FastSurfer-CC: Head Pose Normalization for Consistent Corpus Callosum & Fornix Segmentation

1868 

Abstract Submission 

Clemens Pollak¹, Andreas Girodi¹, Kersten Diers¹, Santiago Estrada^1,2, David Kügler¹, Martin Reuter^1,3,4

¹AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany, ²Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany, ³A.A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, MA, ⁴Department of Radiology, Harvard Medical School, Boston, MA

Clemens Pollak  
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)
Bonn, Germany

Andreas Girodi  
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)
Bonn, Germany

Kersten Diers  
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)
Bonn, Germany

Santiago Estrada  
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)|Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE)
Bonn, Germany|Bonn, Germany

David Kügler  
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)
Bonn, Germany

Martin Reuter  
AI in Medical Imaging, German Center for Neurodegenerative Diseases (DZNE)|A.A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital|Department of Radiology, Harvard Medical School
Bonn, Germany|Boston, MA|Boston, MA

Corpus callosum (CC), fornix (FN), anterior- and posterior commissures (AC, PC) are white matter bundles, central to the communication between hemispheres, memory recall tasks and olfaction. They also play important roles in various diseases, like epilepsy [1], autism [2], and schizophrenia [3]. These bundles have discrete borders to surrounding tissues in the sagittal plane, but not to the left and right hemisphere. Therefore, segmentation is often performed only in midsagittal slices, which are determined a priori. Existing literature frequently treats the head pose normalization and segmentation tasks separately [4,5]. Finding the midsaggital plane (midplane) is challenging, since the border between hemispheres can be curved. Nonetheless, approximating the midplane accurately is required for consistent segmentation, since errors in its orientation can propagate to inaccurate, or even biased measurements w.r.t. CC and FN size and shape. For this reason, we develop a two-step pipeline that first finds the best-suited midplane for CC and FN segmentation, and second, segments CC and FN. Additionally, the second step localizes AC and PC, resulting in the AC-PC line (bicommissural line). We combine this localization with the midplane to standardize the brain orientation on all three axes, which can aid further analysis and quality control.

We develop and evaluate our methods on a dataset of 280 T1-weighted brain MRIs, from 7T, 3T, and 1.5T scanners of the HCP, ADNI, OASIS, ABIDE-II, MIRIAD, la5c, Rhineland-, and in-house studies. To find the midplane, we consider four methods: a) Freesurfer's mri_cc, b) finding the left-right symmetry axis by mid-space registration of Freesurfer's mri_robust_register to a left-right flipped version of the same volume c) registering the center points of segmentation labels to those of a template (centroid_template) d) same as c), but using the flipped segmentation as a template and mapping into mid-space, similar to b) (centroid_flipped). Two blinded experts compared 40 midplanes for each method pair and marked which plane best separates CC, FN, AC, and PC into equal parts. We selected 20 midplane pairs according to the highest difference between them and 20 pairs randomly. The plane-difference was estimated by the volume between planes within the brain, approximated by a left-right oriented cylinder centered at the RAS coordinate origin. As reference standard for the segmentation task, two experts manually segmented CC, FN, AC, and PC in an area 2.5mm around the midplanes derived by mri_cc. The resulting 158 labeled volumes were then split into training (97), validation (31), and test set (30). On these labels, we train a 3D Swin Unet-transformer (S-UNETR) [6] for joint segmentation of CC and FN as well as prediction of heatmaps for the location of AC and PC [8].

We compare the four candidate methods (Fig. 1) and conclude that method centroid_template outperforms mri_robust_register and mri_cc outperforms centroid_flipped. No clear ranking can be found between mri_cc and centroid_template. Since the centroid registration is very fast (2 sec. avg.) compared to mri_cc's registration (52 sec. avg.), we choose to include it in our final pipeline. For the joint segmentation and localization sub-task, we find that S-UNETR outperforms other benchmarked architectures and mri_cc (Fig. 2).

We propose a two-step pipeline for fast and accurate CC and FN segmentation, that includes finding the midplane and performing a joint segmentation and localization. Through localization of AC and PC it also provides an anatomically motivated brain orientation, which matches other standardization techniques, like the rigid registration to the Talairach space. In the future, we aim to predict the midplane with the S-UNETR, further improving upon the existing methods, and to distribute our method as part of the FastSurfer neuroimaging software suite [9].

Image Registration and Computational Anatomy

Methods Development ¹

Segmentation and Parcellation ²

Computational Neuroscience

Data analysis

Machine Learning

MRI

Segmentation

Other - Structural MRI; Corpus Callosum; Fornix; Commisures; Toolbox; Tool; Registration; Standardization; AI