Super-Field MRI Synthesis for Infant Brains Enhanced by Dual Channel Latent Diffusion

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Portable ultra-low-field (uLF) MRI systems offer expanded accessibility in resource-limited settings, particularly for neonates and infants.
• However, uLF MRI image quality is inferior to high-field (HF) MRI, limiting its clinical and research utility.
• Addressing this gap is crucial for improving diagnostic capabilities in underserved populations.

Purpose of the Study:

• To introduce Super-Field Network (SFNet), a novel deep learning model designed to generate high-field (HF) comparable magnetic resonance imaging (MRI) images from ultra-low-field (uLF) inputs.
• To enhance a small dataset of infant uLF-HF MRI data using latent diffusion for improved model training.
• To evaluate SFNet's performance against state-of-the-art methods for image synthesis and brain tissue segmentation.

Main Methods:

• Development of SFNet, a custom swinUNETRv2 with generative adversarial network components, for synthesizing super-field (SF) images from uLF MRIs.
• Novel application of latent diffusion models to create dual-channel (uLF-HF) paired MRIs for dataset augmentation.
• Comparative analysis using Fréchet inception distance, perceptual similarity scores, peak signal-to-noise ratio (PSNR), and Dice similarity coefficients for white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) segmentation.

Main Results:

• SFNet, trained on the latent diffusion-enhanced dataset, achieved state-of-the-art results with a Fréchet inception distance of 9.08 ± 1.21, perceptual similarity of 0.11 ± 0.01, and PSNR of 22.64 ± 1.31.
• Segmentation of WM, GM, and CSF in the developing infant brain showed strong overlap with true HF images (Dice coefficients: 0.71 ± 0.1, 0.79 ± 0.2, 0.73 ± 0.08, respectively).
• SFNet-generated segmentations demonstrated significant improvements over uLF-based segmentations, with 166%, 107%, and 106% increases for WM, GM, and CSF, respectively.

Conclusions:

• SFNet effectively generates super-field (SF) images from uLF MRI data, achieving quality comparable to high-field (HF) MRI.
• The integration of latent diffusion for data augmentation significantly enhances SFNet's performance, particularly for small or underrepresented datasets.
• SF MRI technology holds significant promise for advancing health equity by improving the diagnostic utility of low-cost, portable MRI systems in resource-limited settings and for vulnerable populations.