novel whole-brain, high-resolution (0.5 mm isotropic) T1-w tor (Sanchez-Panchuelo et al. 2012; Sánchez-Panchuelo et al. MPRAGE sequence (MS7T) was used that is adapted to be 2014) and auditory (De Martino et al. 2015;Dicket al. 2012) sensitive to myelin content within grey matter (Fracasso et al. cortical regions have been investigated with functional MRI, 2016). This scan was previously validated with histology also in relation to local myelination (Lutti et al. 2014). (Fracasso et al. 2016) and therefore considered apt to validate However, scanning large groups of subjects using ultra-high the profiles computed from the standard 3 T T1-w scans. field MRI is not straightforward because of the limited avail- In particular, the primary visual cortex (V1) is used to as- ability of ultra-high field scanners as well as the additional sess (MRI) acquisitions and methods that focus on financial costs involved. myeloarchitecture (Hopf 1955, 1956, 1957; Waehnert et al. A large body of studies that used MRI conducted at con- 2016). The reason for this is that in V1, layer 4 (containing ventional field strengths (e.g. 3 T) revealed cortical thinning the myelinated band called the outer line of Baillarger and in in, for instance, schizophrenia (Van Haren et al. 2011; V1 also known as the stria of Gennari; see Supplementary Fig. Wheeler et al. 2014;Xiao etal. 2015), depression (Tu et al. S1) is heavily myelinated and even visible to the naked eye. 2012) and bipolar disorder (Elvsashagen et al. 2013; Lan et al. The signal from V1 is often contrasted to that of the secondary 2014). Although the resolution of scans that were used to visual cortex (V2) which lies adjacent to V1 but has a different measure cortical thickness was considered fine enough to laminar organization. Besides comparing V1 with V2, we compute the inner (white matter) and outer (pial) boundary included V7, which is a (occipital) visual area as well but of the cortex, it was too coarse to obtain information on indi- has a lower myelin content compared to V1 and V2 (Glasser vidual layers, leaving the question whether this thinning was and Van Essen 2011). Four primary motor and sensorimotor due to a mechanism that affects all cortical layers of a partic- regions were also included in this study. Three of these regions ular region similarly, or that specific cortical layers were more (BA1, BA2, BA3b) lie within the primary somatosensory cor- affected than others. The highest resolution that can be tex and one region (BA4) is part of the motor cortex. achieved for these standard T1-weighted (T1-w) MRI scans In a first quantitative analysis we use hierarchical crisp is primarily determined by the available scanning time and the clustering based on dynamic time warping to determine if static magnetic field strength on which the scanner operates profiles computed in 3 T data cluster in a way similar to the and is typically in the order of 0.8–1.0 mm isotropic for 3 profiles computed in 7 T data. In a second quantitative analy- Tesla (T). A number of studies have used MRI at conventional sis we compare the profiles computed for the visual areas field strength to obtain information on the laminar structure. using prior knowledge. Here we expect that the signal mea- These studies, however, required special types of MRI scans, sured in the middle of the profile is the highest for V1 because with long acquisition times (Glasser and Van Essen 2011; of the line of Gennari in V1, followed by the signal measured Mangeat et al. 2015), only focussed on a particular part of at the same location for V2. For V7 we expect that the signal is the brain (Bridge et al. 2005; Walters et al. 2003), or did not relatively lowest. If the clustering patterns are significantly directly image the layers but only inferred the involvement of similar (first analysis) and the relative signal differences are Neuroinform (2018) 16:181–196 183 significant (second analysis) for 7 T profiles as well as 3 T Additional bootstrapping is applied over the alignment and profiles (and with the same ordering) then we consider this as averaging procedure to both increase stability and to be able an indication that 3 T profiles indeed bear information similar to assess stability of the average profile. These steps are ex- to 7 T profiles. In a third analysis we assess the possible plained in more detail, below. The software is currently under influence of acquisition-related Gibb’sringing. development but the code used in this study is available on request. Deconvolution Methods Deconvolution is a mathematical operation that removes Outline of the Method ‘blurring’ effects and was performed as follows. First, the resolution of the original volume (either 3T-T1 or MS7T) The inner (white to grey matter) and outer (pial) boundaries of was doubled in all three directions using nearest neighbour the cortex are computed from the T1-w scan acquired at 3 T interpolation. Next, 3D deconvolution was applied using the using FreeSurfer (Fischl 2012). These boundaries are then Parallel Iterative 3D Deconvolution toolbox (Wendykier superimposed on an upsampled and 3D–deconvoluted version 2009) (implemented in ImageJ / FIJI) with the following set- of the original 3 T scan (see Fig. 1) after which a set of signal tings: Wiener Filter Preconditioned Landweber method; 1 it- profiles is computed for each cortical region. The eration; 3D Gaussian-shaped point spread function (PSF) with deconvolution step increases the detail in extracted profiles a full-width at half maximum of 5 voxels in a volume with size at the expense of a reduction in signal to noise ratio (SNR). 25x25x25 voxels (Fig. 1f); anti-ringing step on; divergence To compensate for possible misalignment between profiles, detection on; gamma = 0. The anti-ringing step reduces ring- the set of profiles are aligned using parametric time warping ing at the boundaries of the volume. (ptw) (Eilers 2004; Gerretzen et al. 2015) and then averaged (increasing SNR) resulting in one average profile. With ptw, one profile is selected as the best representative for the set of Cortical Profile Sampling profiles, and all other profiles are then shifted and scaled lin- early towards this best representative profile. A major advan- Cortical profiles for a selected brain region were sampled tage of ptw is that it is computation friendly both in terms of using a combination of Freesurfer and in-house developed memory usage and computation time (Eilers 2004), and is software that was written in R (R Core Team 2014), and in therefore well-suited to align large sets of profiles. C using the minc software library (http://www.bic.mni.mcgill. Fig. 1 Result of deconvolution of 3 T and 7 T scans, coronal slice. Effect scan. Freesurfer white (red) and pial (yellow) surfaces are overlaid. of deconvolution shown in a coronal section of the occipital lobe for one Insets display signal intensity sampled along displayed orange ruler participant for the area of interest displayed in the white rectangle in (a). (marked by the white arrow), with 3DSlicer 4.4.0, which is 4.2 mm in The four panels on the right indicate the area of interest in an original (a) length and runs from white matter to CSF. Please note that this sample is and deconvoluted (b) 3 T T1w 0.8 mm isotropic scan and in an original drawn in a 2D projection and is not necessarily perpendicular to the (d) and deconvoluted (e) 7 T myelin-sensitive T1w 0.5 mm isotropic cortical mantle. (f) depicts the PSF used 184 Neuroinform (2018) 16:181–196 ca). Freesurfer label files were created using the thresholded laplacian equation (De Martino et al. 2013)orthe equi- Brodmann Area (BA) map for the regions BA1, BA2, BA3b, volume model (Waehnert et al. 2014). BA4a, V1 and V2 (Desikan et al. 2006; Fischl et al. 2004, To prevent loss of detail in the computed average cortical 2008). The label for V7 was created using the PALS B12 profile, a highly accurate alignment of the set of profiles is Visuotopic atlas (https://surfer.nmr.mgh.harvard.edu/fswiki/ required. Here the computation of the alignment transforma- PALS_B12). These label files were used to create pairs of tions using parametric time warping was not performed directly coordinates on the inner and outer boundary of the cortex on the profile data but on a version of the profiles from which (linkage nodes, see (Lerch and Evans 2005)). Thus, a signal the contribution of the slowly varying signal baseline was re- profile is here defined as the T1 signal measured along a moved. This was done to ensure that the alignment is based on straight line running from a point on the inner boundary to local changes in signal intensities that may represent laminar the corresponding point on the outer boundary (where the information rather than the less informative slowly varying correspondence by the points is created by the expansion of baseline of the profiles. For this, we modelled the baseline the outer surface from the inner surface). Information on local using a cubic spline with the degrees of freedom set to 7 and cortical curvature was retrieved from the standard Freesurfer subtracted the baseline from the original signal resulting in the output. Cortical thickness was calculated based on the version of the profile used to compute the transformations. Euclidian distance between inner and outer boundary coordi- To align a complete set of profiles, first, a single profile was nate pairs. Intensity values were sampled at 100 equidistant selected for which the weighted cross correlation