AUTOMATED BML SEGMENTATION USING UNET FOR TOTAL KNEE REPLACEMENT PREDICTION

Knee osteoarthritis (OA) is a leading cause of disability worldwide and often progresses to total knee replacement (TKR). Bone marrow lesions (BMLs), visible as hyperintense regions on MRI, are strong imaging predictors of TKR progression. Although manual BML segmentation is considered the gold standard, it is highly time-consuming and impractical for large-scale clinical use. Deep learning-based segmentation offers a scalable alternative, but it remains unclear whether automatically extracted BML features provide the same predictive value as expert annotations.

This study evaluates the clinical feasibility of using a UNet-based segmentation model for automated BML quantification and investigates whether UNet-derived volumetric features can effectively replace manual annotations for TKR prediction.

Two datasets from the Osteoarthritis Initiative (OAI) were used: Dataset 1 included 300 patients with manual annotations for training, validation, and testing, while Dataset 2 included 1,393 patients for external testing. CLAHE preprocessing was applied to enhance MRI contrast. Separate UNet models were trained for femur, tibia, and patella segmentation using a combined binary cross-entropy and Dice loss. Predicted masks were used to extract 2D and 3D BML volumetric features from medial and lateral compartments. These automated features were compared with manual features using six classifiers: Logistic Regression, SVM, Random Forest, CNN, XGBoost, and Decision Tree. Segmentation performance was evaluated using Dice scores and Pearson volume correlation, while classification performance was assessed using AUC, F1-score, and accuracy.

On Dataset 1, UNet achieved Dice scores of 0.8203 (femur), 0.7387 (tibia), and 0.8264 (patella), with strong Pearson volume correlations above 0.83. On the external Dataset 2, 2D Dice remained robust at 0.7713 (femur), 0.7235 (tibia), and 0.7817 (patella), with Pearson volume correlations all exceeding 0.70, demonstrating cross-dataset generalizability. In TKR prediction, automated UNet-derived features matched or outperformed manual annotations across multiple classifiers. Random Forest achieved the best performance with an AUC of 0.950 using automated features, compared to 0.866 using manual features. SVM also improved from 0.906 to 0.948 AUC when using automated features.

UNet-derived BML volumetric features provide predictive performance comparable to, and in some cases better than, expert manual annotations for TKR prediction. The strong volumetric agreement across datasets suggests that automated segmentation captures clinically meaningful lesion information, even when boundary-level overlap is moderate. These findings support the use of automated UNet-based pipelines as a scalable alternative to manual BML quantification for MRI-based TKR risk assessment in knee OA patients.