CNN MRI Tag Tracking via Synthetic Bloch Data

The patent replaces traditional mechanical post-processing tracking methods with a deep learning-based automated system. A neural network model is trained to directly predict tag line positions and motion paths from tagged MRI images, eliminating the need for manual segmentation and tracking corrections. This substitution transforms a complex, error-prone manual process into an automated, reliable computational process that requires minimal user input.

The patent implements self-service by enabling the system to automatically perform tag tracking without requiring expert user intervention. The deep learning model autonomously identifies tag lines, tracks their motion through the cardiac cycle, and calculates strain measurements. The system serves itself by incorporating synthetic data generation and automated model training, making the entire workflow self-sufficient and eliminating dependence on manual operator skills.

The patent applies copying by creating synthetic training data that replicates real tagged MRI images and their corresponding ground truth motion paths. Instead of requiring大量 real patient data with manually annotated ground truth, the system generates synthetic copies through simulation. A digital twin of the tagging process is created, where virtual tag lines are imposed on synthetic cardiac images, producing unlimited training examples with known ground truth motion paths for supervising the neural network.

The patent implements preliminary action by pre-generating and storing synthetic training data with known ground truth before actual tag tracking is performed. The synthetic dataset is created in advance through simulation, containing diverse examples of tag line positions, motions, and fading patterns. This pre-prepared training corpus enables the neural network to be trained offline, so that during clinical use, only inference is needed without requiring additional manual annotation or ground truth generation.

The patent applies preliminary action by pre-training the neural network model on extensive synthetic data that includes various noise patterns, artifacts, and fading conditions. The model learns robust feature extraction and tracking patterns during offline training, enabling fast inference during actual use. Once trained, the model performs tracking in real-time or near-real-time without requiring manual supervision during the actual tag tracking process, thus reducing both manual effort and computation time during clinical operation.

The patent uses disposable synthetic training data that can be generated indefinitely without degradation. Instead of relying on limited real patient data that would need to be repeatedly annotated, the system uses computationally generated synthetic images that can be produced at will. These synthetic training examples are computationally inexpensive to generate and provide unlimited training samples, making the training process efficient and the resulting model robust to various imaging conditions.