is highest points along a profile running from the inner to the outer with all other profiles (a standard function in the ptw package). boundary using linear interpolation. The sampling of data This profile was then regarded as the best representative profile points was extended in both directions of the profiles with of the set of profiles and was subsequently used as target profile 30 points in each direction to allow for small shifts and scaling to which all other profiles were aligned. Shift and linear scale during the alignment of the set of profiles. transformation values were calculated using ptw with weighted cross correlation (triangle width parameter set to 20) as optimi- sation metric (Bloemberg et al. 2010; Meerts et al. 2004). The Profile Alignment parameter search space was not restricted. The alignment pa- rameters were applied to the original cortical profiles, and miss- As previously argued (Koopmans et al. 2011), simply averag- ing values, created at the beginning or end of a profile that has ing over a set of profiles results in a loss of detail. This may be been shifted inward, were replaced with the respectively first or due to small errors made during boundary extraction (e.g. white last value of the original profile. After alignment, the individual to gray matter boundary), for example due to suboptimal scan profiles were averaged, resulting in one average cortical profile. quality (e.g. subject motion), which could lead to suboptimal placement of inner and outer boundaries and thereby to trans- Bootstrapping lations and scaling deviances of profiles. Another reason may be the fact that differences in cortical curvature induce differ- One concern may be that of model shine-through, meaning that ences in relative thickness of the individual layers. That is, the the shape of the average profile is largely determined by the superficial layers are thinner in gyri compared to the deeper shape of the target profile. Although such model shine-through layers (and vice versa for the sulci) (Waehnert et al. 2014). would not render the method invalid (as the choice of the target To correct for the effects of the former — the translation profile is purely data-driven) it may reduce stability as the and scaling deviances — we developed a procedure based on choice of the target profile becomes crucial. To increase stabil- parametric time warping (ptw) (Eilers 2004) using the ‘ptw’ ity of the computed average profile, bootstrapping was applied package implemented in R (Gerretzen et al. 2015). The effects over the alignment and averaging step. A total of 500 bootstrap of the latter can be mitigated by averaging only over profiles samples were created. For each bootstrap sample, a random set that were computed for brain areas with a similar cortical of profiles (with the same number of profiles as the original set) curvature. We therefore included only profiles for a region was drawn with replacement from the original set of profiles. for which the corresponding cortical curvature was within This sample was then aligned to the best reference profile se- ±1 standard deviation from the mean cortical curvature of that lected from this sample and averaged, yielding an average pro- region. Further homogenization is obtained by only including file. Finally, computing the mean over the 500 average profiles profiles from parts of the region with a similar cortical thick- resulted in one bootstrap analysis mean (BAM) profile. ness. Here, ± 0.5 standard deviation from the mean thickness An additional advantage of the bootstrap procedure is that was used as selection criterion. We note that alternative is it can be used to obtain information on the level of stability methods to correct for local cortical curvature exist and may of features that reflect the level of detail along the BAM pro- be used in future implementations, for instance methods based file. One could select salient features, compute them for each on the heat conduction equation (Annese et al. 2004, 2005), a bootstrap sample average profile and then determine their Neuroinform (2018) 16:181–196 185 distributions. For example, local maxima (‘peaks’) and mini- sampling (1 mm isotropic voxels) after which Rician noise was ma (‘valleys’) detected for each bootstrap sample average added (SNR 10% of the background signal). The result is shown profile can serve as salient features. These peaks and valleys in Fig. 2c. are computed from a smoothed version of the input (cubic To mimic variations due to suboptimal pial and/or white mat- spline, degrees of freedom set to 15) to remove possible ter boundary detection, the begin and end coordinates of the 360 noise-related minima and maxima, and are defined as the po- individual profiles were varied by adding an offset (modelled sitions where the second derivative equals zero. The distribu- with Gaussian noise, sigma 0.2 mm) to the in-plane coordinates tion of the peaks and valleys computed over the 500 bootstrap of the individual profiles yielding the second sampling scheme sample average profiles can then be used to assess the stability shown in Fig. 2d. The 3D deconvolution step used in the com- of the peaks and valleys in the corresponding BAM profile. putation of the average profiles (for both sampling schemes) was identical to the one described above. Note that, although the profiles were sampled from a single (2D) slice, we must use a Simulation Experiment 3D model to be able to apply the 3D deconvolution step. To generate the BAM profile (the red solid line) and to assess the To assess the feasibility of the method we created a 3D spher- stability of the result, bootstrapping was applied as described ical model (isotropic voxel size 0.5 mm, 200x200x200 voxels) above. of the cortex serving as ground truth (Fig. 2a). A volume was To assess the contribution of the deconvolution and alignment created with a sphere containing 10 consecutive shells with step separately, BAM profiles were also generated with the ap- different intensity values (800, 700, 600, 680, 600, 680, 600, plication of alignment only or deconvolution only. 550, 450, 400, respectively) and a background value of 200 to simulate a cortex with two high signal intensity bands. A cross section of this cortex model is shown in Fig. 2b. Participants The ground truth cortical profile (the dashed black line) was computed by averaging 360 equally distributed (with a 1 degree Three healthy volunteers (all male, mean age 28.8 years (SD = interval) profiles within the middle slice (slice #100). To simulate 1.0)) participated in the study after having signed informed con- scanner-induced noise the model was blurred with an isotropic sent. All experimental procedures were conducted in accordance 3D Gaussian blurring kernel (sigma = 1 mm) followed by down- with the 1964 Declaration of Helsinki (most recently amended in Fig. 2 Results of simulation experiment. The black solid line shown in The applied realistic sampling scheme is shown in (d), in which white (a) represents the ground truth cortical profile and the red solid line lines represent lines along which the profile is sampled, here in 360 represents the computed BAM profile. The intermediate steps are samples with normal distributed random variation in coordinates of both shown with dashed lines: the blue line represents sampling without begin and endpoints. (e) displays distributions of peaks and valleys posi- alignment nor deconvolution, the green line represents alignment only, tions detected from the bootstraps samples for the computed BAM profile and the orange line represents deconvolution only. The mid-slice of the (the red solid line) noise-free and noisy model volume is shown in (b) and (c)respectively. 186 Neuroinform (2018) 16:181–196 2008, Seoul), and approved by the ethics committee of the motion between scans, first, a linear (rigid and affine) transfor- University Medical Center Utrecht. mation was computed that registers the intermediate scan to the MS7T scan using mutual information as optimisation metric. MRI Image Acquisition Next, the intermediate scan was registered to the 3T-T1 scan using a combination of linear (rigid and affine) registration with The 3 T MRI T1-w data were acquired using a 3 Tesla Philips cross correlation as optimisation metric followed by nonlinear Achieva scanner (Philips Healthcare, Best, Netherlands), with an (SyN) warping with cross correlation as optimisation metric. 8-channel head coil, and a 3D MPRAGE sequence (number of The transformations were combined and then used to align the excitations per inversion 180; TR/TE 10 ms/4.6 ms; flip-angle MS7T to the 3T-T1 scan, using Hamming windowed sinc 8°; FOV 240x240x160 mm; 200 slices, 0.8 mm isotropic voxel interpolation. size; SENSE parallel imaging factor 1.4 (RL) and 1.7 (AP) in both phase-encoding directions; inversion time 964.4 ms; gradi- Cortical Segmentation and Surface Construction ent non-linearity correction; total scan duration 602 s), hereafter referred to as 3T-T1. FreeSurfer was used for a cortical segmentation of and automatic The 7 T MRI T1-w data were acquired using a 7 Tesla Philips inner (white/grey matter) boundary and outer (pia mater) bound- Achieva scanner (Philips Healthcare, Best, Netherlands) with a ary delineation in 3 T data (v5.3.0, http://surfer.nmr.mgh.harvard. 