Model-Based Neural Reconstruction for Optoacoustic and Ultrasonic Imaging

The patent applies preliminary action by pre-training the neural network offline using simulated training data generated from the forward model. This pre-computation phase creates a ready-to-use reconstruction model that can rapidly process real measurement data without requiring complex iterative calculations during live imaging, thus achieving real-time speed while maintaining quality.

The patent uses copying by creating synthetic training data through simulation that replicates real measurement conditions. The neural network is trained on these copied simulated datasets, allowing it to learn reconstruction patterns without requiring extensive real patient data, thereby enabling fast real-time reconstruction of actual images while preserving diagnostic quality.

The patent replaces the mechanical/iterative reconstruction system with a neural network-based system. Instead of using complex iterative algorithms that require extensive computational resources to process high-resolution, high-data-rate images, the system uses a pre-trained neural network that has learned efficient reconstruction patterns, thereby reducing real-time computational complexity while maintaining full data utilization.

The complex computational work is performed in advance during the offline training phase. The neural network learns optimal reconstruction strategies for handling high-data-rate, high-resolution images during training, so that during real-time operation, the system can efficiently process incoming data without bearing the full computational burden, thus reducing real-time complexity while preserving data fidelity.

The patent uses copying by training the neural network on extensively processed simulated data that has been reconstructed using accurate iterative methods. The network learns from these high-quality reference reconstructions, capturing accurate image features during the offline training phase, enabling fast real-time inference without sacrificing the accuracy benefits of iterative reconstruction.

The patent substitutes the iterative mechanical reconstruction process with a neural network-based system. The complex iterative optimization process is replaced by a neural network that has been trained to produce similar accurate results through learned patterns, thereby achieving comparable reconstruction accuracy with dramatically reduced processing time suitable for real-time applications.

The patent applies dynamics by making the speed of sound parameter adjustable and adaptable during the training and reconstruction process. The neural network can be trained with different speed of sound values to match different tissue types, and this parameter can be dynamically tuned for live imaging applications, enabling the system to adapt to various imaging conditions without requiring complex hardware modifications.

The patent uses parameter changes by allowing the speed of sound parameter to be varied during training and operation. Different speed of sound values can be selected based on the specific tissue being imaged, and the neural network is trained to accommodate these parameter variations, thereby achieving versatile dynamic focusing capability while keeping the system architecture relatively simple through software-based parameter adjustment.