<p>The growth plate and articular cartilage are essential for children’s bone development. Precise segmentation of cartilage in MRI images enables the extraction of quantitative indicators for health assessment and risk identification. Therefore, developing high-precision automatic segmentation models is of great importance for monitoring cartilage development and enabling early intervention. However, in pediatric knee joint MRI images, there are significant variations in the size and shape of the cartilage, the cartilage’s gray value is close to that of the surrounding tissue or synovial fluid, and the boundaries are often fuzzy. To address these challenges, this paper proposes a new UNet++-based segmentation model, BDU-Net. In this model, an edge-preserving enhancement module (EPEM) is designed based on ordinary differential equations (ODE), with the Runge–Kutta second-order (RK2) method introduced to model and strengthen complex textures and contour regions. The edge perception ability is further improved through dynamic feature-weighted fusion. In addition, a multi-scale feature extraction module(MSFEM) is integrated into the bridge section to enhance the joint modeling of global context and local details, thereby improving the model’s ability to focus on and represent key regions. Experiments on three pediatric knee cartilage datasets (PC, MCC, LCGP) demonstrate that BDU-Net outperforms existing state-of-the-art methods in segmentation accuracy, edge preservation, and noise suppression. The proposed method achieves IoU values of 0.7519, 0.8283, and 0.8485 on the three datasets, while the best results from the compared methods are 0.7456, 0.8184, and 0.8352. It also achieves strong results in qualitative analysis and expert scoring, showing clear performance advantages and application potential.</p>