32-channel receive head coil (Nova Medical, Wilmington, MA, edu/) (Fischl 2012). The results were visually inspected and man- USA). First, a conventional T1-w 3D–MPRAGE scan was ac- ually corrected if needed. In order to use the inner and outer quired (voxel size 0.8 mm isotropic, number of excitations per boundaries that were computed from the 3T-T1 scan to sample inversion 300; TR/TE 7 ms / 2.9 ms; flip-angle 8°; inversion time the MS7T scan the latter was registered towards the former. 1200 ms; time delay between inversion pulses 3500 ms; FOV Regions of interest were created by default by FreeSurfer’s 200x250x190.4; acceleration using SENSE 2.8; 238 slices; total recon-all pipeline. The thresholded Brodmann Area annotation scan duration 296.3 s). This scan has the same resolution and a (except for V7) from FreeSurfer was used (Fischl et al. 2008). similar contrast to the scan acquired at 3 T and served as an The distinction of these areas in the atlas and annotation relies on ‘intermediate’ scan to register the 3 T scan with the 7 T scans. information from ten post-mortem brains, and made generaliz- Second, a 3D proton density (PD) scan was acquired (voxel size able to other brains on basis of cortical folding. Note, however, 1.0 mm isotropic, TR/TE 5.7 ms / 2.8 ms; flip angle 1°; FOV that cytoarchitecture cannot be measured with MRI but that clas- 200x250x190; SENSE 1.8 (anterior-posterior) × 1.8 (right-left); sification of cytoarchitecture can only be inferred via 190 slices; total scan duration 49 s) which was used to correct the cytoarchitecture-myeloarchitecture correspondence, and is espe- 7 T T1-w scans for B1 field inhomogeneities (Marques et al. cially difficult in long stretches of cortex such as BA 1,2,3 and 4. 2010; Van de Moortele et al. 2009). Next, three myelin- sensitive T1-w MPRAGE scans were acquired (voxel size Contribution of Deconvolution and Alignment 0.5 mm isotropic, number of excitations per inversion 300; TR/ TE 7.5 ms / 3.5 ms; flip-angle 8°; inversion time 1200 ms; time The method combines deconvolution and profile realignment. delay between inversion pulses 6000 ms; FOV 200x250x180; To investigate the effect of the different processing steps, the SENSE 2.5 (anterior-posterior) × 2.5 (right-left); 360 slices; total automatic procedure was applied with and without scan duration 446.3 s, see (Fracasso et al. 2016)). These three deconvolution of the volume, and with and without parametric myelin-sensitive scans were aligned and averaged to increase time warping (ptw) alignment of the profiles, to the 3T-T1 and SNR. No correction for gradient non-linearity was applied for MS7T scan for one subject (PP3) in the left V1 region. the 7 T scans. To increase homogeneity, this average scan was then divided by a smoothed version (12 mm FWHM Gaussian Clustering kernel) of the PD scan resulting in one scan to which we hereafter refer to as MS7T. To quantitatively assess the reproducibility for BAM profiles computed from the 3 T with respect to the 7 T data, we perform 7 T to 3 T Image Registration two different analyses. In the first (fully data-driven) analysis we use a hierarchical crisp clustering algorithm based on the Because the 3T-T1 and MS7T scan differ in contrast and may be dynamic time warping distance (DTW) implemented in slightly deformed with respect to each other, as they were ac- dtwclust (https://cran.r-project.org/ package = dtwclust). The quired on different scanners, the intermediate scan (see above) aim is to determine to what extent the 7 T BAM profiles was used to optimise the registration. The registration was per- contain information specific to the various brain regions and formed using ANTs (Avants et al. 2011) and prior to the regis- to determine if clustering of the 3 T data produces similar tration all scans were corrected for effects of B0 inhomogeneity results. The DTW distance is a stretch-insensitive measure of usingN4(Tustisonet al. 2011). To correct for possible head the ‘inherent difference’ between two profiles (Giorgino 2009), Neuroinform (2018) 16:181–196 187 operating on information from the complete BAM profiles. An For this we compute the pairwise distance (V1 - V2) at the advantage of using DTW-based clustering is therefore that it center of the profiles (X = 50) and compute if it is significantly does not require us to define salient features beforehand, mak- larger than zero for both 7 T and 3 T data. ing these results more generalizable. This type of clustering In contrast to primary visual regions, higher order visual re- does however require us to determine the number of clusters gions like V7 have a lower myelin content (Glasser and Van beforehand. Because the primary goal is to determine to what Essen 2011). For cortical profiles computed from myelin- extent the BAM profiles computed from 3 T scans contain sensitive MRI scans this should result in relatively fast drop-off similar information to the 7 T BAM profiles we decided to in signal intensity in the direction from the white matter boundary use the GAP statistic to compute the optimal number clusters towards the pial boundary. Here we will measure the differences for the BAM profiles computed at 7 T. The cluster pattern between the V7 profile and the primary visual regions V1 and obtained for the 7 T BAM profiles will then serve as ground V2 at the same position (X = 50) as in the previous comparison. truth for the clustering results obtained with the same number Before assessing the intensity value of a BAM profile at that of clusters for the 3 T data. The amount of overlap between 7 T position, the BAM profiles were first aligned to each other. and 3 T cluster patterns is then computed using the adjusted Rand index (ARI) (Rand 1971; Vinh et al. 2010). The ARI is Gibbs Ringing bounded above by 1, and scaled in such a way that 0 represents the amount of overlap obtained by chance. Please note that the One concern may be that the peaks or valleys in the average ARI is not bounded at 0 but can be negative as well. profiles are not linked to myeloarchitectural properties but that To determine the level of significance of the amount of they are in fact the product of Gibbs ringing as these artefacts overlap we use randomization sampling with replacement show up as decaying bands in the vicinity of sharp transitions (1000 samples for the 3 T clustering) to compute a distribution in the image (e.g. cortex boundaries) and could be for the bootstrap ARI values under the null hypothesis. If the misinterpreted as laminar information. Gibbs ringing is not only ARI of the original cluster pattern is larger than 95% of the ARI of concern in post-processing (in the deconvolution step) but values in the distribution then the overlap is deemed signifi- especially in image acquisition. cant). We consider this an indication that 3 T BAM profiles are For acquisition-related Gibbs ringing the ‘size’ of the rings indeed similar to the BAM profiles computed at 7 T. Note that (that is, the frequency of the decaying over- and undershoot) is here we use randomization testing with replacement instead of directly related to the voxel size of the original scan (Haacke permutation testing. For each sample, the original BAM labels and Brown 1999). Because the voxel size of the MS7T scan is are assigned to 42 profiles that are randomly selected (with 0.5 mm isotropic and 0.8 mm isotropic for the 3T-T1 scan one replacement) from the original set of 3 T BAM profiles after would expect that artificial peaks (and valleys) due to Gibbs which clustering is applied and the ARI between the resulting ringing would be systematically found at different locations clustering pattern and the 7 T clustering pattern is computed. along the cortical depth when comparing the average profiles The reason for this is that the computed distance function for between the 3 T and 7 T scans. To determine if there is a clustering is invariant under permutation of the profiles and as a systematic difference in peak locations between 3 T and 7 T consequence the same profiles will form the same clusters, average profiles we computed the distance for two different although the cluster number itself may change. Of further note peaks with respect to the rightmost valley (distance I and is that because ARI is not a true distance measure (Vinh et al. distance II in Fig. 6) in V1. Values for distance I and distance 2010) we cannot use it to compute for instance effect sizes or to II are computed for both 3 T and 7 T profiles. If these values directly compare clustering results from 7 T data and 3 T results show a significant difference between 3 T and 7 T (computed with respect to a third ground truth clustering. using a paired t-test) then this is an indication that peaks (or valleys) are due to acquisition-related Gibbs ringing. V1 Versus V2 Versus V7 Comparison In the second analysis we do use a priori knowledge on differ- Results and Discussion ences between V1 and V2 as well as known differences in myelin concentration between higher order (V7) and lower Simulation Experiment order (V1, V2) visual cortices. The major difference between adjacent areas V1 and V2 is that layer 4 in V1 is heavily my- In Fig. 2 the results of the simulation are shown to demonstrate elinated (the line of Gennari) (Glasser and Van Essen 2011). the feasibility of the method. Simply averaging without re- This difference is often used to assess myelin-sensitive MRI alignment or deconvolution (the blue dashed line) results in scan sequences (Fracasso et al. 2016). Here we investigate if a low-detailed BAM profile in which the two separate high- this difference (expected to be maximal at approximately the intensity bands seen in the model (black solid line) cannot be center of the profiles) can be detected both in 7 T and 3 T data. identified. Averaging without realignment (but with 188 Neuroinform (2018) 16:181–196 deconvolution) is depicted by the orange dashed line while the alignment only (Fig. 3b and f) show a more pronounced local result of averaging with realignment but without the maximum for both the 3T-T1 and MS7T scans (with MS7T deconvolution step is depicted by the green dashed line. The showing a more pronounced local maximum than 3T-T1) than red solid line denotes the BAM profile computed using the the BAM profiles computed without deconvolution or ptw proposed method. In this average profile two peaks can be alignment (Fig. 3a and e). Figure 3c and g show the results detected which coincide with the high intensity bands in the when averaging is done over the profiles extracted from the model. Note, however, that in the BAM profile an additional deconvoluted scans but without the ptw alignment step. small peak is found on the left. Especially for MS7T this appears to result in a more pro- The distribution of the peaks and valleys computed for each of nounced curvature of the BAM profile in comparison to the the 500 bootstrap samples that were used to generate the BAM BAM profiles shown in Fig. 3a and b. Finally, Fig. 3dandh profile are presented in Fig. 2e. These peak and valley distribu- show the BAM profiles computed with deconvolution and ptw tions give us information on which of the peaks and valleys in the alignment. Both figures show more detail (e.g. multiple peaks) BAM profile can be reliably detected and may be linked to true suggesting that it is indeed the combination of deconvolution physical properties of the cortex. Note for instance that for this and realignment that leads to substantial more detail in the simulation there is virtually no overlap between peaks and val- computed BAM profiles. However, similar to the profiles leys suggesting that their positions are relatively constant and shownin3a–c and 3e–g it is evident that the peaks are more they could therefore serve as salient features. pronounced in Fig. 3h in comparison to 3d, suggesting that the Comparing the average model profile (black solid line) with 7 T data is (as expected) of better quality. Note that the peaks the reconstructed BAM profile (solid red line) suggests that no may be displaced when comparing the rightmost column artificial peaks or valleys are introduced. However, we do see a (where both alignment and deconvolution are applied) with slight over- and undershoot of the signal in the immediate vi- the other profiles in the other columns. This is because the best cinity of a sharp signal change (at depth X = 23). This shows reference profiles (chosen for the alignment in each bootstrap) that one should interpret peaks and valleys adjacent to a very are likely to differ (both in location and scaling) between the large signal intensity change (for T1-w contrast typically the BAM profiles shown in the rightmost column (where align- pial or white matter boundary) with caution. Besides this peak, ment is applied) and the rest. The location of a best reference no other spurious peaks or valleys were found, indicating that profile with respect to the cortex is determined by the estimated the effects of Gibbs ringing due to deconvolution are limited. inner and outer cortical boundaries computed with Freesurfer, From the ground truth (the black solid line) it can be seen which may not always be completely correct. Although the that there are three small peaks and two small valleys in the effect of the chosen reference profile on the shape of final declining slope on the right side of the profile, with X = 155 at BAM profile is limited (thus limited model shine-through) this is most likely not the case for the absolute location and size of its centre. The reconstructed BAM profile actually fails to show the middle of three small peaks (i.e. a false negative) the BAM profile. This also applies to the comparison of BAM although it can be observed from the peak/valley distribution profiles from 3 T versus 7 T. (panel e) that the peak is found in a considerable number of the When comparing results shown in Fig. 3 with the simula- bootstrap profiles (the second green peak from the right). tion (Fig. 2) one can see that in the simulation the various Apparently, this small peak is averaged out in the BAM profile computed profiles do line up with the ground truth. suggesting that, although the method can be used to detect the However, the presented simulation is a rather basic simulation larger maxima, more subtle details may still be missed. and is included to show the feasibility of the proposed method We note that (especially for the in vivo data) a one-to-one and to show that effects of Gibbs ringing on the final BAM correspondence between peaks and valleys over the bootstrap profile are limited. The fact that the peaks found in the simu- profiles may not exist as the number of peaks and valleys per lation using the full procedure are nicely aligned with the bootstrap profile may vary. Therefore, it is not straightforward ground truth may actually reflect the basic nature of the sim- to use the peak/valley distributions to calculate for instance ulation. The noise in the in vivo data may be of a more com- confidence intervals for a specific peak or valley location in plex nature and the variation in laminal makeup within a the BAM profile and for this reason we just present the histo- Brodmann area is likely larger than zero (as is assumed in grams of peaks and valleys found. the simulation). For example, in the simulation experiment alignment without deconvolution results in a fairly good (al- Contribution of Deconvolution and Alignment beit less pronounced) approximation of the ground truth while in the in vivo data (shown in Fig. 3) this is clearly not the case. The results of applying the automatic procedure to the left V1 We further note that we chose not to realign the bootstrap for subject 3 (PP3) with and without deconvolution, and with results before averaging them to compute the final BAM pro- and without alignment, and applying the bootstrap analysis are file although such a ‘second-level’ realignment step could be included. The motivation for this choice is that the histograms showninFig. 3. The BAM profiles computed with ptw Neuroinform (2018) 16:181–196 189 Fig. 3 Results of method applied on in vivo data in V1. Top row displays deconvoluted data; bottom row (e)–(h) depict result of identical process- 3 T results, with displaying BAM profiles and corresponding peak and ing, but for 7 T data. Note that the histograms are scaled to the maximum valley histograms of (a) non-deconvoluted nonaligned data, (b)aligned, number of peaks or valleys non-deconvoluted data, (c) non-aligned, deconvoluted data; (d) aligned (shown at the bottom of each panel) now also include the pronounced compared to other cortical areas. This is in con- inherent variability of location of the model profiles. cordance with the model profile obtained from the work from Dinse and colleagues. We do, however, recognize that besides similarities there Average Profiles 7 T and 3 T Data are also marked differences between the profiles computed for 3 T and 7 T data. For example, for BA1 in the left hemisphere In Fig. 4 the BAM profiles computed for the cortical areas we see two peaks in the 7 T profiles in subjects #2 and #3 BA1, BA2, BA3b, BA4a, V1, V2 and V7 are shown for all while there is only one peak in the corresponding 3 T profiles. subjects and for both MRI field strengths. For four regions For the right hemisphere in BA3 the profiles for 3 T and 7 T (BA1, BA2, BA3b, BA4a), a theoretical T1 model profile data look different, especially for subjects #1 and #2. For BA2 signal is available (Supplementary Fig. S2;(Dinseetal. (left hemisphere) the 7 T profiles show more detail than the 2015)) shown in the top row of the image for reference pur- 3 T profiles. A similar observation can be made for BA4a, poses. Dinse et al. (2015) used prior knowledge on both were especially for subject #3 the middle peak cannot be de- cytoarchitecture (e.g. cell density) and myeloarchitecture (lo- tected for 3 T while for 7 T the middle peak can be clearly cal myelin content) to model a T1-signal profile per brain area. detected for all three subjects. Overall, it appears the profiles For V1 a reference profile is provided that is based on high computed 3 T and 7 T data are more similar for the left hemi- resolution post-mortem MRI (See Supplemental Fig. 1). For sphere than for the right hemisphere. The histograms comput- V2 and V7 no reference profile is available. Note that because ed for 7 T suggest that, for most areas, peaks and valleys occur of the use of a best representative profile as target profile for in groups of four. For the 3 T histograms the pattern is less alignment the absolute positions of peaks and valleys found in consistent. In sum, this qualitative comparison clearly sug- BAM profiles computed for the same area may differ between gests (as could be expected) that the profiles computed for subjects and scanners. The comparison of BAM profiles 7 T are more detailed and more consistent over subjects and should therefore be made on relative features (e.g. relative hemispheres than the profiles computed for 3 T data. distances between minima and/or maxima). Between 20 and 25% of the FreeSurfer profiles are selected Most of the BAM profiles display multiple peaks and val- for processing