Garden classification of femoral neck fracture using deep-learning algorithm.
Authors
Affiliations (5)
Affiliations (5)
- Department of Biomedical Engineering, Asan Medical Center, University of Ulsan College of Medicine, 88 Olympic-ro 43-gil, Songpa-gu, Seoul, Republic of Korea.
- Department of Orthopedic Surgery, Asan Medical Center, University of Ulsan College of Medicine, 88 Olympic-ro 43-gil, Songpa-gu, Seoul, Republic of Korea.
- Department of Orthopedic Surgery, Chungbuk National University Hospital, Chungbuk National University College of Medicine, Cheongju, Republic of Korea.
- Department of Biomedical Engineering, Asan Medical Center, University of Ulsan College of Medicine, 88 Olympic-ro 43-gil, Songpa-gu, Seoul, Republic of Korea. [email protected].
- Department of Orthopedic Surgery, Asan Medical Center, University of Ulsan College of Medicine, 88 Olympic-ro 43-gil, Songpa-gu, Seoul, Republic of Korea. [email protected].
Abstract
The Garden classification, based on X-ray interpretation and established over 50 years ago, remains the standard clinical classification system for femoral neck fractures (FNFs). Yet, this classification has a high interobserver variability of 70%. We sought to develop a deep-learning algorithm capable of accurately predicting FNF types, using only X-ray images, with performance comparable to that of computed tomography (CT). We retrospectively collected data from 1,588 patients who underwent X-ray and 3D-CT scans and were diagnosed with femoral neck fractures at Asan Medical Center. The input X-ray dataset consisted of paired X-ray images of the hip, with anteroposterior (AP) and lateral views. Using 3D-CT as the reference standard, patients were labeled as Garden types I (n = 378, 23.8%), II (n = 68, 4.3%), III (n = 477, 30.0%), and IV (n = 665, 41.9%). Our algorithm consisted of hip-joint detection followed by Garden classification, for which 12 different deep-learning architectures were evaluated. Algorithm performance was externally validated in 100 patients. Our algorithms showed a 90.6% overall accuracy and 88.6% Dice similarity coefficient, indicating excellent FNF type discernment. Our algorithms could serve as a valuable tool for diagnosing FNF based on X-ray data only, with accuracy comparable to that of CT.