Single-Subject Multi-View MRI Super-Resolution via Implicit Neural Representations

Heejong Kim; Abhishek Thanki; Roel van Herten; Daniel Margolis; Mert R Sabuncu

eess.IVMar 23, 2026

Single-Subject Multi-View MRI Super-Resolution via Implicit Neural Representations

Authors: Heejong Kim, Abhishek Thanki, Roel van Herten, Daniel Margolis, Mert R Sabuncu

Organizations: Department of Radiology, Weill Cornell Medicine, New York, USA · Weill Cornell Graduate School of Medical Sciences, New York, USA · School of Electrical and Computer Engineering, Cornell University and Cornell Tech, New York, USA

Abstract

Clinical MRI frequently acquires anisotropic volumes with high in-plane resolution and low through-plane resolution to reduce acquisition time. Multiple orientations are therefore acquired to provide complementary anatomical information. Conventional integration of these views relies on registration followed by interpolation, which can degrade fine structural details. Recent deep learning-based super-resolution (SR) approaches have demonstrated strong performance in enhancing single-view images. However, their clinical reliability is often limited by the need for large-scale training datasets, resulting in increased dependence on cohort-level priors. Self-supervised strategies offer an alternative by learning directly from the target scans. Prior work either neglects the existence of multi-view information or assumes that in-plane information can supervise through-plane reconstruction under the assumption of pre-alignment between images. However, this assumption is rarely satisfied in clinical settings. In this work, we introduce Single-Subject Implicit Multi-View Super-Resolution for MRI (SIMS-MRI), a framework that operates solely on anisotropic multi-view scans from a single patient without requiring pre- or post-processing. Our method combines a multi-resolution hash-encoded implicit representation with learned inter-view alignment to generate a spatially consistent isotropic reconstruction. We validate the SIMS-MRI pipeline on both simulated brain and clinical prostate MRI datasets. Code will be made publicly available for reproducibility: https://github.com/abhshkt/SIMS-MRI

Figures & tables

Figure 1: Overview of the proposed three-phase coordinate-based multi-view MRI fusion framework. Phase 1: A continuous implicit image representation model with multi-resolution hash encoding, $f_{\theta}$ , is trained on a reference (axial) view to map 3D world coordinates to voxel intensities. Phase 2: A SIREN-based registration network [ 30 ] $g_{\phi}$ learns a dense displacement field that aligns the second (coronal) view to the reference coordinate space while keeping $f_{\theta}$ fixed. Phase 3: The implicit image representation is fine-tuned using both the reference and registered views for information fusion. Inference : The learned continuous representation is queried on an isotropic 3D grid to generate a super-resolved volume, enabling reconstruction of unseen slice orientations.

Figure 2: Qualitative results for the unseen view in simulated two-view brain MRI. The first row presents a comparison across methods, and the second row shows a zoomed-in view of the highlighted region. Yellow and green arrows indicate areas where SIMS-MRI achieves superior structural detail reconstruction.

Method	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	MONAI $\downarrow$
Interpolation (B-spline)	$0.0551$	$0.6723$	$18.8914$	$0.2842$	$0.1960$	$6.17\text{\times}{10}^{-4}$
Interpolation (Linear)	$0.0541$	$0.6779$	$19.0907$	$0.2865$	$0.1991$	$7.13\text{\times}{10}^{-4}$
ECLARE-axial	$0.0435$	$0.7616$	$20.5620$	$0.2687$	$0.2040$	$3.60\text{\times}{10}^{-4}$
ECLARE-fused	$0.0412$	$0.7786$	$21.2464$	$0.2464$	0.1802	$2.91\text{\times}{10}^{-4}$
SMORE-axial	$0.0431$	$0.7633$	$20.5709$	$0.2700$	$0.2037$	$3.57\text{\times}{10}^{-4}$
SMORE-fused	0.0409	0.7816	21.2720	0.2445	0.1787	$3.10\text{\times}{10}^{-4}$

Table 1: Quantitative comparison on simulated two-view brain MRI (unseen view evaluation). ’-axial’ denotes reconstruction using the axial view only, while ’-fused’ indicates co-registered and averaged axial/coronal volumes. (

\uparrow

: higher is better and

\downarrow

: lower is better. Best in bold and second-best in underline .)

