Deep Learning-Based Scapular Morphology Assessment Pipeline for Glenoid Segmentation and Landmark Localization

Scapular morphology plays a critical role in the diagnosis and treatment of shoulder disorders. However, current assessment methods primarily rely on manual annotation of three-dimensional computed tomography (3D CT) scans by clinicians, which are time-consuming, labor-intensive, and prone to inter-observer variability. The objective of this study is to develop and validate a deep learning-based open-source pipeline for automated scapular morphology assessment, aiming to improve accuracy, reproducibility, and clinical efficiency. Descriptive Laboratory Study; Level of evidence, 2. We retrospectively collected 793 CT images of the shoulder from 774 patients (mean age, 60.00 ± 18.62 years). A 2D U-Net model performed high-resolution glenoid segmentation (native-resolution axial CT slices, no down-sampling), and a 3D nnUNet model was employed to localize five anatomical scapular landmarks: trigonum spinae (TS), angulus inferior (AI), processus coracoideus (PC), acromion (AC), and angulus acromialis (AA). Model performance was evaluated using the Dice coefficient and precision-recall metrics for glenoid segmentation and the Euclidean distance between the predicted and ground truth landmark position. This pipeline was designed as an automated open-source framework that can be readily integrated into clinical workflows or research environments through clinician-preferred user interface. The glenoid segmentation model achieved a 3D Dice coefficient of 98.46%, with a precision of 96.59% and a recall of 97.30%. The mean Euclidean localization error for the five scapular landmarks ranged from 1.0 to 2.0 mm. The entire assessment process took average 22.55 s per case. The proposed deep learning pipeline enables accurate, efficient, and reproducible assessment of scapular morphology. The proposed approach facilitates reliable and efficient diagnosis, offering a promising tool for clinical and research applications.