Accelerated, motion robust low-field Neonatal MRI with physics based motion models and generative machine learning

Artificial intelligence system for accelerated and motion-robust reconstruction of low-field neonatal MRI images

This technology uses AI-powered generative models and physics-based motion correction to produce faster, clearer, and more reliable low-field neonatal MRI scans, reducing motion artifacts and scan times while improving image quality without needing sedation or patient transport.

Background

Magnetic Resonance Imaging (MRI) is a critical tool in neonatal medicine, offering non-invasive, high-resolution visualization of the developing brain and other organs in newborns. However, the use of MRI in neonates presents unique challenges, particularly when utilizing low-field systems (below 1.5T), which are favored for their safety, portability, and compatibility with neonatal intensive care units (NICUs). The need for improved neonatal MRI arises from the vulnerability of this patient population: neonates are highly susceptible to the risks associated with sedation and transport required for conventional high-field MRI. Low-field, bedside MRI systems have the potential to revolutionize neonatal care by enabling safer, more accessible imaging directly within the NICU environment. Despite these advantages, current approaches to low-field neonatal MRI are hampered by several significant limitations. Long scan times increase the likelihood of patient motion, leading to severe motion artifacts that can render images non-diagnostic. The inherently low signal-to-noise ratio (SNR) of low-field systems further degrades image quality, making it difficult to discern subtle anatomical details. Traditional MRI reconstruction methods often fail to adequately address motion, as their models do not account for patient movement, and acceleration techniques developed for adult imaging are typically inapplicable due to differences in coil design and anatomy. Additionally, the scarcity of high-quality neonatal MRI datasets limits the effectiveness of data-driven machine learning approaches, further constraining image quality and diagnostic utility. These persistent issues highlight the urgent need for new solutions that can deliver fast, motion-robust, and high-fidelity neonatal MRI without compromising patient safety.

Technology Description

This technology is an advanced AI-driven system for improving low-field neonatal MRI imaging, specifically designed to address the unique challenges of imaging newborns. It integrates a diffusion-based generative neural network with a physics-based motion model to simultaneously accelerate scan times and correct for motion artifacts, which are common in neonatal imaging due to patient movement and low signal-to-noise ratios (SNR). The system is trained on a curated dataset of neonatal MRI scans, enhanced through self-supervised denoising techniques to boost SNR, and leverages class embeddings to allow a single model to generalize across different MRI contrasts and orientations. During image reconstruction, the generative model acts as a statistical prior, while the physics-based model accounts for rigid body motion, enabling an iterative process called Motion-Informed Posterior Sampling (MI-PS) that jointly estimates both the clean image and motion parameters from under-sampled, motion-corrupted data. What differentiates this technology is its holistic approach that combines state-of-the-art deep learning with explicit physical modeling of patient motion, tailored specifically for the neonatal population and low-field MRI systems. Unlike traditional MRI reconstruction methods, which often fail in the presence of motion and are not optimized for the limited data and unique anatomical features of neonates, this solution uses a single, robust generative model capable of handling heterogeneous data and various imaging scenarios. The use of self-supervised denoising significantly enhances the quality of training data, while class embeddings enable the model to adapt to multiple scan types without requiring separate models for each. Experimental results demonstrate superior performance in both accelerated imaging and motion correction compared to conventional methods, with the added benefit of reducing the need for sedation and repeat scans—leading to improved safety, cost savings, and broader accessibility for neonatal care.

Benefits

• Significantly reduces neonatal MRI scan times through accelerated imaging techniques.
• Robustly corrects motion artifacts by integrating physics-based motion modeling with generative AI.
• Enhances image quality and signal-to-noise ratio (SNR) using self-supervised denoising tailored for low-field MRI.
• Generalizes across multiple MRI contrasts and orientations via class embeddings in a single unified model.
• Enables safer neonatal imaging by reducing the need for sedation and patient transport in NICUs.
• Improves diagnostic reliability with higher-fidelity neonatal MRI scans despite data scarcity and motion challenges.
• Potentially reduces healthcare costs by minimizing repeat scans caused by motion artifacts.
• Flexible framework adaptable to various acquisition settings and artifact scenarios for broader clinical applications.

Additional Information

This system integrates a diffusion-based generative neural network with a physics-based motion model to enhance low-field neonatal MRI. Trained on denoised data with class embeddings, it uses Motion-Informed Posterior Sampling to simultaneously estimate clean images and motion parameters from under-sampled, corrupted measurements, enabling accelerated, motion-robust, high-fidelity scans