Deep learning algorithms enable MRI-based scapular morphology analysis with values comparable to CT-based assessments

Figures

• 

  Fig. 1. Anatomical landmarks of the scapula: (1) Anterior- (2) Posterior-. (3) Inferior- points on the lower glenoid rim. (4) Glenoid centre, origin of the scapular coordinates system (circle fit from points 1,2, and 3). (5) Inferior scapular angle. (6) Centre of the trigonum spinae scapulae. (7) Superior glenoid tubercule. (8) anterior lateral corner of the acromion. The glenoid plane is defined using points 1,2, and 3. The coronal scapular plane (CSP) is defined using points 4, 5, and 6. The transverse scapular axis runs through points 4) and 6), corresponding to the z-axis of the scapular coordinate system.

• 

  Fig. 2. Overview of the pipeline for automatic morphology analysis. (a) From diagnostic magnetic resonance image (MRI) data, automatic segmentation is performed on multiple image orientations for the same patient. This information is fused to create isotropic, high-resolution 3D masks of the scapula. (b) From these high-resolution 3D masks, landmarks and axes are automatically detected. The landmark coordinates and axes directions are used for morphological analysis.

• 

  Fig. 3. Methods for multi-planar, high-resolution shape reconstruction. (a) Fusion of separately segmented masks from coronal sagittal and axial magnetic resonance imaging (MRI). (b) Example of a scapula segmentation in different sagittal slices, indicated with different colours showing different regions of 1) High consistency; where the mask border remains consistent over the adjacent image slices, 2) medium consistency; where the mask border differs between adjacent image slices, 3) low consistency, where the mask border is highly inconsistent over adjacent image slices.

• 

  Fig. 4. Visualization of intermediate steps in generating the high-resolution scapula model from multiple MRI planes. The blue crosshair indicates the same glenoid position across all views. (a) In-plane distance map of the scapula segmentation on coronal, transverse and sagittal MRI planes, with distances ranging from − 15 and 15. Bright areas represent high segmentation border distances within the scapula mask, while dark areas indicate large distances outside the mask. (b) Mask border consistency along the out-of-plane direction in the coronal, transverse and sagittal MRI planes, ranging from 0 and 1. White areas denote regions with low border consistency (assigned lower weighting for plane-combination), while dark areas indicate regions with higher consistency and weighting.

• 

  Fig. 5. 3D models from a patient with RMSE of 1.54. (a) 3D model of the HR-MRI-mask, (b) 3D model from CT, (c) color-coded map of absolute error surface distance between HR-MRI and CT 3D-models (blue = 0 mm, to yellow ≥ 5 mm).