due to the thickness and curvature constraints. leys. The general impression is that, except for BA1 and BA2, For detailed information on the initial number of FreeSurfer the middle peak is present in the BAM profiles both at 7 T and profiles and the selected number of profiles per region, see at 3 T. Particularly in the average profiles computed for V1 supplementary Table S1. acquired at 7 T, where this middle peak appears to be more 190 Neuroinform (2018) 16:181–196 Fig. 4 Results from seven cortical areas. BAM profiles of areas BA1, axis of these model profiles are inverted to match the signal intensity and BA2, BA3b, BA4a, V1, V2 and V7. The model profiles for BA 1, 2, 3b cortical depth measurement in our data. Original figures are provided as and 4a are provided as an indication of the expected myelin supplementary Fig. S1 (V1) and S2 (BA1–4). Note that the left V1 profile concentrations and were based on the model profiles presented in shown for subject #3 is a scaled version of the profile as shown in Fig. 3h (Dinse et al. 2015), where PP denotes participant. Here both x- and y- in order to align it with the corresponding profiles for subjects #1 and #2 Clustering Results computed between the 7 T and 3 T clustering patterns with seven clusters (k = 7) is 0.147 (1000 randomization samples; Maximum global GAP statistics computed for 7 T data sug- p = 0.011), which is still significant. When assessed with per- gest that the best number of clusters is 4 (See Fig. 5a). The mutation testing, the significance level for ARI computed with results for the hierarchical dtw-based clustering of the 7 T and four clusters reaches p < 0.00005 and with seven clusters p = the 3 T data, using four clusters (k = 4), are shown in Fig. 5b. 0.003. The ARI computed for 3 Tclustering with 7 Tclustering as the From the clustering pattern for 7 T presented in Fig. 5cit reference is 0.270, which is significantly larger (p <0.0005) can be seen that areas V1, V2 are grouped in the same cluster than zero (reflecting the amount of overlap by chance) sug- implying that these areas have similar features and contain gesting that indeed 3 T and 7 T profiles bear similar informa- similar information. Indeed, from previous literature it is tion. We note for the sake of completeness that the ARI known that the areas are similar in terms of local myelin Neuroinform (2018) 16:181–196 191 Fig. 5 Results from crisp clustering with dynamic time warping distance. presented according to participant (PP) and left and right hemisphere (lh, In (a) the GAP score plots are shown for both 7 T and 3 T data. Clustering rh) per column, and according to area per row. (e) displays BAM profiles results (with k = 4) for BAM profiles computed from seven areas, for 7 T from three participants, two hemispheres, of V1, V2 and V7, at 7 T. Error data and 3 T data are shown in (b). The colour of each block indicates the bars indicate +/− 1 SD at the centre of the profile. (f)displayssimilar belongingness of each BAM profile to which cluster. BAM profiles are BAM profiles, but at 3 T content and that the main difference between V1 and V2 is the hold to a large extent for the clusters computed for the 3 T presence of the line of Gennari in V1. The question is whether scans. However, one should keep in mind that the fact that the this difference can be reliably detected and used to further cluster pattern for 3 T are similar to the 7 T cluster pattern does distinguish between V1 and V2 at 7 T, and to what extent this not imply that the clusterings are based on the same profile can be done at 3 T. Note that BA3b is also part of this cluster, features. Dynamic time warp clustering is not based on salient and is like V1 and V2 also a primary sensory cortex. We features (e.g. peak and valley positions) but takes the complete further observe that V7 is clustered together with BA2. profile into account. Therefore, we cannot rule out that differ- From the reference profiles shown in Fig. 4 it is not immedi- ent weightings of profile features for 3 T and 7 T could still ately clear which of the primary somatosensory regions (BA1, lead to a similar clustering. BA2, BA3b) has a higher myelin concentration but from the profiles plotted in Fig. 4 it appears that indeed BA2 has a V1 Versus V2 Versus V7 Comparison lower overall T1 signal compared to BA1 and BA3b. Primary motor cortex BA4a appears to form a separate cluster. The signal difference between V1 and V2 measured at the This is in concordance with information from the reference middle (X = 50; Fig. 5e/f) of the BAM profiles per subject profiles where it can be seen that BA4a is expected to have the and per hemisphere is significant for both 7 T data (t(df = overall highest T1-signal over the profile. These observations 8.4) = 10.1; p < 0.00001) and 3 T data (t(df = 5.9) = 13.1; 192 Neuroinform (2018) 16:181–196 Table 1 Distance measures p < 0.00002). The signal difference between V7 and V2 mea- sured at the same position is also significant for both 7 T data Subject Volume Hemisphere Distance I Distance II Ratio II/I (t(df = 7.4) = 8.8; p < 0.00005) and 3 T data (t(df = 6.3) = 3.3; p < 0.015). In sum, the results from both analyses strongly 1 3 T Left 0.21 0.57 0.3684211 suggest that indeed profiles computed from 3 T scans bear Right 0.18 0.63 0.2857143 information similar to 7 T profiles albeit that the differences 7 T Left 0.14 0.53 0.2641509 between profiles for the V1, V2 and V7 are more pronounced Right 0.21 0.61 0.3442623 in the 7 T data. 2 3 T Left 0.25 0.62 0.4032258 Right 0.22 0.61 0.3606557 7 T Left 0.14 0.36 0.3888889 Gibbs Ringing Right 0.19 0.57 0.3333333 3 3 T Left 0.21 0.59 0.3559322 Acquisition-related Gibbs ringing in the MS7T volume has Right 0.16 0.54 0.2962963 been previously assessed by Fracasso and colleagues 7 T Left 0.18 0.51 0.3529412 (Fracasso et al. 2016) by drawing profiles through white mat- Right 0.29 0.68 0.4264706 ter at a white-gray matter boundary. The authors based their assessment on the rationale that if profiles sampled through gray matter show maxima that were due to Gibbs ringing, that computed for 3 T and 7 T scans (combining left and right same ringing would have to be present in cortical profiles regions). The mean distances for I and II (left and right com- sampled from white matter. Based on their results Fracasso bined) are for 3 T: 0.19 (SD = 0.03), 0.62 (0.10) and for 7 T: and colleagues concluded that Gibbs ringing could not explain 0.18 (0.04) and 0.54 (0.09) respectively. The mean fraction I/II the reported peaks in gray matter. computed for 3 T scans is 0.31 (SD = 0.03) and for 7 T 0.33 Here we performed an additional analysis to investigate the (0.44). The results for the paired t-tests between 7 T and 3 T possible role of acquisition-related Gibbs ringing. Based on data are t =0.39, p = 0.71 for distance I, t =0.96, p =0.39 for the bootstrap results for peaks and valleys locations two dis- distance II and t = −0.20, p = 0.85 for fraction I/II. Note that tances (denoted I and II in Fig. 6) are computed between two that these measures are all relative measures (distances or peaks and the rightmost valley for the average profiles of V1. fractions between distances) and are therefore invariant to The results are reported in Table 1. To take possible scaling possible effects of model shine-through. differences into account we also compared the fractions I/II The absence of a significant difference itself can of course not serve as proof for a limited effect of acquisition-related Gibbs ringing. In particular because it can be argued that the statistical power (n = 6) is very low. But as the difference in distance in case of Gibbs ringing should be around 38% (vox- el size 0.5 vs 0.8) one could expect that it can be detected with these small group sizes. Therefore, we think that in combina- tion with the evidence put forward in (Fracasso et al. 2016)the fact that there is no significant difference in the various dis- tances computed from 3 T and 7 T volumes with different voxel resolutions may at least serve as an indication that these peaks are not the product of acquisition-related Gibbs ringing. Limitations The central question in this study was: BAre cortical profiles computed from conventional 3 T scans using the proposed method (to a certain extent) similar to cortical profiles com- puted from this particular 7 T MRI acquisition that has previ- ously been validated with histology?