MedGSSR: Generalizable Medical Image Super-Resolution 3D Reconstruction via Hierarchical Feed-forward Gaussian Splatting

Chengkai Wang^1*, Luoyu Hong^2*, Yiting Zhao², Jiamin Wang³, Xiang Feng³, Feiwei Qin², Zhenzhong Kuang², Xuefei Yin⁴, Ali Bashashati^1†, Yanming Zhu^4†

¹University of British Columbia ²Hangzhou Dianzi University ³ShanghaiTech University ⁴Griffith University

* Equal contribution † Corresponding author

ECCV 2026

Paper Supp Code BibTeX

MedGSSR teaser showing the feed-forward Gaussian field paradigm and 2x/4x super-resolution results

MedGSSR introduces a paradigm shift for Medical 3D Super-Resolution. Left: Existing self-supervised methods (relying on isolated optimization) and neural implicit networks (relying on coordinate querying) fail to simultaneously achieve fast inference, robust generalization, and high-quality reconstruction. MedGSSR introduces a generalized feed-forward 3DGS paradigm that explicitly satisfies all three clinical requirements. Right: Cross-dataset visual comparisons (trained on MSD, tested on HCP) at 2× and 4× upsampling scales. Compared to baselines that produce geometric artifacts or over-smoothed textures, MedGSSR accurately restores high-frequency anatomical structures with the lowest error (dark blue in difference maps) while accelerating inference by an order of magnitude.

Abstract

High-resolution volumetric medical imaging is critical for clinical diagnosis, yet acquisition is often limited by scanner hardware, scan time, and for CT, radiation dose. Medical 3D Super-Resolution (Med3DSR) offers a computational alternative, but existing methods commonly rely on per-subject optimization, pretrained priors, or coordinate-based implicit representations, which compromise anatomical fidelity and limit efficiency.

We present MedGSSR, a fully end-to-end feed-forward framework that represents volumes as an explicit 3D Gaussian field for Med3DSR. Unlike coordinate-based implicit functions, our explicit 3D Gaussian representation naturally enhances signal continuity and local high-frequency fidelity. MedGSSR decouples reconstruction into coarse-grained structural preservation and fine-grained textural refinement through a Pyramid Anatomical Encoder and a Hierarchical Gaussian Projector.

To support arbitrary-scale super-resolution, MedGSSR introduces sub-voxel Gaussian decomposition and a Differentiable Gaussian Voxelizer that directly queries the continuous 3D intensity field, reducing discretization artifacts. Extensive experiments on MRI and CT benchmarks show strong reconstruction quality, robust cross-dataset generalization, and improved downstream clinical segmentation.

Method Overview

Given a low-resolution medical volume, MedGSSR first uses a Pyramid Anatomical Encoder to extract complementary coarse and fine volumetric features. A Hierarchical Gaussian Projector then predicts explicit structural and textural 3D Gaussian primitives. Finally, a Differentiable Gaussian Voxelizer renders the continuous Gaussian field onto any target high-resolution grid, enabling arbitrary-scale 3D super-resolution without per-subject optimization.

Contributions

We reformulate medical 3D super-resolution as a generalized feed-forward mapping from sparse LR voxel grids to an explicit continuous 3D Gaussian field.
We propose a Hierarchical Gaussian Projector with sub-voxel decomposition, explicitly decoupling structural preservation from textural refinement for heterogeneous medical images.
We introduce a Differentiable Gaussian Voxelizer that supports arbitrary-scale volumetric rendering directly in 3D.
Extensive MRI and CT experiments demonstrate state-of-the-art intra-domain performance, robust cross-domain generalization, and downstream clinical utility.

Main Results: In-Domain 3D Super-Resolution

We first evaluate reconstruction fidelity within the primary MRI and CT domains. MedGSSR consistently improves high-frequency anatomical reconstruction across 2×, 3×, and 4× scales.

In-domain qualitative comparison on MELA (CT) and MSD (MRI).

