HULFSynth : An INR based Super-Resolution and Ultra Low-Field MRI Synthesis via Contrast factor estimation

TL;DR

This work tackles the gap between High-Field (HF) and Ultra-Low-Field (ULF) MRI by proposing a physics-inspired, bidirectional synthesis framework that operates without HF supervision. A modular forward model estimates tissue-specific degradation factors $m\in\mathbb{R}^3$ from target ULF contrasts using a 3×3 matrix $A$ derived from tissue SNRs, guiding contrast modulation via $\phi$ to produce ULF-like images; an unsupervised Implicit Neural Representation (INR) then jointly predicts HF intensities $\hat{X}$ and tissue segmentations $\hat{S}_t$, enabling HF reconstructions at arbitrary resolutions through $\hat{X}_w=\sum_t \hat{X}\odot\hat{S}_t$. The HF synthesis path uses a Fourier-feature-augmented MLP with Gabor activations to map coordinates to $\hat{X}$ and $\hat{S}_t$, optimized with MAE, a Dice+Cross-Entropy segmentation loss, and TV regularization. Evaluations on synthetic ULF-like data and real 64mT/3T data show substantial WM-GM contrast improvements (up to $52\%$ and $37\%$, respectively) with modest losses in standard image quality metrics, and robustness to target contrast, noise, and seed initialization, while reducing hallucination risk compared to data-driven baselines. Overall, the method offers a trustworthy, physics-informed pathway for cross-field MRI enhancement that can improve accessibility and longitudinal comparability without requiring HF data for training.

Abstract

We present an unsupervised single image bidirectional Magnetic Resonance Image (MRI) synthesizer that synthesizes an Ultra-Low Field (ULF) like image from a High-Field (HF) magnitude image and vice-versa. Unlike existing MRI synthesis models, our approach is inspired by the physics that drives contrast changes between HF and ULF MRIs. Our forward model simulates a HF to ULF transformation by estimating the tissue-type Signal-to-Noise ratio (SNR) values based on target contrast values. For the Super-Resolution task, we used an Implicit Neural Representation (INR) network to synthesize HF image by simultaneously predicting tissue-type segmentations and image intensity without observed HF data. The proposed method is evaluated using synthetic ULF-like data from generated from standard 3T T$_1$-weighted images for qualitative assessments and paired 3T-64mT T$_1$-weighted images for validation experiments. WM-GM contrast improved by 52% in synthetic ULF-like images and 37% in 64mT images. Sensitivity experiments demonstrated the robustness of our forward model to variations in target contrast, noise and initial seeding.

Figures (3)

• Figure 1: High-Field Synthesis pipeline: (A) Given an observed ULF image and its segmentations (blue variables), our model jointly predicts HF-like image intensity and soft tissue segmentations (red variables) without the need for HF supervision. This is achieved by formulating our forward model ($\phi$) within the INR framework. At inference, INR predicts HF-like images at arbitrary resolutions. (B) Loss functions that govern the learning process, where $\mathrm{l_1, l_2, l_3, l_4}$ are tunable hyperparameters.
• Figure 2: Qualitative comparison of HF Synthesis (Axial plane). Rows 1 and 4 show IXI and LMIC subjects, with corresponding Canny edge maps in Rows 2 and 5. Subplots A(1), A(4) are observed ULF inputs and A(3), A(6) are FAST segmentations of ULF inputs. Enhanced WM-GM contrast with sharper edges can be visualized in our method (column D), when compared against baselines (E-G). (Table \ref{['tab:Quantiative']}).
• Figure 3: