Truace
Subscribe

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 cl…

By the Truvace desk2 min read1 sourcePeer-reviewedTRV-2026-1634· cite & verify
Deep Learning-Based Scapular Morphology Assessment Pipeline for Glenoid Segmentation and Landmark Localization
Gray328 by Henry Vandyke Carter. Public domain

In brief

Researchers built an open-source deep learning pipeline that analyzes shoulder CT scans. One model outlines the glenoid, the socket of the shoulder blade, and a second model locates five anatomical landmarks on the scapula. The system was tested on 793 CT images from 774 patients collected retrospectively.

Scapular shape informs how shoulder disorders are diagnosed and treated, and manual measurement is slow and varies between observers. A fast, reproducible tool could ease that burden. The reported results come from a laboratory study of retrospective scans, so benefit in routine care is not yet shown.

Main points

  1. The dataset contained 793 shoulder CT images from 774 patients with a mean age of 60.00 ± 18.62 years.
  2. A 2D U-Net handled glenoid segmentation on native-resolution axial slices with no down-sampling.
  3. A 3D nnUNet located five landmarks: trigonum spinae, angulus inferior, processus coracoideus, acromion, and angulus acromialis.
  4. Glenoid segmentation reached 96.59% precision and 97.30% recall alongside the 98.46% Dice score.

The gain

On retrospective shoulder CT data, the pipeline segmented the glenoid with high overlap and placed five scapular landmarks within 1.0 to 2.0 mm on average. Each case took about 22.55 seconds, versus slow manual annotation.

The rundown

Current practice relies on clinicians manually annotating three-dimensional CT scans. The authors describe this as time-consuming, labor-intensive, and prone to inter-observer variability, which motivated an automated approach.

The pipeline pairs two models. A 2D U-Net segments the glenoid at full resolution, and a 3D nnUNet predicts the positions of five scapular landmarks. Performance was judged by Dice, precision and recall for segmentation, and by Euclidean distance from ground truth for landmarks.

The authors designed it as an open-source framework that can be integrated into clinical workflows or research environments through a clinician-preferred user interface. The study is classed as a descriptive laboratory study, level of evidence 2.

Sources

  1. Peer-reviewedJournal of Imaging Informatics in Medicine2026-10-09

The debate