Data Preparation

Experiments were conducted using two datasets collated from publicly available T1-weighted 1.5-Tesla (T) MRI volumes, which we will refer to as ACPC-MRI-1 and ACPC-MRI-2. Each MRI scan is 256 × 256 × 150 voxels with a cubic millimeter resolution per voxel. The first dataset, ACPC-MRI-1, is comprised of 908 MRI volumes across 687 subjects from the following publicly available data: Open Access Series of Imaging Studies (OASIS) (Marcus et al., 2007) and Mindboggle-101 (Klein et al., 2017). The OASIS dataset is comprised of scans for subjects aged 18 to 96 years, with a subset of subjects aged 60 years and older who were diagnosed with the onset of dementia to moderate Alzheimer’s disease (Marcus et al., 2007). Two to four additional scans are included for the subset of 100 subjects with dementia and 20 healthy subjects (Marcus et al., 2007). The MindBoggle dataset includes T1-weighted MRI volumes from 101 healthy subjects (Klein and Tourville, 2012). The second dataset, ACPC-MRI-2, contains 220 volumes across 158 subjects from the OASIS-3 dataset (LaMontagne et al., 2019). The AC and PC anatomical landmarks were manually localized for a total of 1,128 MRI annotated volumes across ACPC-MRI-1 and ACPC-MRI-2 datasets.

An annotation protocol was established to label the AC and PC anatomical landmarks for each T1-weighted MRI volume. To streamline the annotation process, each volume was aligned to a common MRI template using a rigid registration approach, which is publicly available in the Advanced Normalization Tools (ANTs) software toolkit (Avants et al., 2011). Each normalized MRI volume was imported into the publicly available 3D Slicer Tool¹ and viewed with a predefined AC and PC template overlaid on the volume (Kikinis et al., 2014; Fedorov et al., 2016). Using the 3D Slicer Tool, each annotator manually fine-tuned reference point locations to the posterior and anterior edges of the AC and PC points, respectively, and saved the 3D coordinates for each point to a fiducial markup file per MRI volume (see Figure 1). Without fine tuning the locations, the average 3D Euclidean distances between initial locations provided by our template and our ground truth locations of AC and PC points were 6.91 ± 5.50 mm and 6.44 ± 4.89 mm (N = 1128), respectively. For the ACPC-Dataset-1, a team of five annotators ranging from novice to an experienced neurosurgeon labeled the AC and PC points for a total of 908 T1-weighted MRI volumes across 687 subjects. Multiple annotators labeled a subset of the MRI scans (N = 274), and the resulting 3D coordinates for each landmark were averaged and used as ground truth. For quality control, when multiple labels were available for a given point, the 3D Euclidean distance between points was calculated The average AC and PC distance, across annotators, was 1.14 ± 0.57 mm (2.92 mm max error) and 1.04 ± 0.53 mm (2.94 mm max error), respectively. For distances greater than 2.0 mm, the labels were visually evaluated and adjusted by an expert annotator. For MRI volumes where only a single annotation was available and the annotator was considered a novice, the label was visually inspected and adjusted as needed. Approximately 10% of the labeled points were re-adjusted manually within the Slicer Tool by an expert-level annotator. For ACPC-Dataset-2, an expert-level annotator labeled the AC and PC landmarks of all the volumes, using the same annotation protocol described for ACPC-Dataset-1.

Example of an MRI volume annotated using 3D Slicer software. The volume is displayed in Right-Anterior-Superior (RAS) coordinate space. The annotator manually adjusts fiducial markers overlaid on the AC and PC points from linked 2-D views. For this subject, the AC and PC points are both visible in the same sagittal slice, whereas only the AC point is visible in the axial and coronal slices shown here. Depending on the orientation of the brain within the MRI volume, the AC and PC points are sometimes both visible within a single sagittal and an axial slice. The 3D coordinates for each point are displayed in millimeters in the Cartesian X-Y-Z space. The labels and corresponding 3D coordinates are saved to a fiducial comma-separated values (.fcsv) file.

The DeepNavNet models described in this paper regress from an input MRI volume to an output heatmap volume with spheres centered around 3D coordinates of the AC and PC landmarks within each volume. Therefore, training data preparation included an additional step of creating heatmap volumes with Gaussian spheres, with sigma set to 3, centered around 3D coordinates of the AC and PC landmarks within each MRI volume. Each heatmap was created with the same dimensions as their corresponding MRI volume, and the voxel intensity values were normalized to a maximum value of 1. The ACPC-MRI-1 dataset was used for training, validation and testing, whereas the ACPC-MRI-2 data was used for additional testing only. As such, heatmap creation was only necessary for the ACPC-MRI-1 dataset.