OPTIMIZED DEEP LEARNING METHOD FOR AUTOMATED SEGMENTATION OF BONE MARROW LESIONS

Abstract

INTRODUCTION

Bone Marrow Lesions (BMLs), characterized by high-signal intensity on fat-suppressed MRIs, are associated with the progression of knee osteoarthritis (OA). In early OA or when joint damage is not visible on radiographs, BMLs are predictive markers for progression. However, their irregular distribution, potentially large size, and low-contrast boundaries challenge BML segmentation.

OBJECTIVE

This study introduces a novel training strategy for enhancing automated BML segmentation accuracy

METHODS

We aimed to optimize a deep learning method for automatic BML detection and segmentation in MRI, using the Osteoarthritis Initiative (OAI) dataset split into 70% training (210 participants), 15% validation (45 participants), and 15% testing (45 participants), totaling 1025, 190, and 201 MRIs, respectively. Images were employed using data augmentation like brightness, contrast, and geometric transformations. We applied a closing operation, a morphological technique combining dilation and erosion, to smooth edges, addressing the coarse manual labels that impair training. Several models (U-net, SwinUnetR, AttentionUnet, and U-net++) were trained with single-label (BML) and dual-label (BML + femur bone) outputs. Model performance was measured with the Dice Similarity Coefficient (DSC) for overlap and HD95 for boundary error. Cross-entropy and Dice loss functions improved sensitivity during training, particularly in dual-label channels where the femur bone location helped constrain BML positions. We also applied Pixel-Wise Voting (PWV) to improve segmentation stability and accuracy by averaging results from image variations, reducing false positives, and enhancing final segmentation outcomes.

RESULTS

UNet++ model with dual-label (BML + femur bone) yielded the best accuracy, outperforming U-net, SwinUnetR, and AttentionUnet. Figure 1 shows its predicted region (yellow) overlapping well with the manually labeled BML and aligning with boundaries. Specifically, the dual-label Unet model with PWV improved DSC from 62.21% to 64.88% for BML and to 96.52% for bone, while HD95 dropped to 26.82% for BML and 15.52% for bone. SwinUnetR with dual-label and PWV also showed improved DSC (65.06% to 66.70% for BML; 96.34% for bone) and reduced HD95 to 28.31% for BML and 11.54% for bone. AttentionUnet exhibited notable PWV improvements in bone segmentation. Overall, Unet++ achieved the highest performance with dual-label and PWV, increasing DSC from 66.16% to 68.48% for BML and 96.66% for bone, with the lowest HD95 values.

CONCLUSION

This study employed augmentation strategies, a closing operation, and both single- and dual-label analyses to train four models—Unet, SwinUnetR, AttentionUnet, and Unet++. Cross-entropy loss and Pixel-Wise Voting (PWV) enhanced model performance, with dual-label consistently outperforming single-label, especially with PWV. Our findings highlight the potential of automated segmentation as a powerful tool for researchers.