Classification of Knee Osteoarthritis Based on Transfer Learning Model and Magnetic Resonance Images
Xin Wang, Shuang Liu, Changcai Zhou
Abstract
Magnetic resonance (MR) images can display the bone and cartilage tissue of the knee joint and play an important role in the early classification and diagnosis of knee osteoarthritis (OA). Deep learning has great advantages in feature extraction and has been widely used for classification and detection. Therefore, in this work, the transfer learning-based model is proposed to discriminate the presence or absence of knee OA. The pre-trained model InceptionResNetV2 is used to extract deep features of knee OA from MR images belonging to the OAI-ZIB dataset. The two different optimization algorithms (SGD and RMSprop) are used for training and testing the knee MR image dataset. The performance of the transfer learning model is evaluated using the accuracy, sensitivity, precision, specificity, F1-Score, and Matthews Correlation Coefficient (MCC) evaluation metrics. From the experimental results, the proposed transfer learning-based model using the RMSprop optimizer gets better classification results in the patient-level and image-level dataset partitioning methods, achieving 88.5% and 96.1% accuracy, 92.7% and 97.7% sensitivity, 86% and 94.4% precision, 86.6% and 94.6% specificity, 88.3% and 96% F1-Score, 77.2% and 92.2% MCC. The proposed method outperforms the comparison algorithms, indicating that it is effective in classifying knee OA from MR images.