BDU-Net: An Edge-Segmentation-Oriented U-Shaped Network for Pediatric Knee Joint Segmentation.

Area of Science:

• Medical imaging analysis
• Pediatric orthopedics
• Artificial intelligence in healthcare

Background:

• Accurate segmentation of pediatric knee cartilage in MRI is crucial for assessing bone development and identifying risks.
• Challenges include variations in cartilage size/shape, low contrast, and fuzzy boundaries in pediatric knee MRI.
• High-precision automatic segmentation models are needed for effective cartilage monitoring and early intervention.

Purpose of the Study:

• To develop a novel, high-precision automatic segmentation model for pediatric knee cartilage MRI.
• To address the challenges of segmentation accuracy, edge preservation, and noise suppression in pediatric knee cartilage images.
• To improve the quantitative assessment of cartilage development and enable early detection of potential issues.

Main Methods:

• Proposed BDU-Net, a UNet++-based segmentation model incorporating an edge-preserving enhancement module (EPEM) using ordinary differential equations (ODE) and the Runge-Kutta second-order (RK2) method.
• Integrated a multi-scale feature extraction module (MSFEM) in the bridge section for enhanced global and local feature modeling.
• Employed dynamic feature-weighted fusion to improve edge perception.

Main Results:

• BDU-Net demonstrated superior performance over state-of-the-art methods on three pediatric knee cartilage datasets (PC, MCC, LCGP).
• Achieved high Intersection over Union (IoU) scores: 0.7519 (PC), 0.8283 (MCC), and 0.8485 (LCGP), outperforming comparative methods.
• Showcased significant improvements in segmentation accuracy, edge preservation, and noise suppression, validated by qualitative analysis and expert scoring.

Conclusions:

• The proposed BDU-Net model effectively addresses the challenges of segmenting pediatric knee cartilage in MRI.
• BDU-Net offers clear performance advantages and significant application potential for monitoring cartilage development and enabling early intervention in children.
• The model's ability to enhance edge perception and model complex features contributes to its high accuracy and reliability.