SMS MRI Reconstruction Using Diffusion Plug-and-Play Denoising

The patent replaces conventional mechanical/reconstructive systems (parallel imaging and simultaneous multi-slice reconstruction algorithms) with a deep learning-based denoising diffusion model. This substitution allows the system to achieve high acceleration factors while maintaining image quality, as the diffusion model learns to denoise and reconstruct images from highly accelerated data in a way that conventional algorithms cannot, thereby resolving the contradiction between scan time reduction and image quality preservation.

The patent changes the fundamental parameters of the reconstruction process by introducing a diffusion probabilistic model that operates differently from traditional iterative reconstruction methods. The model uses a denoising network trained on synthetic data with varying acceleration factors, allowing it to adapt to different scan acceleration levels while maintaining image quality. This parameter change enables the system to achieve both fast scan times and high image quality simultaneously.

The patent creates a universal deep learning model that can handle multiple acquisition protocols and coil configurations through a single trained architecture. The denoising diffusion model is trained on synthetic data that encompasses a wide variety of acceleration factors and coil sensitivity patterns, enabling it to generalize to diverse real-world acquisition protocols. This universality allows the model to maintain high reconstruction accuracy across different scanning conditions without requiring protocol-specific models, thus resolving the contradiction between reconstruction performance and generalization capability.

The patent performs preliminary training of the diffusion model on synthetic SMS MRI data with varying acquisition protocols and coil configurations before actual scanning. This preliminary action creates a robust, pre-trained model that captures the essential characteristics of different acquisition scenarios. When the model is applied to real data, it can generalize effectively because it has already learned the relationships between different acquisition parameters and image qualities during training, thereby achieving both high reconstruction accuracy and broad generalization capability.

The patent uses synthetic data generation to create virtual copies of real MRI datasets. Instead of training on actual patient data from multiple institutions, the system generates synthetic SMS MRI data that replicates the statistical properties and anatomical variations of real scans. This copying approach allows the model to learn generalization capabilities from a vast amount of diverse data without the time and computational costs of collecting and processing real multi-center datasets. The synthetic data training process is significantly faster and more computationally efficient while achieving comparable or superior generalization performance.

The patent performs all training and model development as a preliminary action before actual clinical use. By pre-training the diffusion model on synthetic data with diverse acquisition protocols and anatomical variations, the system prepares a generalizable model in advance that can be applied to new patients and scanners without requiring retraining. This preliminary training phase captures the essential patterns and relationships needed for generalization, allowing rapid deployment to clinical settings without the ongoing computational burden of continuous training on diverse real-world data.