Deep Learning-Based Needle Tracking Trained on Bloch-Simulated Data and Evaluated on Clinical Real-Time bSSFP Images

Recently, Deep Learning-based methods were used to track the position and orientation of needles in MR images in real-time. Synthetic training data can be generated in large amounts, without data privacy restrictions, and without the need of animal experiments. Therefore, we have simulated the image acquisition using virtual human phantoms containing randomly placed metallic needles in a Bloch simulator. The synthetic images were used to train a U-net to predict the position and orientation of the needle within the susceptibility artifacts of clinical images in less than $$$90\,\text{ms}$$$.