^ and the generalizability of the results presented herein is therefore limited to these Fig. 6 Overview of distances used in assessment Gibbs ringing. This particular scans. However, we do not want to imply that this figure depicts the distances I and II, as measured in x-steps, divided by particular 7 T scan acquisition is the best possible solution the total 100 steps between white-gray matter and pial boundary, between available to study cortical myeloarchitecture (and is the the rightmost minima and first two peaks to the left of that minima. Results are shown in Table 1 ground truth) as the quality of a scan does not only depend Neuroinform (2018) 16:181–196 193 on the main magnetic field strength of the scanner but is the convergence but could introduce artefacts. The application of result of a complex interplay of many factors, including the the anti-ringing step is merely done as a precaution to limit use of high-performance RF coils, characteristics of the gra- spurious effects caused by features in the close vicinity of the dient system and chosen image contrast. The fact that a 3 T volume’s boundaries. profile is similar to 7 T profile does therefore not automatical- The resolution of the 3 T T-weighted scans in the current ly imply that the 3 T profile is free of artefacts, for instance, as study is relatively high (0.8 mm isotropic) in comparison with both scans are based on T1-weighted contrast and therefore scans typically acquired in large scale studies (e.g. 1 mm iso- T1-related artefacts common to both scan types may remain tropic) and one could question if the proposed method would unnoticed. work for lower resolution scans. But successful application of Although the results of the quantitative analyses presented the method will probably not only depend on resolution but here suggest that the proposed automatic method indeed for instance also on SNR and T1-weighting. We did however yields similar cortical profiles for 3 T compared to 7 T, we perform initial experiments (Supplementary Fig. S3), suggest- do recognise that at the individual profile level marked differ- ing that resolution itself appears not to be a critical factor and ences between 3 T and 7 T are found. The general impression the method can be applied in scans with lower resolutions. from the qualitative analyses is that the 7 T profiles are more The proposed method averages over sets of cortical profiles pronounced (showing clearer peaks and valleys) and are more but here we did not address the issue of optimal set size. consistent over subjects and hemispheres. This could be ex- Determination of the optimal set size is not straightforward pected as the 7 Tscans are not only of higher resolution but are as it depends amongst others on the homogeneity and thick- also optimised for contrast within the cortical gray matter. ness of the cortical layers in the area of interest, the quality of Thus, if possible, the obvious choice is to use a dedicated the cortex delineation, SNR and scanner resolution. If a set is MRI acquisition to investigate cortical myeloarchitecture. too small then the effects of noise will be insufficiently miti- However, if this is not possible (because of the large number gated. On the other hand, there is the risk that a set is too large. of subjects involved or because the scans already have been In that case the cortical architecture of the sampled region is acquired) then the proposed method may be of use to study (to insufficiently homogeneous leading to less detail. Further re- a certain extent) cortical myeloarchitecture at a group level. search is needed to determine optimal (and minimal) set size Furthermore, we would like to stress that we do not claim that for the various brain regions. Of further note, the sampling of the proposed method measures cytoarchitecture and that the the cortex, which in the current study was done along straight proposed method will have the same limitations that apply for lines, could be improved in future implementations by using T1-weighted MRI when used to study myeloarchitecture. sampling schemes that correct for cortical mantle curvature. For instance, corrections based on the heat conduction equa- tion (Annese et al. 2004; Annese et al. 2005). This would Methodological Considerations increase the number of profiles included per region and as a Although the results of these initial experiments are promising consequence profiles could be computed for smaller regions several aspects of the method require further investigation. For enabling the possibility to incorporate more detailed atlases instance, in the current implementation an isotropic Gaussian (Glasser et al. 2016; Hagmann et al. 2008). kernel with a fixed size (in voxels) was used in the Deconvolution is usually implemented using a Fourier deconvolution step. It can be argued that this point spread transformation – a mathematical operation carried out in the function itself does not adequately describe the effects of scan- complex domain. MRI data is complex-valued but usually only ning at a lower image resolution and that other types of the magnitude image is stored and the phase image is ignored. deconvolution kernels (e.g. different shapes, sizes) may be This phase image however may contain additional information more appropriate. We note, however, that here our goal was leading to better deconvolution results. For future studies one not to model the effect of differences in scanner resolution per may therefore consider to save the phase images as well. se but merely to increase image detail. The use of a Gaussian Possible applications of the proposed method can be found blurring kernel has the advantage that it mitigates the potential in studying the frequently reported cortical thinning in psychi- problem of introducing Gibbs ringing artefacts when applying atric diseases. One strategy would be to determine a number of deconvolution (Chandrawansa et al. 2000). Initial experi- characteristic points (e.g. peaks and valleys; one close to white ments (data not shown) suggested that the chosen kernel pa- matter, one approximately at the center of the profile and one rameters are adequate for this proof of concept. close to the pial surface) for the BAM profiles and then define For the other parameters for the Wiener filter Landweber a relative measure based on these characteristic points. For parallel deconvolution algorithm (besides the choice of the example, the fraction between center to white matter distance PSF) we chose the provided default settings but other settings and center to pial distance. Then, for ROIs for which group- may improve the results further. For example, here we chose related differences in thickness were found, a group compar- ison using this relative measure can be conducted to determine to disable the preconditioner which may lead to a speed up of 194 Neuroinform (2018) 16:181–196 if the effects of thinning are evenly distributed over all layers In conclusion, although the exact relation between the infor- or whether the thinning can be attributed to the deeper or more mation extracted from the average cortical profile and the superficial layers. We previously outlined such an approach to myeloarchitecture of the cortex requires further clarification, the study effects of schizophrenia (Mandl et al. 2015). results of our study suggest that this information obtained at 3 T We note that the proposed method may not only be useful to is in good agreement with information obtained at ultra-high field study psychiatric diseases but can also be used to study the de- MRI and can be used to study various architectural aspects of the veloping brain. For instance, brain development is accompanied cortex. The proposed fully automatic method therefore is a step by widespread cortical thinning (Van Soelen et al. 2012)and the forward to study the myeloarchitecture of the human cortex question is if this thinning reflects neuronal pruning (selective in vivo using scans routinely acquired on clinical MRI scanners, removal of neurons and connections) or that it actually reflects both in health and disease. white matter encroachment (where the myelin concentration of the deeper layers increases). In case of pruning the cortex actually thins but in case of encroachment the cortex only appears to Information Sharing Statement become thinner as the deep layers show a T1 signal similar to white matter. Applying the proposed method in a longitudinal The MRI data used in this paper was acquired in the UMC study design could be used to determine which of the two expla- Utrecht and is not part of an online repository. The following nations is correct. openly available software was used: R (RRID:SCR_001905, In the current study we focused on the similarities and differ- https://www.r-project.org), AFNI (RRID:SCR_005927, https:// ences in average cortical profiles computed from T1-w scans but afni.nimh.nih.gov/download), ANTs (RRID:SCR_004757, the proposed automatic method is not limited to T1-w contrast http://stnava.github.io/ANTs/), FreeSurfer (RRID:SCR_001847, per se. Similar to the use of the cortical boundaries computed on https://surfer.nmr.mgh.harvard.edu/fswiki/DownloadAndInstall), the 3T-T1 to process the MS7T scan, the cortical boundaries can FIJI (RRID:SCR_002285, https://imagej.net/Fiji/Downloads), be used to process other types of contrast (e.g. multi T1 (T1 with minc-library (RRID:SCR_002391, http://www.bic.mni.mcgill. multiple inversion times), T2, T2*, diffusion-weighted) scans ca). The in-house developed pipeline is available on request. from the same subject, provided that an accurate alignment can be obtained. In particular composition analysis based on multi T1 Acknowledgements This work was supported in part by a Netherlands Organization for Scientific Research (NWO) Vidi grant 13339 (N.P.) mapping is potentially interesting as it is an alternative way to obtain layer specific information at low resolution and may be Compliance with Ethical Standards The authors declare that they have no combined with the proposed method in the current study conflict of interest. All participants signed an informed consent before partic- (Lifshits et al. 2017). ipating. All experimental procedures were conducted in accordance with the 1964 Declaration of Helsinki (most recently amended in 2008, Seoul), and approved by the ethics committee of the University Medical Center Utrecht. Summary Here we introduced a new automatic method to compute detailed Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http:// average profiles for cerebral cortical regions from