Introduction

The cognitive abilities of parrots are known to be highly developed: they involve sophisticated problem solving, mirror use¹ and tool making to name a few. The parrots’ higher-order cognitive functions are thought to match those exhibited by crows and primates². However, little is known about the anatomy and structural connectivity of the parrot’s brain, the existing avian brain atlases being so far focused on other species such as pigeons³, budgerigars⁴ and quails⁵. In this work, we propose to use the power of diffusion MRI along with the strong gradients of a unique 17.2T preclinical MRI system to deliver a first white matter atlas of the parrot brain.

Materials and methods

Acquisition Protocol – Experiments were performed on a 17.2T horizontal animal MRI system (Bruker BioSpin, Ettlingen, Germany) equipped with strong gradients (1000mT/m-9600T/m/s). A 3D diffusion-weighted Pulsed Gradient Spin Echo echoplanar (EPI) sequence was implemented with the following parameters: 13 shots, TR/TE=250/23.8ms, diffusion gradients characteristics δ/∆=5.0/12.3ms, three diffusion shells including 25/60/90 diffusion directions with b-values of 1500/4500/8000s/mm² and 17 b=0s/mm² reference images, isotropic resolution of 180µm, matrix size=150x150x160, total acquisition time 30h.
Animal cohort – The MRI protocol was applied to three ex-vivo adult Pyrrhura molinae parrots. After intracardiac perfusion (4% paraformaldehyde + Gd-DOTA), brains were collected, fixed by immersion and scanned on the 17.2T MRI system.
Pre-processing – Data were denoised with a non-local means algorithm using the Ginkgo toolbox⁶. The PGSE EPI 3D sequence used with a large number of segments (13) resulted in raw DW images deprived of any visible geometrical distortion.
Template construction – The 17 T2-weighted b=0s/mm² images were averaged and a precise mask of the brain was computed for each individual. Intensity bias was removed using N4 bias field correction⁷ and all three unbiased images were provided to the antsMultivariateTemplateConstruction2 framework⁸. The resulting T2-weighted template was then semi-automatically segmented into left hemisphere, right hemisphere and cerebellum.
Post-processing – Orientation distribution function (ODF) maps were computed using the analytical Q-ball model⁹ (spherical harmonics order 8, regularization factor 0.006). A deterministic regularized streamline tractography¹⁰ was performed to infer a dense whole brain connectogram per animal (~1.856.000 connections) using the following parameters: uniform seeding over a predefined domain of propagation computed from the average b=0s/mm² image (8 seeds per voxel), forward step: 45µm, aperture angle: 30°. A lower GFA threshold of 0.03 was used as a stopping criterion for tractography. The processing pipeline is displayed in Fig.1.
Intra-subject fiber clustering – After subdividing the individual tractograms into 4 subsets (inter-hemispheric, right/left hemispheres, cerebellum fibers), and 3 length groups (1-7/7-13/13-19mm, Fig.2), a density mask was computed for each group with a 180µm resolution, allowing to perform a hierarchical clustering of the connectivity matrix established from a random parcellation obtained within each density mask using a K-Means algorithm (minimum cluster size of 300 voxels, average cluster size of 3000 voxels, connectivity matrix threshold of 1%). All fibers intersecting a given cluster by more than 33% of their length were assigned to a target fascicle, yielding an individual set of fascicles for each parrot. For each fascicle, a centroid was defined corresponding to the fiber being closest to all the other fibers populating the fascicle, thus providing a map of centroids for each individual.
Inter-subject fiber clustering – A diffeormorphic transformation based on the Symmetric Normalization (SyN)¹¹ approach provided by ANTs⁸ was computed to match each individual DW dataset and fascicle map to the template frame described earlier. A second hierarchical density-based HDBSCAN¹² step (normalization factor=1, minimum cluster size=1) was then performed at the level of the population to automatically regroup centroids stemming from different parrots into clusters and to only keep the most representative ones (present in 2 or 3 parrots).
Atlasing – The final step consisted in selecting clusters intersecting the WM regions manually delineated in the T2-weighted template to establish a first Pyrrhura molinae parrot brain white matter atlas based on the nomenclature of [Reiner et. al., 2004]¹³ : anterior commissure (AC), cerebellar fibers, lateral forebrain bundle (LFB), pallial-subpallial lamina (LPS), septopallio-mesencephalic tract (TSM), fronto-arcopallial tract (FA) and occipito-mesencephalic tract (OM).

Results

The dense 1 855 986±169 179 fiber connectomes were disentangled into 25 142±3 690 clusters in the right hemisphere, 27 767±2 816 clusters in the left hemisphere, 7 808±590 clusters in the cerebellum and 3 269±789 inter-hemispheric clusters as a result of the intra-subject fiber clustering step. The inter-subject fiber clustering step emphasized 4 989, 5 660, 1 684 and 494 representative white matter clusters respectively for the right, left hemispheres, cerebellum and inter-hemispheric region. Figures 3 and 4 represent the 4 lateralized bundles, the anterior commissure and the cerebellar fibers in 3D rendering.

Conclusion & Discussion

This study establishes the foundations of a structural connectivity atlas of the parrot brain in which 10 white matter bundles were characterized based on high-resolution high-field diffusion MRI data. Future work will focus on subdividing the FA and OM tracts. The size of the cohort will be increased in order to gain robustness and achieve a higher degree of precision. As such, the number of white matter bundles characterized by the atlas will be extended with delineation of the quintofrontal, thalamopallial, dorsal arcopallial and optic tracts.