Cyclic Self-Supervised Diffusion for Ultra Low-field to High-field MRI Synthesis

TL;DR

This work introduces CSS-Diff, a cycle-constrained diffusion framework for translating low-field MRI into high-field–quality scans while preserving anatomical fidelity. It combines a diffusion-based synthesis chain with a cycle-consistency loss and two self-supervised pretraining modules—Slice-wise Gap Perception and Local Structure Correction—to address spurious detail, slice misalignment, and structural artifacts. Empirical results across multiple datasets and contrasts show state-of-the-art PSNR, SSIM, and LPIPS, along with improved segmentation accuracy and alignment with high-field references. The approach demonstrates robust cross-site generalization and offers practical potential for broadening access to clinically reliable, high-field–like MRI data.

Abstract

Synthesizing high-quality images from low-field MRI holds significant potential. Low-field MRI is cheaper, more accessible, and safer, but suffers from low resolution and poor signal-to-noise ratio. This synthesis process can reduce reliance on costly acquisitions and expand data availability. However, synthesizing high-field MRI still suffers from a clinical fidelity gap. There is a need to preserve anatomical fidelity, enhance fine-grained structural details, and bridge domain gaps in image contrast. To address these issues, we propose a \emph{cyclic self-supervised diffusion (CSS-Diff)} framework for high-field MRI synthesis from real low-field MRI data. Our core idea is to reformulate diffusion-based synthesis under a cycle-consistent constraint. It enforces anatomical preservation throughout the generative process rather than just relying on paired pixel-level supervision. The CSS-Diff framework further incorporates two novel processes. The slice-wise gap perception network aligns inter-slice inconsistencies via contrastive learning. The local structure correction network enhances local feature restoration through self-reconstruction of masked and perturbed patches. Extensive experiments on cross-field synthesis tasks demonstrate the effectiveness of our method, achieving state-of-the-art performance (e.g., 31.80 $\pm$ 2.70 dB in PSNR, 0.943 $\pm$ 0.102 in SSIM, and 0.0864 $\pm$ 0.0689 in LPIPS). Beyond pixel-wise fidelity, our method also preserves fine-grained anatomical structures compared with the original low-field MRI (e.g., left cerebral white matter error drops from 12.1$\%$ to 2.1$\%$, cortex from 4.2$\%$ to 3.7$\%$). To conclude, our CSS-Diff can synthesize images that are both quantitatively reliable and anatomically consistent.

Figures (8)

• Figure 1: The motivation and challenges of the proposed CSS-Diff framework. (a) Motivation: Low-field MRI is portable but blurry and inaccurate, while high-field MRI is precise but costly and immobile. Synthesizing high-field quality from low-field inputs improves clarity and diagnostic reliability. (b) Multi-aspect fidelity gap: The task faces three challenges: spurious details from contrast–resolution disparity, slice-wise gaps from spatial mismatches, and structural corruption with artifacts or false patterns. (c) The CSS-Diff uses a reverse-preserve strategy with self-supervised guidance. It perceives slice-wise gaps, extracts inter-slice features, and enforces local structural constraints. This enables cycle-consistent synthesis of high-field MRI with preserved anatomical fidelity.
• Figure 2: Detailed architecture of the proposed cyclic self-supervised diffusion (CSS-Diff) framework. (a) Progressive self-supervised diffusion gradually enhances MRI quality from low-field to high-field. (b) The framework incorporates slice-wise gap perception (SGP), local-structure correctness (LSC), and adversarial training to guide high-fidelity MRI synthesis. (c) Data Synthesis and Adversarial Training aims to synthesise high-field MRI data from low-field MRI inputs using a synthesis network.
• Figure 3: Visualization result of different baselines for exemplar regions. (top and bottom row of each panel, CSS-Diff$^\dagger$ denotes the CSS-Diff baseline model, while CSS-Diff indicates CSS-Diff with all modules enabled) (a) Synthesis of high-field MRI data from cross-contrast low-field MRI data. (b) Synthesis of ultra high-field MRI from same-contrast low-field MRI.
• Figure 4: Visualization result of SGP and LSC process. (a) SGP enhances inter-slice similarity by pre-training on sequential low-field and high-field MRI data and matching the most similar slices within a randomly shuffled batch. (b) LSC enhances local structures by recovering fine image details from locally masked and rotated images.
• Figure 5: Ablation study on sampling and network parameters of the CSS-Diff framework, evaluated on the paired 64 mT $\rightarrow$ 3 T dataset.