atlasBREX: Automated template-derived brain extraction in animal MRI

www.nature.com/scientificreports OPEN Received: 12 February 2019 Accepted: 3 June 2019 Published: xx xx xxxx atlasBREX: Automated templatederived brain extraction in animal MRI Johannes Lohmeier Hideyuki Okano2,3 1 , Takaaki Kaneko2, Bernd Hamm1, Marcus R. Makowski1 & We proposed a generic template-derived approach for (semi-) automated brain extraction in animal MRI studies and evaluated our implementation with different animal models (macaque, marmoset, rodent) and MRI protocols (T1, T2). While conventional MR-neuroimaging studies perform brain extraction as an initial step priming subsequent image-registration from subject to template, our proposed approach propagates an anatomical template to (whole-head) individual subjects in reverse order, which is challenging due to the surrounding extracranial tissue, greater differences in contrast pattern and larger areas with field inhomogeneity. As a novel approach, the herein introduced brain extraction algorithm derives whole-brain segmentation using rigid and non-rigid deformation based on unbiased anatomical atlas building with a priori estimates from study-cohort and an initial approximate brain extraction. We evaluated our proposed method in comparison to several other technical approaches including “Marker based watershed scalper”, “Brain-Extraction-Tool”, “3dSkullStrip”, “Primatologist-Toolbox”, “Rapid Automatic Tissue Segmentation” and “Robust automatic rodent brain extraction using 3D pulsecoupled neural networks” with manual skull-stripping as reference standard. ABX demonstrated best performance with accurate (≥92%) and consistent results throughout datasets and across species, age and MRI protocols. ABX was made available to the public with documentation, templates and sample material (https://www.github.com/jlohmeier/atlasBREX). Brain extraction, also referred as skull-stripping or whole-brain segmentation, describes the process of extracting the brain from the surrounding extracranial tissue. In MRI studies, it is common that this procedure is implemented at an early stage, as it plays an important role for further processing, such as spatial normalisation, surface reconstruction and structural analysis1. Manual delineation is considered technical standard, but it demands high time investment, experience and neuroanatomical knowledge. Hence, there is need for automated technical alternatives, which are less operator-dependent. Several (semi-) automated (hybrid) approaches were developed for human neuroimaging thus far1, but present a high degree of specialisation due to a priori estimates. Therefore, established technical approaches for human neuroimaging are often not compatible with animal MRI and the adaption can be demanding due to interspecies differences in brain size, shape and tissue contrast as well as differences in MRI scanners, magnetic field strengths, radiofrequency coils and MRI protocols. A common challenge in skull-stripping animal MR-neuroimaging is the presence of more severe field inhomogeneity, which is attributable to non-uniformity in radiofrequency coils. As illustrated in Fig. 1, both pattern of occurrence (see heterogeneous gradient) as well as the severity of distortion are subject to variation, which affects the performance of processing algorithms that infer information from image intensity. Further challenges arise from low-resolution images (see Fig. 1a,c), such as in functional and diffusion MRI studies. In recent years, neuroimaging studies with animal models, such as macaque2, marmoset3,4 and rodent5, gained in significance due to their contributions to understanding the central nervous system6,7. To date, however, there are only few technical approaches available that can be applied to animal MRI, such as Marker based watershed scalper (MBWSS)8, Brain-Extraction-Tool (BET)9, 3dSkullStrip (3DSS) as part of the Analysis of Functional NeuroImages (AFNI) package10, Primatologist-Toolbox (PRIMA)11, Rapid Automatic Tissue Segmentation (RATS)12 and Robust automatic rodent brain extraction using 3D pulse-coupled neural networks (PCNN3D)13. 1 Charité Universitätsmedizin Berlin, Radiology, Berlin, Germany. 2Center for Brain Science Institute, RIKEN, Marmoset Neural Architecture, Wako-shi, Saitama, Japan. 3Department of Physiology, Keio University School of Medicine, Tokyo, Japan. Correspondence and requests for materials should be addressed to J.L. (email: johannes. ) or H.O. (email: ) Scientific Reports | (2019) 9:12219 | https://doi.org/10.1038/s41598-019-48489-3 1 www.nature.com/scientificreports/ www.nature.com/scientificreports Figure 1. Challenges in skull-stripping animal MR-neuroimaging. Illustration of common difficulties when skull-stripping animal MRI (top-left: marmoset [9.4 T Biospec, Bruker, Germany]; top-right: rat [9.4 T Biospec, Bruker, Germany]; bottom-left: rhesus macaque [3 T Prisma, Siemens, Germany]; bottom-right: marmoset [3 T Prisma, Siemens, Germany]), such as low image resolution (a,c), strong field inhomogeneity (a–d) or greater field-of-view (b,d) with larger areas of non-brain tissue. However, it is common that results are below standard and require further manual intervention. Hence, there is demand for more robust brain extraction algorithms in animal MRI. Therefore, we proposed a generic template-derived approach for animal neuroimaging: We present atlasBREX (ABX), a semi-automated processing pipeline that propagates skull-stripping of an anatomical template built from the study-cohort after rigid and non-rigid deformation to each individual subject (see Fig. 2). First, in a practical and unbiased manner, an anatomical study-specific template is computed from all individual subjects (see Fig. 2, step 1) using an iterative hierarchical group-wise registration framework, Atlas Building by Self-Organized Registration and Bundling (ABSORB)14. Next, the study-specific anatomical template is subject to manual (hybrid) skull-stripping (see Fig. 2, step 2). In the following steps, rigid and non-rigid deformation fields are computed from template- to target-space (see Fig. 2, step 5 and 7), which are applied to the template mask in order to compute a subject-specific brain mask (see Fig. 2, step 6 and 8). While spatial transformation is conventionally performed after image preprocessing (including brain extraction) with the objective to align images within or across individuals, our proposed approach is based on the backpropagation of an anatomical template to the (whole-head) individual subject, which is challenging due to the surrounding extracranial tissue, greater differences in contrast pattern and larger areas with strong field inhomogeneity (see Fig. 1). Therefore, as a novel approach, the herein introduced brain extraction algorithm derives whole-brain segmentation using rigid and non-rigid deformation based on unbiased anatomical atlas building with a priori estimates from the study-cohort and an initial approximate brain extraction. The (...truncated)