A multi-task learning-based network for detecting three-dimensional osteometric landmarks of the shoulder joint

Background Mirror-based preoperative planning for total shoulder arthroplasty relies on accurate three-dimensional (3D) localization of bony landmarks on the contralateral, anatomically intact shoulder. Manual annotation is time-consuming and shows inter- and intra-observer variability. The aim of this study was to develop and evaluate a multi-task deep-learning network for automated 3D bony landmark detection on contralateral shoulder CT. Methods In this single-center retrospective study, 179 patients with intact contralateral shoulder CT scans were split at the patient level into training (n = 125), validation (n = 18), and test (n = 36) sets. We developed a dual-decoder 3D BiFormer-UNet for humerus and scapula segmentation and heatmap-based localization of 14 anatomical landmarks. A cross-task feature-fusion (CtFF) module transferred bone-structural guidance from the segmentation decoder to the landmark decoder. The network was trained end-to-end using a composite loss comprising hard-negative-mining heatmap regression, a geometric structure constraint, distance-aware similarity learning, and auxiliary segmentation loss. Landmark localization was evaluated using MRE and SDR at thresholds of 2.0, 2.5, 3.0, and 4.0 mm, while segmentation was assessed using the Dice coefficient. Per-case MRE was compared with four 3D baseline models using paired tests with Holm–Bonferroni correction. Results In the 36-case test set, the proposed model achieved an MRE of 2.1 ± 1.2 mm (95% CI, 1.7–2.5 mm) and SDRs of 72.0%, 81.8%, 90.3%, and 96.7% at thresholds of 2.0, 2.5, 3.0, and 4.0 mm, respectively. Its MRE was significantly lower than those of 3D U-Net, V-Net, 3D HRNet adaptation, and H3DE-Net (all Holm–Bonferroni-adjusted P < 0.05). On the test set, the auxiliary segmentation branch achieved mean Dice coefficients of 0.83 ± 0.05 for the humerus and 0.78 ± 0.07 for the scapula. Internally defined center landmarks, particularly the humeral head rotation center, anatomical neck center, and proximal and distal medullary canal centers, showed larger localization errors (2.3–2.6 mm) than most surface landmarks. Conclusion The proposed multi-task framework demonstrated promising 3D landmark localization performance on intact contralateral shoulder CT by integrating global anatomical context with bone-structural guidance. It may serve as an upstream localization component for mirror-based shoulder arthroplasty planning, although external validation is required before clinical use.