Deep Learning for Automatic Bone Segmentation of the Pelvis using MRI with T1 VIBE Dixon for FAI Patients

Purpose or Learning Objective: Femoroacetabular impingement (FAI) can cause hip pain and osteoarthritis (OA) in young patients of childbearing age. Diagnosis and surgical planning for FAI patients is based on static two-dimensional imaging. Magnetic resonance imaging (MRI)-based three-dimensional (3D) dynamic hip impingement simulation enables a patient-specific diagnosis of intra-and extra-articular FAI. But manual segmentation of MRI-based 3D models is time consuming; therefore, automatic segmentation was investigated. We aimed to explore (1) the difference between manual and automatic MRI-based 3D models of the hip joint, and (2) correlate hip range of motion (ROM) using deep learning.

Methods or Background: A controlled retrospective study approved by the institutional review board involving 30 symptomatic FAI patients (60 hips) and 19 asymptomatic volunteers (38 hips) was performed. All patients and volunteers underwent pelvic computed tomography (CT) scans and 3-T MRI of the hip (49 hips) including T1 volumetric interpolated breath-hold examination (VIBE) Dixon of the pelvis. Mean age of the patients was 27 ± 9 years, and 50% underwent surgical treatment.

Automatic segmentation of MRI-based 3D models using machine learning was compared with manual (semiautomatic) segmentation. For automatic segmentation, a convolutional neural network was used, and threefold cross validation was performed. Dice coefficient was calculated for 98 hips using manual and automatic MRI-based 3D models. Impingement-free ROM was compared between CT- and automatic MRI-based 3D models.

Results or Findings: (1) Dice coefficient of 30 FAI patients was 94% for the pelvis and 97% for the proximal femur. Dice coefficient of 19 volunteers was 93% for the pelvis and 96% for the proximal femur. (2) Correlation for impingement-free flexion (r = 0.93, p < 0.001) and extension (r = 0.99, p < 0.001) was excellent for FAI patients. Mean difference for flexion and internal rotation in 90 degrees of flexion was 3 ± 4 degrees and 3 ± 4 degrees, respectively.

Conclusion: Automatic MRI-based 3D models can replace manual segmentation for patients with FAI. Based on these results, we will use automatic MR-based 3D models in our future clinical routine. This allows a fast radiation-free and patient-specific preoperative diagnosis and surgical planning of open hip preservation surgery and hip arthroscopy for patients of childbearing age with FAI.

Publication History

Article published online:
02 June 2022