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

Chengkai Wang; Luoyu Hong; Yiting Zhao; Jiamin Wang; Xiang Feng; Feiwei Qin; Zhenzhong Kuang; Xuefei Yin

Figure 1: 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\times

and

4\times

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.

Figure 2: Overview of MedGSSR. Given an LR input volume, the Pyramid Anatomical Encoder (PAE) built on a 3D U-Net extracts a coarse feature map and a

2\times

upsampled fine feature map via input-residual connection and PixelShuffle. The Hierarchical Gaussian Projector (HGP) performs sub-voxel Gaussian decomposition by predicting multiple Gaussian primitives per voxel with lightweight MLP heads, producing structural Gaussians from the coarse stream and textural Gaussians from the fine stream. These primitives are then aggregated by the Differentiable Gaussian Voxelizer (DGV) through differentiable rasterization to form a continuous volumetric intensity field, from which the HR volume is rendered.

Paradigm	Methods	$2\times$			$3\times$			$4\times$
Paradigm	Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
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 [ 6 ]	32.60	0.9715	0.0513	30.07	0.9524	0.0878	28.26	0.9309	0.1271
Neural Implicit	ArSSR [ 29 ]	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

Table 1: Comparison of intra-domain 3D super-resolution performance on the MSD (MRI) dataset. Models are both trained and evaluated on the MSD dataset.

Paradigms	Methods	$2\times$			$3\times$			$4\times$
Paradigms	Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
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 [ 6 ]	37.12	0.9654	0.1201	35.42	0.9235	0.1863	33.94	0.8957	0.2456
Self-supervised	NAB-GS [ 20 ]	-	-	-	-	-	-	34.13	0.9518	-
Neural Implicit	ArSSR [ 29 ]	38.53	0.9644	0.0914	36.02	0.9413	0.1576	34.63	0.9244	0.2261

Table 2: Comparison of intra-domain 3D super-resolution performance on the MELA (CT) dataset. Models are both trained and evaluated on the MELA dataset.

Figure 3: Qualitative comparison of intra-domain 3DSR on MELA (CT) and MSD (MRI). Axial views are shown at

2\times

,

3\times

and

4\times

scales. MedGSSR better recovers fine-grained skeletal and vascular structures, avoiding the smoothing artifacts of ArSSR and the noise commonly observed in other methods.

Paradigms	Methods	$2\times$			$3\times$			$4\times$
Paradigms	Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
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 [ 6 ]	36.91	0.9716	0.0724	34.48	0.9547	0.1362	33.12	0.9382	0.1587
Neural Implicit	ArSSR [ 29 ]	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

Table 3: Comparison of cross-domain 3D super-resolution performance on the HCP (MRI). Models are trained on the MSD dataset and evaluated on the HCP dataset.

Paradigms	Methods	$4\times$			$6\times$			$8\times$
Paradigms	Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Traditional	Trilinear	31.96	0.9404	0.1051	26.73	0.9027	0.1706	23.96	0.8436	0.2605
Traditional	Cubic	32.88	0.9407	0.0898	28.51	0.9189	0.1654	26.07	0.8579	0.2517
Self-supervised	CuNeRF [ 6 ]	38.03	0.9628	0.1252	33.76	0.9164	0.2465	30.50	0.8690	0.3510
Neural Implicit	ArSSR [ 29 ]	35.92	0.9471	0.0918	32.88	0.9332	0.1191	29.38	0.8994	0.2145
FF-3DGS	MedGSSR	41.20	0.9818	0.0468	37.74	0.9698	0.0765	35.24	0.9233	0.1253

Table 4: Comparison of cross-domain 3D super-resolution performance on the UHRCT (CT). Models are trained on the MELA dataset and evaluated on the UHRCT dataset.

Figure 4: Qualitative comparison of cross-domain 3DSR on the HCP (MRI). All models are trained on the MSD and evaluated on the unseen HCP. The error maps (dark blue indicates lower error) highlight that MedGSSR produces the sharpest anatomical details with minimal reconstruction deviation compared to the baselines.

Methods	CSF	GM	WM	dGM	BS+CB	Average
Trilinear	0.4937	0.7265	0.8617	0.8098	0.6701	0.7124
Cubic	0.5410	0.7616	0.8860	0.8616	0.8518	0.7804
CuNeRF [ 6 ]	0.3668	0.5330	0.7347	0.1285	0.0458	0.3618
ArSSR [ 29 ]	0.6782	0.8318	0.9187	0.9140	0.5612	0.7908
Ours (MedGSSR)	0.5779	0.8406	0.9295	0.9407	0.8006	0.8178

Table 5: Comparison of Dice scores. The segmentation is performed using a pre-trained nnUNet on

4\times

super-resolved MRI volumes (HCP). Best results are highlighted in bold.

Figure 5: Downstream brain tissue segmentation on the HCP dataset (

4\times

SR). The structural fidelity provided by MedGSSR translates to precise automated boundary delineation, avoiding the tissue-bleeding artifacts observed in the others.

Figure 6: Ablation on Structure-Texture Decoupling on the HCP dataset (

4\times

SR). The coarse branch captures global anatomy, and adding the fine branch restores high-frequency textures, improving sharpness and quantitative metrics.

Settings	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
$m=1$	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

Table 6: Ablation on sub-voxel count

m

under intra-domain setting on the HCP dataset (4

\times

SR). Setting

m=4

delivers the best balance between accuracy and computational cost. Best results are highlighted in bold.

