Murphy MP, Killen CJ, Winfrey SR, Schmitt DR, Hopkinson WJ, Wu K, Brown NM. Artificial Intelligence Autonomously Measures Cup Orientation, Corrects for Pelvis Orientation, and Identifies Retroversion From Antero-Posterior Pelvis Radiographs.
J Arthroplasty 2023;
38:S319-S323. [PMID:
36893991 DOI:
10.1016/j.arth.2023.02.076]
[Citation(s) in RCA: 0] [Impact Index Per Article: 0] [Reference Citation Analysis] [Abstract] [Key Words] [MESH Headings] [Track Full Text] [Journal Information] [Submit a Manuscript] [Subscribe] [Scholar Register] [Received: 11/30/2022] [Revised: 02/24/2023] [Accepted: 02/26/2023] [Indexed: 03/11/2023] Open
Abstract
BACKGROUND
Measuring cup orientation is time consuming and inaccurate, but orientation influences the risk of impingement and dislocation following total hip arthroplasty (THA). This study designed an artificial intelligence (AI) program to autonomously determine cup orientation, correct for pelvis orientation, and identify cup retroversion from an antero-posterior pelvic radiographs.
METHODS
There were 2,945 patients between 2012 and 2019 identified to have 504 computed tomographic (CT) scans of their THA. A 3-dimensional (3D) reconstruction was performed on all CTs, where cup orientation was measured relative to the anterior pelvic plane. Patients were randomly allocated to training (4,000 x-rays), validation (511 x-rays), and testing (690 x-rays) groups. Data augmentation was applied to the training set (n = 4,000,000) to increase model robustness. Statistical analyses were performed only on the test group in their accuracy with CT measurements.
RESULTS
AI predictions averaged 0.22 ± 0.03 seconds to run on a given radiograph. Pearson correlation coefficient was 0.976 and 0.984 for AI measurements with CT, while hand measurements were 0.650 and 0.687 for anteversion and inclination, respectively. The AI measurements more closely represented CT scans when compared to hand measurements (P < .001). Measurements averaged 0.04 ± 2.21°, 0.14 ± 1.66°, -0.31 ± 8.35°, and 6.48° ± 7.43° from CT measurements for AI anteversion, AI inclination, hand anteversion, and hand inclination, respectively. AI predictions identified 17 radiographs as retroverted with 100.0% accuracy (total retroverted, n = 45).
CONCLUSION
The AI algorithms may correct for pelvis orientation when measuring cup orientation on radiographs, outperform hand measurements, and may be implemented in a timely fashion. This is the first method to identify a retroverted cup from a single AP radiograph.
Collapse