Deep Learning Denoising for MRI Image Quality

The deep learning network is trained in advance using a large dataset of noisy and denoised MR images to learn the mapping between noisy inputs and clean outputs. This preliminary training phase allows the network to capture noise patterns and denoising strategies, so that during actual medical imaging applications, the network can rapidly apply learned denoising operations without requiring complex real-time computations.

The patent uses multiple copies of the same deep learning network architecture (e.g., multiple U-Net models with different initialization or training configurations) to process the same input image. These parallel network copies provide diverse denoising perspectives and can be aggregated to produce more robust results, improving image quality while distributing the computational load across multiple identical structures rather than one highly complex variable architecture.

The patent extracts only the essential training data needed for effective network training by selectively acquiring multiple excitation images at specific k-space lines rather than acquiring complete multiple images for every training sample. This extraction approach provides sufficient statistical information about noise characteristics while minimizing the total acquisition time required for training data collection.

The patent acquires more excitation images than the absolute minimum required for training (excessive action) to ensure robust noise characterization, but only at strategically selected k-space lines rather than uniformly across the entire k-space. This partial application of excessive action provides adequate training data for high denoising accuracy while keeping the time penalty acceptable by concentrating acquisitions in critical regions.