A Novel 2.5-Dimensional Deep Learning Model for "Bone-on-Bone" Detection on Magnetic Resonance Imaging in Medial Unicompartmental Knee Arthroplasty Candidates.
Authors
Affiliations (4)
Affiliations (4)
- Department of Orthopaedic Surgery, China-Japan Friendship Hospital, Beijing 100029; Department of Orthopedics, Shenzhen Second People's Hospital/First Affiliated Hospital of Shenzhen University Health Science Center, Shenzhen, Guangdong 518035.
- Department of Orthopaedic Surgery, China-Japan Friendship Hospital, Beijing 100029; Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, 100730.
- Department of Orthopaedic Surgery, China-Japan Friendship Hospital, Beijing 100029.
- Department of Orthopaedic Surgery, China-Japan Friendship Hospital, Beijing 100029. Electronic address: [email protected].
Abstract
Accurate identification of "bone-on-bone" (BoB) osteoarthritis is critical for patient selection for medial unicompartmental knee arthroplasty (mUKA); however, the assessment remains subjective. We aimed to develop and validate a deep learning model using multi-sequence knee magnetic resonance imaging (MRI) for automated detection of isolated medial compartment BoB osteoarthritis in UKA candidates, thereby reducing avoidable indication-related failures. We retrospectively collected preoperative knee MRI data from 191 patients (64 patients who had BoB osteoarthritis, 62 patients who had partial cartilage loss, and 65 patients who had normal cartilage) who underwent mUKA or knee arthroscopy between January 2022 and January 2025. We developed an automated pipeline comprising a cartilage-segmentation model followed by a multi-sequence MRI-based 2.5-dimensional (2.5D) deep-learning model integrating adjacent slices and multiplanar views. The dataset was divided into training (70%) and test (30%) sets. Model performance was evaluated on the test set and compared with models using different feature-construction strategies. Orthopaedic surgeons who had varying experience independently evaluated classification performance. The 2.5D deep-learning model demonstrated superior diagnostic performance on the test set, with a micro area under the receiver operating characteristic curve (micro-AUC) of 0.893, compared with models based solely on clinical data (0.810) or radiomic features (0.851). Combining the deep learning model output with radiomic and clinical features further improved performance (micro-AUC = 0.918). The fused model's classification accuracy was comparable to that of senior orthopaedic surgeons and significantly higher than that of junior surgeons (P = 0.027 and 0.010, respectively). It also reduced interobserver variability and produced results substantially faster than manual interpretation. A multi-sequence 2.5D deep learning model reliably detects medial compartment BoB osteoarthritis on MRI. Its objective, reproducible output may enhance diagnostic accuracy and consistency, providing a standardized reference for cartilage assessment and supporting preoperative clinical decision-making for UKA.