Method	MAE $\downarrow$	SSIM $\uparrow$	PSNR $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	MONAI $\downarrow$
SIMS-MRI	0.0247	0.8877	24.6369	0.2362	0.1845	$2.18\text{\times}{10}^{-4}$
w/o FreeNeRF & w/ HE	0.0281	0.8461	23.6493	0.2254	0.1744	$1.64\text{\times}{10}^{-4}$
w/o FreeNeRF & w/ FM	$0.0327$	$0.8077$	$23.3116$	$0.3577$	$0.2345$	$5.31\text{\times}{10}^{-4}$

Table 2: Ablation study on simulated two-view brain MRI. HE: Hash Encoding; FM: Fourier Mapping. Bold and underlined values indicate the best and second-best results.

Figure 3: Results for prostate MRI. The violin plots show the MONAI metric (MONAI: Medical Open Network for Artificial Intelligence model-based perceptual metric). Each row corresponds to a representative subject, selected from slices near the median of the distribution shown in the violin plots (color-coded and marked in the plots).

Explore similar work

Prior-Guided Implicit Neural Representations for Single-Subject Diffusion MRI Super-Resolution
Sep 1, 2026Abdulkader Ghandoura, Marsil Zakour, William Consagra +1Implicit Neural Representations Medical Image Super-Resolution
Psychiatry Neuroimaging Laboratory Brigham and Women’s Hospital Harvard Medical School Boston, MA, USA · School of Computation, Information and Technology Technical University of Munich Munich, Germany · Department of Statistics University of South Carolina Columbia, SC, USA
Physics-Driven 3D Gaussian Rendering for Zero-Shot MRI Super-Resolution
Mar 10, 2026Shuting Liu, Lei Zhang, Wei Huang +2Medical Image Super-Resolution Magnetic Resonance Imaging
School of Artificial Intelligence, Sichuan University, Chengdu 610065, China · Institute of High Performance Computing, A*STAR, Singapore 138632
Decoupling Multi-Contrast Super-Resolution: Self-Supervised Implicit Re-Representation for Unpaired Cross-Modal Synthesis
May 9, 2025Yinzhe Wu, Hongyu Rui, Fanwen Wang +5Medical Image Super-Resolution Unpaired Image-to-Image Translation
Department of Bioengineering and I-X, Imperial College London, London SW7 2AZ, United Kingdom · Cardiovascular Research Centre, Royal Brompton Hospital, London, United Kingdom · Bioengineering Department and Imperial-X, Imperial College London, London W12 7SL, U.K. +3

Sep 1, 2026J/K move · Enter open · S save

Prior-Guided Implicit Neural Representations for Single-Subject Diffusion MRI Super-Resolution

Abdulkader Ghandoura, Marsil Zakour, William Consagra, Yogesh Rathi

Psychiatry Neuroimaging Laboratory Brigham and Women’s Hospital Harvard Medical School Boston, MA, USA · School of Computation, Information and Technology Technical University of Munich Munich, Germany · Department of Statistics University of South Carolina Columbia, SC, USA

Implicit Neural Representations Medical Image Super-Resolution Transfer Learning Diffusion MRI

Open paper AskPDF arXiv

Resolving complex fiber geometries in brain white matter requires high-resolution diffusion MRI at the cost of long acquisition times. This leads many clinical protocols to opt for low-resolution scans, making downstream microstructure estimation and tractography challenging. Implicit neural representations (INRs) can model the diffusion signal continuously, enabling native single-subject super-resolution by querying the network at arbitrary spatial coordinates, yet existing methods often suffer from long training times and lack a mechanism to incorporate anatomical priors to regularize super-resolution by constraining the space of plausible reconstructions. To address these limitations, we propose a novel transfer-learning framework that pre-trains an INR on a high-resolution template and then adapts it to subject-specific scans via registration and fine-tuning. For

4\times

through-plane super-resolution from 5 mm to 1.25 mm on Human Connectome Project (HCP) data, our method reduces NRMSE by 36-49% and increases FSIM by 24-43% over a recent baseline with

6\times

faster training, outperforming competing INR-based methods across both image quality and domain-specific metrics. Code is available on the project page at https://abdulkaderghandoura.github.io/research/msc-thesis/ .

Abstract

Explore similar work

Prior-Guided Implicit Neural Representations for Single-Subject Diffusion MRI Super-Resolution

Physics-Driven 3D Gaussian Rendering for Zero-Shot MRI Super-Resolution

Decoupling Multi-Contrast Super-Resolution: Self-Supervised Implicit Re-Representation for Unpaired Cross-Modal Synthesis