Automated landmark-based symmetric and standard alignment of skull base structures on CT

Abstract

Introduction: Symmetry and standard alignment are crucial in both clinical interpretation and research on head CT studies. Registration to a standard template is the traditional method for alignment, yet registration does not guarantee precise alignment of any given structure. This study introduces a method for aligning skull base structures while still achieving a standard anterior commissure-posterior commissure (AC-PC)-like orientation on head CT studies using landmarks, specifically the cochleas and nasal bridge.

Methods: A retrospective study was conducted using head CTs from various General Electric scanners. Landmarks were manually annotated, and a 3D U-Net was trained for landmark identification. Landmark-based alignment was then performed on a test dataset and assessed in two different ways: whole head and skull base alignment. Whole head alignment was assessed quantitatively by expert review. Skull base alignment was then assessed at the cochleas, comparing their alignment between this landmark-based technique and registration to a template.

Results: This landmark-based technique significantly improved whole head and skull base alignment of head CT studies. Whole head alignment reduced average deviations of 5, 11, and 4° in the axial, sagittal, and coronal planes to 1, 5, and 2° respectively. Meanwhile, skull base alignment assessed via the cochlea was also improved relative to traditional registration. For the landmark technique, the cochleas were deviated from perfect by a mean of 0.552 and 0.511 mm along the y and z axes compared to 2.110 and 2.506 mm with registration.

Conclusion: This study demonstrates a simple landmark-based technique for aligning the cochleas on head CT studies while approximating whole head AC-PC orientation, which has applications in both clinical and research settings, particularly for studies focused on the skull base.

Keywords: AC-PC line; Alignment; Head CT; Landmark; Registration; U-Net.

Fig. 1

Schematic representation of 3D U-Net utilized for landmark identification. Specifically, 128 × 128 × 80 initial 3D image resolution with 5 channels scaling from 16 to 256 with strides of 2 and 2 residual units per layer (R) are depicted.

Fig. 2

Landmark alignment technique. CT images are shown in a bone window setting to better demonstrate skull base structures. Initial CT (A) is shown in a plane including all three landmarks (black circles on nasal bridge, right cochlea, left cochlea). First, axial rotation to align the cochleas on the y axis (B), then sagittal rotation to align the nasal bridge and right cochlea on the z axis (C), then coronal rotation to align the cochleas on the z axis (D).

Fig. 3

Manual whole head alignment validation measurement technique. CT images are shown in a soft tissue window setting to better demonstrate measurement landmarks on this unaligned head CT. (A) Axial alignment (degrees angulation from the dashed line between the anterior and posterior attachments of the falx) is off by 4.64° in the clockwise direction for a value of +5 (A). (B) Sagittal alignment (degrees angulation from the dashed AC-PC line) is off by 25.3° in the clockwise direction for a value of +25. (C) Coronal alignment (degrees angulation from the dashed line drawn between the center of the sella inferiorly and superior sagittal sinus superiorly) is off by 5.26° in the counter-clockwise direction for a value of −5 (C).

Fig. 4

Assessment of Landmark versus Registration Skull Base Alignment. This box plot demonstrates the millimeters (mm) of deviation from perfect symmetric alignment of the cochleas along the y and z axes by original, not absolute values. (Reg = Registration Alignment, LM = Landmark Alignment).

Fig. 5

Assessment of Landmark Whole Head Alignment Accuracy. This box plot demonstrates the whole head absolute degrees of malalignment in each plane on the original unaligned acquisition versus the landmark aligned images. (UA = unaligned, A = landmark aligned, Ax = axial, Sag = sagittal, Cor = coronal).