CT Image Artifact Correction Using Dual CNN Models

The patent replaces conventional iterative artifact correction algorithms with a trained neural network model that processes images in a single forward pass. The neural network, trained on paired artifact-containing and artifact-free images, learns to directly map artifact-ridden images to corrected versions, eliminating the need for time-consuming iterative optimization and achieving both high correction effectiveness and fast processing speed.

The patent performs preliminary training of the neural network model using large datasets of artifact-containing and artifact-free image pairs before actual correction tasks. This pre-training phase enables the model to learn artifact patterns and correction mappings in advance, so that during actual use, only a single forward propagation is needed without iterative refinement, dramatically reducing processing time while maintaining correction quality.

The patent replaces complex iterative correction algorithms with a trained neural network that performs correction through a single forward pass. The neural network architecture (such as U-Net or residual networks) with appropriate loss functions provides a unified, elegant solution that is both conceptually simpler and computationally more efficient than traditional iterative methods, while achieving superior or comparable correction effectiveness.

The patent transforms the artifact correction problem from an iterative optimization process with multiple adjustable parameters into a direct mapping problem solved by a neural network with fixed weights after training. By changing from algorithmic parameter tuning to learned parameter representation, the system achieves simpler implementation and faster execution while maintaining or improving correction quality.