Deep learning of personalized priors from past MRI scans enables fast, quality-enhanced point-of-care MRI with low-cost systems

TL;DR

This work tackles the barrier of accessible MRI by leveraging personalized priors from past high-field scans to accelerate and enhance low-field MRI. It introduces ViT-Fuser, a dual-transformer framework that fuses high-field priors with current low-field measurements using a Gram-based feature-space regularization and a hybrid pixel-space loss, enabling robust reconstruction across vendors, field strengths, and pulse sequences. Across simulated, phantom, and in-vivo data—including an ultra-low-field out-of-distribution case—the method demonstrates superior image quality, preserved anatomy, and resilience to varying SNR and undersampling. The approach has practical implications for fast, diagnostic-quality imaging at the point of care and lays groundwork for broader multi-field-strength MRI workflows and accessible longitudinal monitoring.

Abstract

Magnetic resonance imaging (MRI) offers superb-quality images, but its accessibility is limited by high costs, posing challenges for patients requiring longitudinal care. Low-field MRI provides affordable imaging with low-cost devices but is hindered by long scans and degraded image quality, including low signal-to-noise ratio (SNR) and tissue contrast. We propose a novel healthcare paradigm: using deep learning to extract personalized features from past standard high-field MRI scans and harnessing them to enable accelerated, enhanced-quality follow-up scans with low-cost systems. To overcome the SNR and contrast differences, we introduce ViT-Fuser, a feature-fusion vision transformer that learns features from past scans, e.g. those stored in standard DICOM CDs. We show that \textit{a single prior scan is sufficient}, and this scan can come from various MRI vendors, field strengths, and pulse sequences. Experiments with four datasets, including glioblastoma data, low-field ($50mT$), and ultra-low-field ($6.5mT$) data, demonstrate that ViT-Fuser outperforms state-of-the-art methods, providing enhanced-quality images from accelerated low-field scans, with robustness to out-of-distribution data. Our freely available framework thus enables rapid, diagnostic-quality, low-cost imaging for wide healthcare applications.

Figures (9)

• Figure 1: Overview of our proposed personalized imaging paradigm combining high-field and low-field MRI. We introduce a novel clinical workflow where a high-quality baseline scan is acquired at a highly equipped facility using a standard high-field MRI system (e.g., $1.5T$ or $3T$), and is followed by repeated, accessible low-field scans (e.g., $47mT$) at the point of care. While low-field MRI offers low-cost, portable imaging, it suffers from long scan durations and degraded image quality due to low SNR and reduced tissue contrast. Our approach leverages prior information from the baseline high-field scan to enhance follow-up low-field scans using a deep learning model, ViT-Fuser. This AI framework extracts personalized features from the prior scan and integrates them during image reconstruction, enabling accelerated, diagnostic-quality follow-up scans. The paradigm enables frequent, high-quality monitoring for patients requiring longitudinal care, without repeated hospital visits.
• Figure 2: Our framework and training strategy for the proposed multi-field-strengh clinical workflow. Our training scheme (a) involves a pre-training stage via a simulation environment, utilizing longitudinal high-field data, overcoming the lack of low-field data. Additionally, we illustrate the proposed hybrid loss that combines both pixel space loss and feature space loss, which is calculated differently per training stage. We also demonstrate the inference process (b), where features extracted from a prior high-field MRI scan are leveraged to enhance the reconstruction of a current low-field scan. The diagram depicts the feature extraction, fusion, and reconstruction stages. Finally, we provide an overview of the different datasets of our experiments (c).
• Figure 3: ViT-Fuser reconstruction compared to other methods: Accuracy and sharpness preservation for different SNR settings. Comparison of MoDL Aggarwal2019, ViT Lin2022, and ViT-Fuser (ours) for accelerated imaging for in-vivo Glioblastoma data Suter2022 with 3-fold equispaced undersampling at SNR levels of $5dB$ (bottom row) and $10dB$ (top row). ViT-Fuser achieves improved fine-detail reconstruction and better texture preservation as SNR decreases from $10dB$ to $5dB$. These results indicate that using priors could be beneficial for improving image quality in low SNR regimes.
• Figure 4: ViT-Fuser reconstructions compared to other methods: Longitudinal change and similarity preservation experiments. Comparison of MoDL Aggarwal2019, ViT Lin2022, and ViT-Fuser (ours) for in-vivo Glioblastoma data Suter2022 with 8-fold Poisson Disc undersampling with SNR of $10dB$. ViT-Fuser demonstrates decent reconstruction of longitudinal anatomical changes (subject 1, yellow arrow) and temporal similarities (subject 2).
• Figure 5: Robustness to varying SNR levels and acceleration factors. Top: SSIM, LPIPS, and CMMD metrics versus varying SNR level (dB) for 3-fold equispaced sampling. Bottom: metrics versus acceleration factor using 2D Poisson-disc undersampling at $10dB$ SNR. Note that the SNR axis is in dB, i.e., the shift from $0dB$ to $20dB$ corresponds to a 100-fold linear increase in SNR. In all cases, our method achieves the best results (highest SSIM, lowest LPIPS and CMMD) and demonstrates high robustness (most moderate slope) across the studied spectrum, i.e. for varying SNR or acceleration factors.