Autoencoder MRI Reconstruction with Multi-Contrast Input

The patent combines multiple contrast MRI images (T1-weighted, T2-weighted, FLAIR) into a unified reconstruction framework. The neural network simultaneously processes multiple contrast images as input channels, merging the information from different contrasts to produce a single high-quality reconstructed image, thereby resolving the contradiction between simplicity and quality.

The patent uses a composite approach by integrating multiple types of MRI contrast information (different weighting images) into a single reconstruction model. This composite input structure allows the network to leverage complementary information from various contrasts, improving reconstruction quality without significantly increasing operational complexity.

The patent implements a universal neural network architecture that can handle multiple contrast types through a single model. The network is designed with multi-channel input capability that universally processes different MRI contrasts (T1, T2, FLAIR) using the same computational framework, avoiding the need for separate processing pipelines for each contrast type.

The patent performs preliminary alignment and normalization of multiple contrast images before they enter the neural network. This preprocessing step ensures that the multi-contrast inputs are properly registered and scaled, reducing the computational burden during the main reconstruction process and mitigating the increase in processing complexity.

The patent replaces traditional mechanical/image-processing reconstruction algorithms with a neural network-based approach. This substitution uses learned patterns from training data to reconstruct images, replacing conventional iterative mathematical methods with a data-driven model that reduces artifacts while maintaining reasonable computational efficiency.

The patent changes the fundamental parameters of the reconstruction process by using a trained neural network model with fixed weights and biases. Once trained, the network processes images through a fixed computational graph with optimized parameters, achieving high diagnostic reliability without requiring extensive computational resources during inference.

The patent segments the reconstruction task into distinct functional components within the neural network: an encoder that extracts features from multiple contrast inputs, a decoder that reconstructs the image, and skip connections that preserve spatial information. This segmentation of the model architecture makes the complex deep learning approach more manageable and efficient.

The patent introduces skip connections as intermediary pathways that directly transmit feature maps from encoder layers to corresponding decoder layers. These intermediaries preserve important spatial and contextual information that would otherwise be lost in the encoding-decoding process, enabling high-quality reconstruction without requiring excessively deep or complex network architectures.