T1-w scans creativecommons.org/licenses/by/4.0/), which permits unrestricted use, routinely acquired at a clinical 3 T MRI scanner in vivo (3T- distribution, and reproduction in any medium, provided you give T1). To validate this new method, we compared BAM profiles appropriate credit to the original author(s) and the source, provide a link for various cortical regions (BA1, BA2, BA3b, BA4a, V1 and to the Creative Commons license, and indicate if changes were made. V2) computed from 3 T T1-w scans, with BAM profiles com- puted from 7 T myelin-sensitive T1-w scans (MS7T). The com- puted BAM profiles for the 3T-T1 and MS7T scans result in References similar clusterings. As expected, a significant higher relative sig- nal at the center of the profile (the approximate location of the Abdollahi, R. O., Kolster, H., Glasser, M. F., Robinson, E. C., Coalson, T. line of Gennari) is found for V1 compared to V2 for both 3 Tand S., Dierker, D., Jenkinson, M., Van Essen, D. C., & Orban, G. A. (2014). Correspondences between retinotopic areas and myelin 7 T data. Moreover, comparison of the BAM profile for V7 maps in human visual cortex. NeuroImage, 99,509–524. https:// (which has a lower myelin concentration than V1 and V2) with doi.org/10.1016/j.neuroimage.2014.06.042. the V1 and V2 profiles indeed shows a relatively fast signal drop Annese, J., Pitiot, A., Dinov, I. D., & Toga, A. W. (2004). A myelo- from the white matter boundary and a significant lower signal at architectonic method for the structural classification of cortical areas. NeuroImage, 21,15–26. https://doi.org/10.1016/j. the middle of the BAM profile for both 3 T and 7 T data. neuroimage.2003.08.024. Comparison of peak positions in BAM profiles of V1 computed Annese, J., Gazzaniga, M. S., & Toga, A. W. (2005). Localization of the from 3 T and 7 T scans suggests that the peaks have a location human cortical visual area MT based on computer aided histological that is irrespective of volume resolution and are therefore not the analysis. Cerebral Cortex, 15,1044–1053. https://doi.org/10.1093/ cercor/bhh205. result of Gibbs ringing. Neuroinform (2018) 16:181–196 195 Avants, B. B., Tustison, N. J., Song, G., Cook, P. A., Klein, A., & Gee, J. Elvsashagen,T.,Westlye,L.T.,Boen,E.,Hol,P.K.,Andreassen,O.A., Boye, B., & Malt, U. F. (2013). Bipolar II disorder is associated with C. (2011). A reproducible evaluation of ANTs similarity metric per- formance in brain image registration. NeuroImage, 54,2033–2044. thinning of prefrontal and temporal cortices involved in affect regula- https://doi.org/10.1016/j.neuroimage.2010.09.025. tion. Bipolar Disorders, 15,855–864. https://doi.org/10.1111/bdi. Bloemberg, T. G., Gerretzen, J., Wouters, H. J. P., Gloerich, J., van Dael, M., Wessels,H. J.C. T.,vanden Heuvel,L.P.,Eilers,P.H.C.,Buydens, L. Fischl, B. (2012). FreeSurfer. NeuroImage, 62,774–781. https://doi.org/ M. C., & Wehrens, R. (2010). Improved parametric time warping for 10.1016/j.neuroimage.2012.01.021. proteomics. Chemometrics and Intelligent Laboratory Systems, 104, Fischl, B., Van Der Kouwe, A., Destrieux, C., Halgren, E., Ségonne, F., Salat, 65–74. https://doi.org/10.1016/j.chemolab.2010.04.008. D. H., Busa, E., Seidman, L. J., Goldstein, J., Kennedy, D., Caviness, V., Bridge, H., Clare, S., Jenkinson, M., Jezzard, P., Parker, A. J., & Makris,N., Rosen, B.,&Dale, A.M.(2004). Automatically Matthews, P. M. (2005). Independent anatomical and functional Parcellating the human cerebral cortex. Cerebral Cortex, 14,11–22. measures of the V1/V2 boundary in human visual cortex. Journal https://doi.org/10.1093/cercor/bhg087. of Vision, 5,1–1. https://doi.org/10.1167/5.2.1. Fischl, B., Rajendran, N., Busa, E., Augustinack, J., Hinds, O., Yeo, B. T. Brodmann, K. (1909). Vergleichende Lokalisationslehre der Grosshirnrinde. T., Mohlberg, H., Amunts, K., & Zilles, K. (2008). Cortical folding doi:https://doi.org/10.1097/00005053-191012000-00013. patterns and predicting cytoarchitecture. Cerebral Cortex, 18, 1973– 1980. https://doi.org/10.1093/cercor/bhm225. Chandrawansa, K., Ruymgaart, F. H., & Van Rooij, A. C. M. (2000). Controlling the Gibbs phenomenon in noisy deconvolution of irregular Fracasso, A., Van Veluw, S. J., Dumoulin, S. O., Zwanenburg, J. M., & multivariable input signals. Journal of Applied Mathematics and Petridou, N. (2016). Lines of Baillarger in vivo and ex vivo; myelin Stochastic Analysis, 13,1–14. contrast across lamina at 7T MRI and histology. NeuroImage, 133, 163–175. Charney, D.S., Nestler, E.J., Sklar, P., & Buxbaum, J.D. (Eds.) (2013. Fukunaga, M., Li, T.-Q., van Gelderen, P., de Zwart, J. A., Shmueli, K., Yao, Neurobiology of mental illness, 4th edn. New York: Oxford B., Lee, J., Maric, D., Aronova, M. A., Zhang, G., Leapman, R. D., University Press. Schenck, J.F.,Merkle,H.,&Duyn,J.H.(2010).Layer-specificvaria- Clark, V. P., Courchesne, E., & Grafe, M. (1992). In vivo myeloarchitectonic tion of iron content in cerebral cortex as a source of MRI contrast. analysis of human striate and extrastriate cortex using magnetic reso- Proceedings of the National Academy of Sciences of the United States nance imaging. Cerebral Cortex, 2,417–424. of America, 107, 3834–3839. https://doi.org/10.1073/pnas.0911177107. De Martino, F., Zimmermann, J., Muckli, L., Ugurbil, K., Yacoub, E., & Gerretzen, J., Eilers, P., Wouters, H., & Bloemberg, T. (2015). R package Goebel, R. (2013). Cortical depth dependent functional responses in Bptw.^ https://cran.r-project.org/web/packages/ptw/index.html. humans at 7T: Improved specificity with 3D GRASE. PLoS One, 8, 30–32. https://doi.org/10.1371/journal.pone.0060514. Giorgino, T. (2009). Computing and visualizing dynamic time warping alignments in R :The dtw package. Journal of Statistical Software, De Martino, F., Moerel, M., Xu, J., Van De Moortele, P. F., Ugurbil, K., 31,1–24. https://doi.org/10.18637/jss.v031.i07. Goebel, R., Yacoub, E., & Formisano, E. (2015). High-resolution Glasser, M. F., & Van Essen, D. C. (2011). Mapping human cortical areas mapping of myeloarchitecture in vivo: Localization of auditory in vivo based on myelin content as revealed by T1- and T2-weighted areas in the human brain. Cerebral Cortex, 25,3394–3405. https:// MRI. The Journal of Neuroscience, 31, 11597–11616. https://doi.org/ doi.org/10.1093/cercor/bhu150. 10.1523/JNEUROSCI.2180-11.2011. Desikan, R. S., Segonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., Buckner, R. L., Dale, A. M., Maguire, R. P., Hyman, B. Glasser,M.F.,Coalson,T. S., Robinson,E. C., Hacker,C.D.,Harwell, J., T., Albert, M. S., & Killiany, R. J. (2006). An automated labeling system Yacoub,E.,Ugurbil,K.,Andersson,J.,Beckmann,C.F.,Jenkinson,M., for subdividing the human cerebral cortex on MRI scans into gyral Smith, S. M., & Van Essen, D. C. (2016). A multi-modal parcellation of based regions of interest. NeuroImage, 31,968–980. https://doi.org/10. human cerebral cortex. Nature, 536,171–178. https://doi.org/10.1038/ 1016/j.neuroimage.2006.01.021. nature18933. Haacke, M., & Brown, R. (1999). MRI physical principles and sequence Dick,F.,Tierney,A.T.,Lutti,A.,Josephs,O.,Sereno,M.I.,& Weiskopf,N. design. New York: John Wiley & Son. (2012). In vivo functional and myeloarchitectonic mapping of human primary auditory areas. The Journal of Neuroscience, 32, 16095– Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C. J., Van 16105. https://doi.org/10.1523/JNEUROSCI.1712-12.2012.In. Wedeen, J., & Sporns, O. (2008). Mapping the structural core of human cerebral cortex. PLoS Biol, 6, 1479–1493. https://doi.org/10.1371/ Dinse, J., Härtwich, N., Waehnert, M. D., Tardif, C. L., Schäfer, A., journal.pbio.0060159. Geyer, S., Preim, B., Turner, R., & Bazin, P.-L. (2015). A Cytoarchitecture-driven myelin model reveals area-specific signa- Hopf, A. (1955). Über die Verteilung myeloarchitektonischer Merkmale tures in human primary and secondary areas using ultra-high reso- in der isokortikalen Schläfenlappenrinde beim Menschen. Journal lution in-vivo brain MRI. NeuroImage, 1–17. https://doi.org/10. fürHirnforschung,2,36–54. 1016/j.neuroimage.2015.04.023. Hopf, A. (1956). Über die Verteilung myeloarchitektonischer Merkmale Duyn, J. H., & Schenck, J. (2016). Contributions to magnetic susceptibility in der Stirnhirnrinde beim Menschen. Journal für Hirnforschung, 3, 79–104. of brain tissue. NMR in Biomedicine. https://doi.org/10.1002/nbm. 3546. Hopf, A. (1957). Über die Verteilung myeloarchitektonischer Merkmale Duyn, J. H., van Gelderen, P., Li, T.-Q., de Zwart, J. A., Koretsky, A. P., & in der isokortikalen Scheitellappenrinde beim Menschen. Journal Fukunaga, M. (2007). High-field MRI of brain cortical substructure für Hirnforschung, 3,79–104. based on signal phase. Proceedings of the National Academy of Koopmans, P. J., Barth, M., Orzada, S., & Norris, D. G. (2011). Multi-echo Sciences of the United States of America, 104, 11796–11801. fMRI of the cortical laminae in humans at 7 T. NeuroImage, 56, 1276– https://doi.org/10.1073/pnas.0610821104. 1285. https://doi.org/10.1016/j.neuroimage.2011.02.042. Eickhoff, S., Walters, N. B., Schleicher, A., Kril, J., Egan, G. F., Zilles, K., Lan, M. J., Chhetry, B. T., Oquendo, M. A., Sublette, M. E., Sullivan, G., Watson, J. D. G., & Amunts, K. (2005). High-resolution MRI re- Mann, J. J., & Parsey, R. V. (2014). Cortical thickness differences flects myeloarchitecture and cytoarchitecture of human cerebral cor- between bipolar depression and major depressive disorder. Bipolar tex. Human Brain Mapping, 24,206–215. https://doi.org/10.1002/ Disorders, 16,378–388. https://doi.org/10.1111/bdi.12175. hbm.20082. Lerch, J. P., & Evans, A. C. (2005). Cortical thickness analysis examined Eilers, P. H. C. (2004). Parametric time warping. Analytical Chemistry, through power analysis and a population simulation. NeuroImage, 76,404–411. https://doi.org/10.1021/ac034800e. 24,163–173. https://doi.org/10.1016/j.neuroimage.2004.07.045. 196 Neuroinform (2018) 16:181–196 Lifshits, S., Tomer, O., Shamir, I., Barazany, D., Tsarfaty, G., Rosset, S., surface-based morphometry study. Psychiatry Res – Neuroimaging, 202,206–213. https://doi.org/10.1016/j.pscychresns.2011.07.011. & Assaf, Y. (2017). Resolution considerations in imaging of the cortical layers. NeuroImage,1–9. https://doi.org/10.1016/j. Tustison, N. J., Cook, P. A., & Gee, J. C. (2011). N4ITK: Improved N3 neuroimage.2017.02.086. bias correction. IEEE Transactions on Medical Imaging, 29, 1310– 1320. https://doi.org/10.1109/TMI.2010.2046908.N4ITK. Lorio, S., Kherif, F., Ruef, A., Melie-Garcia, L., Frackowiak, R., Ashburner, Van de Moortele, P. F., Auerbach, E. J., Olman, C., Yacoub, E., Ugurbil, J., Helms, G., Lutti, A., & Draganski, B. (2016). Neurobiological origin K., & Moeller, S. (2009). T1 weighted brain images at 7 tesla unbi- of spurious brain morphological changes: A quantitative MRI study. ased for proton density, T2* contrast and RF coil receive B1 sensi- Human Brain Mapping, 37, 1801–1815. https://doi.org/10.1002/hbm. tivity with simultaneous vessel visualization. NeuroImage, 46,432– 446. https://doi.org/10.1016/j.neuroimage.2009.02.009. Lutti, A., Dick, F., Sereno, M. I., & Weiskopf, N. (2014). Using high- Van Haren, N. E. M., Schnack, H. G., Cahn, W., Van den Heuvel, M. P., resolution quantitative mapping of R1 as an index of cortical Lepage, C., Collins, L., Evans, A. C., Hulshoff Pol, H. E., & Kahn, R. myelination. NeuroImage, 93(Pt 2), 176–188. https://doi.org/10.1016/j. S. (2011). Changes in cortical thickness during the course of illness in neuroimage.2013.06.005. schizophrenia. Archives of General Psychiatry, 68,871–880. Mandl, R.C.W., Brouwer, R.M., Van den Heuvel, M.P., Scheeuwe, T.W., Van Soelen, I. L. C., Brouwer, R. M., Van Baal, G. C. M., Schnack, H. G., Cahn, W., Kahn, R.S., & Hulshof-fPol, H.E. (2015). Schizophrenia- Peper, J. S., Collins, D. L., Evans, A. C., Kahn, R. S., Boomsma, D. related thinning of the outer cortical laminae in Brodmann areas 44, I., & Hulshoff Pol, H. E. (2012). Genetic influences on thinning of 45 revealed using MRI at 3T. In: Colorado Springs: ICOSR. doi: the cerebral cortex during development. NeuroImage, 59, 3871– https://doi.org/10.1093/schbul/sbv010.abstractnr2116177. 3880. https://doi.org/10.1016/j.neuroimage.2011.11.044. Mangeat, G., Govindarajan, S. T., Mainero, C., & Cohen-Adad, J. (2015). Vinh, N. X., Epps, J., & Bailey, J. (2010). Information theoretic measures Multivariate combination of magnetization transfer, T2* and B0 for Clusterings comparison: Variants, properties, normalization and orientation to study the myelo-architecture of the in vivo human correction for chance. Journal of Machine Learning Research, 11, cortex. NeuroImage, 119,89–102. https://doi.org/10.1016/j. 2837–2854. neuroimage.2015.06.033. Vogt, O., & Vogt, C. (1903). Zur anatomischen Gliederung des Cortex Marques, J. P., Kober, T., Krueger, G., van der Zwaag, W., van de cerebri. Journal für Psychologie und Neurologie, 2,160–180. Moortele, P.-F., & Gruetter, R. (2010). MP2RAGE, a self bias- von Economo, C.F., & Koskinas, G.N. (1925). Die cytoarchitektonik der field corrected sequence for improved segmentation and T1- hirnrinde des erwachsenen menschen. Vienna: Julius Springer. mapping at high field. NeuroImage, 49, 1271–1281. https://doi. Waehnert, M. D., Dinse, J., Weiss, M., Streicher, M. N., Waehnert, P., org/10.1016/j.neuroimage.2009.10.002. Geyer, S., Turner, R., & Bazin, P.-L. (2014). Anatomically motivat- Meerts, W. L., Schmitt, M., & Groenenboom, G. C. (2004). New applications ed modeling of cortical laminae. NeuroImage, 93(Pt 2), 210–220. of the genetic algorithm for the interpretation of high-resolution spectra. https://doi.org/10.1016/j.neuroimage.2013.03.078. Canadian Journal of Chemistry, 82,804–819. https://doi.org/10.1139/ Waehnert, M.D., Dinse, J., Schäfer, A., Geyer, S., Bazin, P.L., Turner, R., v04-037. Tardif, C.L. (2016). A subject-specific framework for in vivo Nieuwenhuys, R. (2013). The myeloarchitectonic studies on the human ce- myeloarchitectonic analysis using high resolution quantitative rebral cortex of the Vogt-Vogt school, and their significance for the MRI. NeuroImage, 125, 94–107. doi:https://doi.org/10.1016/j. interpretation of functional neuroimaging data. Brain Structure & neuroimage.2015.10.001. Function, 218,303–352. https://doi.org/10.1007/s00429-012-0460-z. Wagstyl, K., Ronan, L., Goodyer, I. M., & Fletcher, P. C. (2015). Cortical R Core Team (2014). R: A Language and Environment for Statistical thickness gradients in structural hierarchies. NeuroImage, 111,241– Computing. R Foundation for Statistical Computing, Vienna, 250. https://doi.org/10.1016/j.neuroimage.2015.02.036. Austria. http://www.R-project.org/. Walters, N. B., Egan, G. F., Kril, J. J., Kean, M., Waley, P., Jenkinson, M., Rand, W. M. (1971). Objective criteria for the evaluation of clustering & Watson, J. D. G. (2003). In vivo identification of human cortical methods. Journal of the American Statistical Association, 66,846–850. areas using high-resolution MRI : An approach to cerebral structure Sanchez-Panchuelo, R. M., Besle, J., Beckett, A., Bowtell, R., Schluppeck, – Function correlation. PNAS, 100,2981–2986. D., & Francis, S. (2012). Within-digit functional parcellation of Wasserthal, C., Brechmann, A., Stadler, J., Fischl, B., & Engel, K. (2014). Brodmann areas of the human primary somatosensory cortex using Localizing the human primary auditory cortex in vivo using struc- functional magnetic resonance imaging at 7 tesla. The Journal of tural MRI. NeuroImage, 93(Pt 2), 237–251. https://doi.org/10.1016/ Neuroscience, 32, 15815–15822. https://doi.org/10.1523/jneurosci. j.neuroimage.2013.07.046. 2501-12.2012. Weiss, M., Geyer, S., Lohmann, G., Trampel, R., & Turner, R. (2011). Sánchez-Panchuelo, R. M., Besle, J., Mougin, O., Gowland, P., Bowtell, Quantitative T1 mapping at 7 tesla identifies primary functional R., Schluppeck, D., & Francis, S. (2014). Regional structural differ- areas in the living human brain. Proceedings on International ences across functionally parcellated Brodmann areas of human pri- Society for Magnetic Resonance in Medicine, 19(19), 2360. mary somatosensory cortex. NeuroImage, 93,221–230. https://doi. Wendykier, P. (2009). Parallel Iterative Deconvolution. https://imagej.net/ org/10.1016/j.neuroimage.2013.03.044. Parallel_Iterative_Deconvolution. Sereno, M. I., Lutti, A., Weiskopf, N., & Dick, F. (2013). Mapping the human Wheeler, A. L., Chakravarty, M. M., Lerch, J. P., Pipitone, J., Daskalakis, cortical surface by combining quantitative T1 with Retinotopy. Cerebral Z. J., Rajji, T. K., Mulsant, B. H., & Voineskos, A. N. (2014). Cortex, 23,2261–2268. https://doi.org/10.1093/cercor/bhs213. Disrupted prefrontal interhemispheric structural coupling in schizo- Stüber, C., Morawski, M., Schäfer, A., Labadie, C., Wähnert, M., Leuze, C., phrenia related to working memory performance. Schizophrenia Streicher, M., Barapatre, N., Reimann, K., Geyer, S., Spemann, D., & Bulletin, 40,914–924. https://doi.org/10.1093/schbul/sbt100. Turner, R. (2014). Myelin and iron concentration in the human brain: A Xiao, Y., Lui, S., Deng, W., Yao, L., Zhang, W., Li, S., Wu, M., Xie, T., quantitative study of MRI contrast. NeuroImage, 93,95–106. https:// He, Y., Huang, X., Hu, J., Bi, F., Li, T., & Gong, Q. (2015). Altered doi.org/10.1016/j.neuroimage.2014.02.026. cortical thickness related to clinical severity but not the untreated Trampel, R., Ott, D. V. M., & Turner, R. (2011). Do the congenitally blind disease duration in schizophrenia. Schizophrenia Bulletin, 41,201– have a stria of gennari? First intracortical insights in vivo. Cerebral 210. https://doi.org/10.1093/schbul/sbt177. Cortex, 21,2075–2081. https://doi.org/10.1093/cercor/bhq282. Zilles, K., & Amunts, K. (2015). Anatomical basis for functional special- Tu,P.C., Chen,L. F.,Hsieh, J.C.,Bai,Y. M.,Li, C.T., &Su, T.P.(2012). ization. In K. Uludağ &K. Uğurbil (Eds.), fMRI: From nuclear Regional cortical thinning in patients with major depressive disorder: A spins to brain functions (pp. 27–66). New York: Springer US.
Neuroinformatics – Springer Journals
Published: Jan 19, 2018
It’s your single place to instantly
discover and read the research
that matters to you.
Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.
All for just $49/month
Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly
Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.
Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.
Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.
All the latest content is available, no embargo periods.
“Hi guys, I cannot tell you how much I love this resource. Incredible. I really believe you've hit the nail on the head with this site in regards to solving the research-purchase issue.”Daniel C.
“Whoa! It’s like Spotify but for academic articles.”@Phil_Robichaud
“I must say, @deepdyve is a fabulous solution to the independent researcher's problem of #access to #information.”@deepthiw
“My last article couldn't be possible without the platform @deepdyve that makes journal papers cheaper.”@JoseServera