Deep Learning Contrast-Enhanced MRI Image Generation

The patent uses deep learning models to create synthetic copies of high-quality contrast-enhanced images from low-dose input images. The generative adversarial networks (GANs) learn to replicate the appearance and diagnostic features of full-dose images without requiring actual high-dose contrast agents, thus producing artifact-free images that match the quality of higher contrast doses while maintaining safety by avoiding extra contrast dose administration

The patent transforms the input low-dose images by adjusting key parameters such as contrast enhancement levels, noise characteristics, and intensity distributions to match the statistical properties of high-dose images. This parameter transformation allows the model to generate images with improved signal-to-noise ratio and contrast-to-noise ratio without physically administering more contrast agent

The patent replaces the mechanical/physical approach of extending scan time to improve image quality with a computational approach using deep learning algorithms. Instead of acquiring more data through prolonged scanning, the system uses artificial intelligence to enhance and reconstruct images from existing low-dose, short-duration scans, thereby eliminating the trade-off between scan time and image quality

The patent performs preliminary training of deep learning models using datasets containing both low-dose and high-quality reference images. This pre-training enables the model to learn the transformation patterns and enhancement strategies needed to improve image quality from short scans, so that during actual clinical use, high-quality images can be generated rapidly without requiring prolonged scan times

The patent introduces an intermediary processing stage between the low-dose input images and the final output, where generative adversarial networks act as a mediator to transform and enhance the images. The GAN architecture includes a generator that creates enhanced images and a discriminator that evaluates their quality, working together to produce artifact-free images that maintain diagnostic accuracy while reducing contrast agent retention

The patent employs a composite deep learning architecture combining multiple model types including U-Net-based generators, adversarial discriminators, and auxiliary loss functions. This composite approach integrates different computational techniques to simultaneously reduce artifacts, enhance contrast, and preserve anatomical structures, thereby achieving high image quality without the streaks and artifacts that plague simpler deep learning methods

The patent trains deep learning models on diverse datasets comprising images from multiple scanner vendors, different clinical sites, and various imaging protocols. The model learns universal features and transformation patterns that are transferable across different scanning conditions and hardware platforms, enabling a single model to generalize effectively without requiring site-specific customization or increasing operational complexity

The patent employs a modular deep learning architecture that segments the image processing task into distinct functional components such as noise reduction, contrast enhancement, and artifact suppression. Each module can be independently trained and optimized on specific datasets, then combined to form a comprehensive model that achieves generalizability across different sites and scanners while maintaining manageable complexity through functional decomposition