A Deep Learning Approach for Automatic 3D Segmentation of Hip Cartilage and Labrum from Direct Hip MR Arthrography

Abstract

Abstract The objective was to use convolutional neural networks (CNNs) for automatic segmentation of hip cartilage and labrum based on 3D MRI. In this retrospective single-center study, CNNs were used to develop a fully automated segmentation model for hip cartilage and labrum from MRI. Direct hip MR arthrographies (01/2020-10/2021) were selected from 100 symptomatic patients. Institutional routine protocol included a 3D T1 mapping sequence, which was used for manual segmentation of hip cartilage and labrum. 80 hips were used for training and the remaining 20 for testing. Model performance was assessed with six evaluation metrics including Dice similarity coefficient (DSC). In addition, model performance was tested on an external dataset (40 patients) with a 3D T2-weighted sequence from a different institution. Inter-rater agreement of manual segmentation served as benchmark for automatic segmentation performance. 100 patients were included (mean age 30 ± 10 years, 64% female patients). Mean DSC for cartilage was 0.92 ± 0.02 (95% confidence interval [CI] 0.92–0.93) and 0.83 ± 0.04 (0.81–0.85) for labrum and comparable (p = 0.232 and 0.297, respectively) to inter-rater agreement of manual segmentation: DSC cartilage 0.93 ± 0.04 (0.92–0.95); DSC labrum 0.82 ± 0.05 (0.80–0.85). When tested on the external dataset, the DSC was 0.89 ± 0.02 (0.88–0.90) and 0.71 ± 0.04 (0.69–0.73) for cartilage and labrum, respectively. The presented deep learning approach accurately segments hip cartilage and labrum from 3D MRI sequences and can potentially be used in clinical practice to provide rapid and accurate 3D MRI models.