AI-Based 3D Myocardial Blood Flow Mapping with 82Rb PET

The patent replaces traditional mechanical/computational methods with AI-based deep learning models (U-Net architecture) to automate the generation of 3D parametric maps. The neural network learns to directly map PET images to MBF parameter maps, eliminating the need for time-consuming manual segmentation and kinetic modeling while achieving comparable or superior precision.

The patent uses AI models to create digital copies of the complex physiological processes. The deep learning model learns from training data the relationship between PET images and MBF parameters, then replicates this relationship to generate accurate 3D parametric maps rapidly without repeating the complex computational steps of traditional methods.

The patent replaces traditional computational methods with AI-based deep learning models (U-Net architecture) to automate the generation of 3D parametric maps. The neural network learns to directly map PET images to MBF parameter maps, eliminating the need for time-consuming manual segmentation and kinetic modeling while achieving comparable or superior precision.

The patent incorporates feedback mechanisms through the training process, where the model continuously refines its ability to accurately segment myocardium and calculate MBF parameters. The training data includes ground truth labels that provide feedback on accuracy, enabling the model to learn from errors and improve stability in challenging regions like high arterial blood concentration areas.

The patent employs sophisticated segmentation techniques through the U-Net architecture that divides the 3D myocardium into numerous small volumetric elements (voxels). This fine-grained segmentation allows the model to detect and quantify small regional flow defects that would be invisible in coarser traditional methods, while maintaining overall measurement precision through the integrated kinetic modeling.

The patent transitions from traditional 2D polar map representations to 3D volumetric parametric maps. This dimensional change enables the visualization and detection of small regional flow defects in three-dimensional space, providing superior spatial resolution and the ability to identify subtle perfusion abnormalities that are lost in two-dimensional projections.

The patent performs preliminary action by training the AI model offline using extensive training data and computational resources. Once trained, the model weights are fixed and can be applied to new patients rapidly without requiring complex real-time computations. This preliminary training phase captures the complexity of the task, allowing for fast inference during actual clinical use.

The patent uses AI models to create digital copies of the complex physiological processes. The deep learning model learns from training data the relationship between PET images and MBF parameters, then replicates this relationship to generate accurate 3D parametric maps rapidly without repeating the complex computational steps of traditional methods.