SIINR: Structurally Informed Implicit Neural Representations for super-resolution with uncertainty quantification of clinical quality diffusion MRI datasets

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Medical Imaging · Depth: Expert, extended

Summary

SIINR (Structurally Informed Implicit Neural Representations) is a novel framework designed for super-resolution and uncertainty quantification of clinical diffusion Magnetic Resonance Imaging (dMRI) datasets. It addresses the challenge of low out-of-plane resolution in clinical acquisitions by combining a supervised 3D U-net as a prior with a self-supervised implicit neural representation (INR). The U-net is trained on 428 subjects from diverse open-access dMRI datasets, totaling 37,820 volumes, using a 70-20-10 train-validation-test split. The INR component, which uses a four-layer MLP with a 1024 hidden dimension and Fourier feature encoding (n_p 5000, sigma^2 3.5, l_max 8), integrates the high-resolution prior and original low-resolution data, enabling joint spatial and angular modeling, data consistency, and analytic posterior distributions for uncertainty. Validated on test sets and clinical cases, including subjects with multiple sclerosis and brain lesions, SIINR consistently outperforms standard interpolation methods in quantitative error metrics and anatomical fidelity, particularly for super-resolution factors up to 4.

Key takeaway

For Machine Learning Engineers developing dMRI analysis pipelines, SIINR offers a robust solution for enhancing low-resolution clinical data. You should consider integrating SIINR to improve structural detail in thick-slice acquisitions and to provide crucial uncertainty quantification for derived metrics like ADC and FA. This framework allows for more reliable interpretation of neuroimaging results, especially in pathological cases, by flagging uncertain regions.

Key insights

SIINR combines a supervised U-net with a self-supervised INR for dMRI super-resolution and uncertainty quantification.

Principles

• Fuse supervised priors with self-supervised INRs for robust reconstruction.
• Leverage INRs for joint spatial and angular domain modeling.
• Quantify uncertainty using analytic posterior distributions from INRs.

Method

SIINR uses a 3D U-net for initial super-resolution, then an INR (MLP with Fourier features) is fit per subject, balancing U-net output and low-resolution data via a combined MSE loss, enabling uncertainty quantification.

In practice

• Apply SIINR to enhance clinical dMRI data with thick slices.
• Use uncertainty maps to assess reliability of derived neuroimaging metrics.
• Adapt the supervised component for specific upsampling ratios or tasks.

Topics

• Diffusion MRI
• Super-resolution
• Implicit Neural Representations
• Uncertainty Quantification
• Deep Learning
• Clinical Neuroimaging

Code references

• tomhend/SIINR-public

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• SIINR: Structurally Informed Implicit Neural Representations for super-resolution with uncertainty quantification of clinical quality diffusion MRI datasets — SIINR enhances clinical dMRI super-resolution by fusing U-net priors with INRs, quantifying uncertainty for robust interpretation.
• Anatomy-Guided Residual Motion Diffusion for Controllable 4D Cardiac MRI Synthesis — A cascaded latent diffusion model synthesizes controllable, anatomically consistent 4D cardiac MRI, improving AI model robustness with limited data.
• A physics-informed foundation model for quantitative diffusion MRI — PIGMENT uses a physics-informed generative prior to enable reliable quantitative diffusion MRI from sparse, heterogeneous data.
• Benchmarking MRI Representations for Deep Learning-Based Focal Cortical Dysplasia Segmentation — Optimizing MRI representation significantly improves deep learning-based Focal Cortical Dysplasia segmentation.
• Contrast-Informed Augmentation and Domain-Adversarial Training for Adult-to-Neonatal MR Reconstruction Generalization — Contrast-informed augmentation and domain-adversarial training enhance deep learning MR reconstruction generalization from adult to neonatal data.
• PhyMRI-SR: Toward Physics-Aware MRI Image Super-Resolution — MRI super-resolution benefits from physics-aware reconstruction that optimizes resolution-SNR trade-offs, not just upscaling fixed inputs.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.