AFIDs: a standardized framework for evaluating anatomical correspondence between primate brains
Borna Mahmoudian, Nikoloz Sirmpilatze, Mohamad Abbass, Sarah Allarakhia, Greydon Gilmore, Geetika Gupta, Katja Heuer, P. Christiaan Klink, Roberto Toro, & Jonathan Lau
‪Establishing accurate spatial correspondences across brains is a fundamental step in neuroscience research. Current methods use sophisticated algorithms to perform registrations between brains, providing a mapping between modalities, subjects, template spaces, and even different species. However, the accuracy of these mappings varies widely depending on factors such as the image quality, the registration algorithm, and its parameters (Klein et al., 2009). To quantify registration accuracy, a point-based set of 32 anatomical landmarks, termed anatomical fiducials (AFIDs), has recently been described and validated for human magnetic resonance imaging (MRI) datasets (Lau et al., 2019).
‪In this study, we sought to bridge the gap between human and animal models in neuroscience research by extending the AFIDs framework to a widely used non-human primate model, the macaques.
Inspired by our work with human data, we first determined whether the same set of 32 AFIDs could be placed in commonly used macaque MRI datasets and developed an online protocol for training new users. Facilitated by the BrainWeb collaborative space, we recruited 8 raters and set out to validate the protocol on a set of publicly available macaque neuroimaging datasets and brain templates: D99 , INIA19, MNI macaque, NMTv1.3, and Yerkes19 (more information on the templates here) as well as 8 single subjects (available as part of the PRIME-DE collection, Milham et al. 2018). Every rater placed a set of 32 AFIDs per brain using the open source software 3D Slicer (Fedorov et al 2012). To control for rater reliability, raters placed the AFIDs on each scan twice, in two independent sessions.
Below you can see how an example AFID - superior interpeduncular fossa (SIPF) - is being placed in 3D Slicer, with the help of a target image from the protocol.
We followed the same protocol for validation as used in the original human study, quantifying the point placement accuracy as AFLE (AFID localization error). The utility of the AFIDs protocol was demonstrated by using these placed AFIDs to quantify template-to-template registration error, termed AFRE (AFID registration error), using pre-computed warps available through the RheMAP project (Sirmpilatze and Klink, 2020).
Using our extensive illustrated protocol for the precise placement of 32 anatomical fiducials in macaque brains, users were able to place the AFIDs with voxel-level accuracy, i.e. AFLE = 0.42 +/- 0.25 mm across all templates and AFIDs.
The AFRE for template-to-template linear (affine) registration was 1.03 +/- 0.49 across all template pairs. The addition of a non-linear registration step (ANTs Symmetric Diffeomorphic registration, Avants et al., 2008) reduced the AFRE to 0.68 +/- 0.43.
‪The AFIDs protocol has been successfully extended to macaques, allowing us to precisely place the same original set of point landmarks first described in humans. This open protocol holds value for a broad number of applications, including creating correspondences across macaque brains, refined stereotactic targeting of deep brain structures, and teaching neuroanatomy. The point-to-point distances between AFIDs may also serve as a means for quantifying cross-species anatomical similarities and differences and is the subject of ongoing work. The raw data and code are available on GitHub.