MSD (MRI)

Paradigm	Method	2×			3×			4×
Paradigm	Method	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓
Traditional	Trilinear	33.31	0.9670	0.1134	30.77	0.9377	0.2026	29.08	0.9100	0.2558
Traditional	Cubic	33.99	0.9725	0.0981	31.13	0.9422	0.2083	29.27	0.9118	0.2711
Self-supervised	CuNeRF	32.60	0.9715	0.0513	30.07	0.9524	0.0878	28.26	0.9309	0.1271
Neural Implicit	ArSSR	32.87	0.9742	0.0581	29.56	0.9442	0.0730	28.51	0.9315	0.0853
FF-3DGS	MedGSSR	35.91	0.9821	0.0337	33.97	0.9700	0.0706	32.10	0.9541	0.1246

MELA (CT)

Paradigm	Method	2×			3×			4×
Paradigm	Method	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓
Traditional	Trilinear	37.51	0.9530	0.1561	34.83	0.9166	0.2302	33.84	0.8843	0.3291
Traditional	Cubic	37.64	0.9636	0.1179	35.14	0.9237	0.2325	34.09	0.8888	0.3263
Self-supervised	CuNeRF	37.12	0.9654	0.1201	35.42	0.9235	0.1863	33.94	0.8957	0.2456
Self-supervised	NAB-GS	-	-	-	-	-	-	34.13	0.9518	-
Neural Implicit	ArSSR	38.53	0.9644	0.0914	36.02	0.9413	0.1576	34.63	0.9244	0.2261
FF-3DGS	MedGSSR	42.08	0.9738	0.0705	39.95	0.9558	0.1193	37.31	0.9362	0.1472

Note: For NAB-GS, due to the unavailability of its open-source code, we directly report the 4× performance quoted from their original paper.

Generalization Results: Cross-Domain 3D Super-Resolution

We evaluate models trained on MSD directly on the unseen HCP dataset, without fine-tuning. MedGSSR maintains strong reconstruction fidelity under MRI domain shift.

Cross-domain qualitative comparison on HCP

Cross-domain HCP (MRI) results. Models are trained on MSD and evaluated on HCP.

HCP (MRI): MSD → HCP

Paradigm	Method	2×			3×			4×
Paradigm	Method	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓
Traditional	Trilinear	36.45	0.9425	0.1048	33.75	0.9221	0.1634	30.76	0.9026	0.2675
Traditional	Cubic	36.47	0.9588	0.0885	34.01	0.9289	0.1454	31.86	0.9125	0.2483
Self-supervised	CuNeRF	36.91	0.9716	0.0724	34.48	0.9547	0.1362	33.12	0.9382	0.1587
Neural Implicit	ArSSR	36.07	0.9655	0.0615	33.56	0.9429	0.0892	31.87	0.9220	0.1141
FF-3DGS	MedGSSR	40.81	0.9849	0.0674	37.94	0.9702	0.0945	35.84	0.9533	0.1253

Extended HCP 4× Cross-Domain Comparison

Method	PSNR ↑	SSIM ↑	LPIPS ↓	FLOPs (G)	Peak Mem. (GB)	Params (M)	Inference Time (ms)
MTVNet	34.16	0.9246	0.1368	1359.50	1.75	22.61	1083.18
SuperFormer	33.87	0.9142	0.1446	1888.20	2.06	19.66	1592.91
CuNeRF	33.12	0.9382	0.1587	2061.58	0.12	0.98	259.76
ArSSR	31.87	0.9220	0.1141	955.15	8.15	6.28	105.98
MedGSSR (Ours)	35.84	0.9533	0.1253	385.87	1.29	18.13	56.52

Clinical Validation

We evaluate whether higher-quality super-resolved MRI volumes improve downstream brain tissue segmentation under the challenging 8× cross-domain setting.

Downstream brain tissue segmentation on HCP under the more challenging 8× cross-domain setting.

Ablation Study

We study the role of structure-texture decoupling, sub-voxel count, Gaussian support, frequency supervision, and reduced training data.

Sub-voxel Count

Setting	PSNR ↑	SSIM ↑	LPIPS ↓
m = 1	35.98	0.9760	0.0423
m = 2	36.06	0.9771	0.0388
m = 4 (Default)	36.60	0.9870	0.0363
m = 8	36.52	0.9862	0.0365

HCP Cross-Domain Ablations

Variant	PSNR ↑	SSIM ↑	LPIPS ↓
Ours	35.84	0.9533	0.1253
Ours + FFT Loss	35.72	0.9472	0.1078
Ours w/ 1σ truncation	31.05	0.9142	0.1608
Ours w/ 50% training data	34.26	0.9405	0.1452
Ours w/ 75% training data	35.49	0.9512	0.1283

BibTeX

@inproceedings{wang2026medgssr,
  title     = {MedGSSR: Generalizable Medical Image Super-Resolution 3D Reconstruction via Hierarchical Feed-forward Gaussian Splatting},
  author    = {Wang, Chengkai and Hong, Luoyu and Zhao, Yiting and Wang, Jiamin and Feng, Xiang and Qin, Feiwei and Kuang, Zhenzhong and Yin, Xuefei and Bashashati, Ali and Zhu, Yanming},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}