Ultrasound-guided musculoskeletal minimally invasive procedures offer clear clinical benefits, yet remain technically challenging to master. Existing deep learning-based ultrasound simulation methods often focus on static, fixed-plane images, lacking real-time and dynamic capabilities essential for clinical training. To address this, we propose RDG-USIS: a Real-time, Dynamic, and Generalizable UltraSound Image Simulation algorithm designed to enhance training for minimally invasive procedures. We first introduce a CT-ultrasound data acquisition method with 3D spatial constraints, generating paired CT-ultrasound data by resampling CT volumes according to ultrasound probe pose. RDG-USIS consists of two stages: structural segmentation using a Total-Segmentator based on nnU-Net, followed by ultrasound simulation via an improved convolutional ray-casting algorithm. A CycleGAN model then translates the simulated images into realistic ultrasound style, trained using both synthetic and real data. We benchmark RDG-USIS against Pix2Pix and Diffusion models, and conduct ablation studies. The results show superior structural accuracy and visual realism. Finally, the model is integrated into a multi-modal training system and evaluated in clinical tasks, confirming its real-time performance and generalizability. The source code is available at https://github.com/JZK00/RDG-USIS .