Diffusion Probabilistic Generative Models for Accelerated, in-NICU Permanent Magnet Neonatal MRI

TL;DR

This study tackles the challenge of long NICU MRI scans with a 1T permanent-magnet system by developing diffusion probabilistic generative models trained on real-world, low-SNR neonatal data. A novel pipeline combines multi-contrast class embeddings, self-supervised denoising, and posterior-sample averaging to reconstruct under-sampled data, decoupling the prior from the measurement model. The approach is validated on a real NICU dataset with a clinical reader study, showing that approximately 1.5× acceleration yields clinically acceptable images and non-inferior structure delineation for several brain regions. The work demonstrates the practical potential of diffusion-based priors for robust, motion-tolerant, accelerated in-NICU neonatal MRI, with implications for broader adoption and motion-correction integration.

Abstract

Purpose: Magnetic Resonance Imaging (MRI) enables non-invasive assessment of brain abnormalities during early life development. Permanent magnet scanners operating in the neonatal intensive care unit (NICU) facilitate MRI of sick infants, but have long scan times due to lower signal-to-noise ratios (SNR) and limited receive coils. This work accelerates in-NICU MRI with diffusion probabilistic generative models by developing a training pipeline accounting for these challenges. Methods: We establish a novel training dataset of clinical, 1 Tesla neonatal MR images in collaboration with Aspect Imaging and Sha'are Zedek Medical Center. We propose a pipeline to handle the low quantity and SNR of our real-world dataset (1) modifying existing network architectures to support varying resolutions; (2) training a single model on all data with learned class embedding vectors; (3) applying self-supervised denoising before training; and (4) reconstructing by averaging posterior samples. Retrospective under-sampling experiments, accounting for signal decay, evaluated each item of our proposed methodology. A clinical reader study with practicing pediatric neuroradiologists evaluated our proposed images reconstructed from 1.5x under-sampled data. Results: Combining all data, denoising pre-training, and averaging posterior samples yields quantitative improvements in reconstruction. The generative model decouples the learned prior from the measurement model and functions at two acceleration rates without re-training. The reader study suggests that proposed images reconstructed from approximately 1.5x under-sampled data are adequate for clinical use. Conclusion: Diffusion probabilistic generative models applied with the proposed pipeline to handle challenging real-world datasets could reduce scan time of in-NICU neonatal MRI.

Figures (20)

• Figure 1: The proposed training and inference pipeline. Images from different orientations and contrasts are combined into a single dataset using class embeddings. A denoiser, trained in a self-supervised fashion, denoises the dataset before training. Diffusion posterior sampling reconstructs an image from under-sampled k-space by averaging 5 posterior samples. The proposed methods enable training of diffusion models for accelerated reconstruction on noisy and limited in-NICU neonatal MRI data.
• Figure 2: (a) The top row shows two training samples from our dataset and the bottom row shows the corresponding training samples after applying our denoiser trained in a self-supervised fashion. (b,c,d,e) prior samples generated by our trained model when conditioned on class embeddings of fse axial, sagittal, coronal, and se axial. Our model uses all available training data to learn a statistical prior over neonatal MR images.
• Figure 3: Violin plots comparing NRMSE of posterior sampling reconstructions on $R=2$ under-sampled data using diffusion models trained on just a single image class, trained on all data, and trained on all data with class embeddings. Results on all contrasts and orientations suggest that a diffusion model trained by combining all data with class embeddings yields the best quantitative performance.
• Figure 4: Example images from the experiment comparing posterior sampling reconstructions on $R = 2$ under-sampled data using diffusion models trained on just a single image class, trained on all data, and trained on all data with class embeddings. This is a specific illustration of the quantitative conclusion that using all data combined with class embeddings to train the diffusion model yields best performance.
• Figure 5: Violin plots comparing NRMSE of reconstructions on $R=1.5$ under-sampled data using baseline L$_1$ and diffusion models trained on all data with and without denoising. While using a learned prior provides benefit, little quantitative difference exists between models trained with and without denoising pre-training.