3T-to-7T MRI Synthesis With Spatial Alignment Network

The patent introduces a spatial alignment network (SAN) as an intermediary component between the generator and the final output. The SAN takes the generated 7T image and the target 7T image as inputs, computes a displacement field, and applies spatial transformation to correct misalignments. This intermediary module resolves the contradiction by providing additional alignment processing beyond standard linear registration, thereby improving both spatial alignment precision and image accuracy without requiring expensive 7T scanning.

The patent employs a generative adversarial network (GAN) to create a virtual copy of the 7T imaging system's output. The generator network is trained to map 3T input images to 7T output images by learning the complex nonlinear transformations between the two modalities. This copying approach allows the system to produce 7T-quality images from inexpensive 3T scanners, effectively replicating the high-resolution output without requiring access to expensive 7T hardware.

The patent performs preliminary action by pre-processing and aligning the paired 3T and 7T training images before feeding them to the GAN. The spatial alignment network is trained in advance on registered image pairs to learn the typical displacement patterns and anatomical variations. This preliminary alignment and training phase enables the model to quickly generate high-quality 7T images during inference without requiring time-consuming post-processing or re-alignment, thus improving generation quality while managing acquisition time efficiently.