Variants	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours	$\mathbf{35.84}$	$\mathbf{0.9533}$	${0.1253}$
Ours $+$ FFT Loss	${35.72}$	$0.9472$	$\mathbf{0.1078}$
Ours w/ $1\sigma$ truncation	$31.05$	$0.9142$	$0.1608$
Ours w/ 50% training data	$34.26$	$0.9405$	$0.1452$
Ours w/ 75% training data	$35.49$	${0.9512}$	$0.1283$

Table 7: Ablation study on effects of frequency-domain supervision, Gaussian truncation radius, and training data under the cross-domain setting on the HCP dataset (

4\times

SR). Best results are highlighted in bold.

Figure 7: Additional visual comparisons for intra-domain 3D super-resolution on the MSD (MRI) dataset. Results are shown across axial, coronal, and sagittal views at 2

\times

, 3

\times

, and 4

\times

upsampling scales, with zoomed-in regions for detailed inspection. From left to right: Trilinear, Cubic, CuNeRF, ArSSR, Ours, and GT. MedGSSR better preserves cortical folds, tissue boundaries, and local anatomical details than competing methods.

Figure 8: Additional visual comparisons for intra-domain 3D super-resolution on the MELA (CT) dataset. Results are shown across axial, coronal, and sagittal views at 2

\times

, 3

\times

, and 4

\times

upsampling scales, with zoomed-in regions for detailed inspection. From left to right: Trilinear, Cubic, CuNeRF, ArSSR, Ours, and GT. Compared with the baselines, MedGSSR produces clearer boundaries and more faithful recovery of fine anatomical structures across all super-resolution.

Figure 9: Additional visual comparisons for cross-domain 3D super-resolution on the unseen HCP (MRI) dataset [ 26 ] . Models are trained exclusively on the MSD dataset and evaluated at 2

\times

, 3

\times

, and 4

\times

upsampling scales. Results are shown across axial, coronal, and sagittal views, with zoomed-in regions for detailed inspection. Compared with competing methods, MedGSSR better preserves tissue boundaries, cortical structures, and local anatomical details under dataset shift.

Figure 10: Additional visual comparisons for cross-dataset 3D super-resolution on the unseen UHRCT (CT) dataset [ 7 ] . Models are trained on MELA and evaluated at 4

\times

, 6

\times

, and 8

\times

upsampling scales. Following the benchmark protocol, results are shown on the axial plane, with zoomed-in regions for detailed inspection. MedGSSR remains more robust than competing methods as the resolution gap increases, preserving clearer anatomical boundaries and finer local structures.

Methods	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FLOPs (G)	Peak Mem. (GB)	Params (M)	Inference Time (ms)
MTVNet	$\underline{34.16{\pm}2.86}$	$0.9246{\pm}0.0267$	$0.1368{\pm}0.0416$	$1359.50$	$1.75$	$22.61$	$1083.18$
SuperFormer	$33.87{\pm}3.02$	$0.9142{\pm}0.0273$	$0.1446{\pm}0.0435$	$1888.20$	$2.06$	$19.66$	$1592.91$
CuNeRF	$33.12{\pm}3.56$	$\underline{0.9382{\pm}0.0365}$	$0.1587{\pm}0.0557$	$2061.58$	$\mathbf{0.12}$	$\mathbf{0.98}$	$259.76$
ArSSR	$31.87{\pm}3.37$	$0.9220{\pm}0.0324$	$\mathbf{0.1141{\pm}0.0513}$	$\underline{955.15}$	$8.15$	$\underline{6.28}$	$\underline{105.98}$
MedGSSR(Ours)	$\mathbf{35.84{\pm}2.73}$	$\mathbf{0.9533{\pm}0.0224}$	$\underline{0.1253{\pm}0.0378}$	$\mathbf{385.87}$	$\underline{1.29}$	$18.13$	$\mathbf{56.52}$

Table 8: Extended quantitative and computational comparison under cross-domain

4\times

3D SR on the HCP dataset. In addition to the baselines reported in the main paper, MTVNet and SuperFormer are included as transformer-based baselines. Computational cost is reported for all methods. Best and second-best results are bolded and underlined, respectively.

Figure 11: Additional visual comparisons for downstream brain tissue segmentation under cross-domain on the unseen HCP dataset based on 4

\times

super-resolved MRI volumes. Results are shown across axial, coronal, and sagittal views, with zoomed-in regions for detailed inspection. Compared with competing methods, MedGSSR yields segmentation masks that are more consistent with the ground truth, with clearer boundaries and more faithful anatomical structures.

Figure 12: Additional visual comparisons for downstream brain tissue segmentation under cross-domain on the unseen HCP dataset based on 8

\times

super-resolved MRI volumes. Results are shown across axial, coronal, and sagittal views. This more challenging setting serves as an additional stress test beyond the main experiments. As the upsampling factor increases, the competing methods exhibit substantial degradation in structural consistency, whereas MedGSSR still preserves more faithful anatomical regions and produces segmentation masks that remain noticeably closer to the ground truth.

Variants	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Ours	$\mathbf{37.31}$	$\mathbf{0.9362}$	$\mathbf{0.1472}$
Ours w/ $m=1$	$36.42$	$0.9234$	$0.1564$
Ours w/ $1\sigma$ truncation	$33.47$	$0.8971$	$0.1827$
Ours w/ 75% training data	$37.04$	$0.9338$	$0.1502$

Table 9: Additional CT ablation on the MELA dataset under the intra-domain

4\times

3D SR setting. The results show trends consistent with the MRI ablations and evaluate sub-voxel decomposition, Gaussian truncation radius, and reduced training data. Best results are highlighted in bold.

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

Abstract

Figures & tables

Explore similar work

Structure-Adaptive Sparse Diffusion in Voxel Space for 3D Medical Image Enhancement

MedSR-Vision: Deep Learning Framework for Multi-Domain Medical Image Super-Resolution

Pixel-Level Residual Diffusion Transformer: Scalable 3D CT Volume Generation