¹Department of Orthopaedic Surgery, Inselspital, University of Bern, Bern, Schweiz
²University Institute of Diagnostic, Interventional and Paediatric Radiology, Inselspital, University of Bern, Bern, Schweiz
³Radiology Department, Balgrist University Hospital, Zürich, Schweiz

Text

Objective and research question: Hip deformities alter joint biomechanics, leading to pain and early degeneration, particularly in younger patients. Joint-preserving surgery relies on accurate diagnosis and correction. Conventional imaging, including radiographs, CT, and MRI, struggles to capture the three-dimensional (3D) structure of the hip. Modern contrast-enhanced 3D MRI morphologic sequences, combined with artificial intelligence (AI) algorithms, now enable automatic segmentation of the hip labrum and cartilage. Ensuring the generalizability of these methods also for non- contrast MRI protocols is crucial for their potential application in clinical settings.The objective of this study is to train and validate a deep learning algorithm (3D nnU-Net) for fully automated segmentation and generation of MRI-based 3D models of the hip labrum and cartilage using non-contrast hip MRI.

Material and methods: Manual segmentation of the hip labrum and cartilage was performed on non-contrast 3D double-echo steady-state (T2 DESS) sequences (voxel resolution: 0.3 × 0.3 × 1 mm) using Amira 6.1 software (FEI, Hillsboro, Oregon, USA) in the axial-oblique plane to establish ground truth. Data were obtained from 45 asymptomatic volunteers (mean age 30 ± 6 years, 56% female). The dataset was split into 25 cases for training and 20 cases for testing of automatic segmentation on unseen data, with model performance evaluated using 5-fold cross-validation. Model evaluation metrics included the dice similarity coefficient (DSC), absolute and relative surface area differences between manual and automatic segmentation. Data normality was tested with the Kolmogorov-Smirnov test, and paired t-tests were used for statistical comparisons.

Results: Mean DSC was 0.94 ± 0.01 (95% CI: 0.94–0.95) for cartilage and 0.81 ± 0.04 (95% CI: 0.80–0.83) for labrum. Labrum surface area differed significantly (p = 0.004) with an absolute difference of 76.9 ± 63.8 mm² (95% CI: 47.0–106.8) and relative difference of 8.8 ± 5.1% (95% CI: 6.4–11.2).For cartilage, the area showed no significant difference (p = 0.277), with an absolute difference of 65.9 ± 45.9 mm² (95% CI: 44.4–87.4) and a relative difference of 3.1 ± 2.0% (95% CI: 2.2–4.1).

Discussion and conclusion: Excellent performance was achieved in the automatic segmentation of the hip labrum and cartilage, with DSC values of 0.81 for the labrum and 0.94 for cartilage, comparable to published results for contrast-enhanced MRI-based segmentation (DSC 0.83 and 0.92, respectively). The surface area of the labrum differed minimally between manual and automatic methods (p=0.004), representing a mean relative difference of 8.8%.