Complex-Valued MRI Denoising for Non-Uniform Noise Artifacts

The patent transforms the denoising problem from spatial domain filtering to frequency domain processing by changing the parameter space. The complex-valued CNN operates on frequency-domain representations of MRI data, allowing adaptive noise suppression that preserves fine structures better than traditional spatial filters. This parameter transformation enables the system to handle non-uniform noise patterns effectively.

The patent replaces mechanical filter-based denoising systems with an intelligent neural network system. Instead of using fixed mathematical filters (NLM, BM3D), the system employs a trained complex-valued CNN that learns optimal denoising strategies from training data. This substitution allows the system to adapt to different noise patterns and preserve diagnostic features that mechanical filters would remove.

The patent implements self-service by enabling the CNN to learn from synthetic training data generated from simulated noise patterns. The system creates its own training dataset by programmatically adding various noise types and levels to synthetic MRI images, eliminating the need for extensive collection of real noisy clinical data. This self-generated training approach makes the system easily deployable without requiring access to large annotated datasets.

The patent uses copying by creating synthetic training examples that replicate real-world noise conditions. Instead of requiring actual noisy patient data, the system generates copies of synthetic MRI images with added noise patterns that mimic clinical scenarios. These synthetic copies serve as effective training data, allowing the network to learn denoising capabilities without accessing sensitive or difficult-to-obtain real patient data.