SRUG: A Fusion-Driven Generator Network for Medical Image Translation

Xihe Qiu; Yang Dai; Xiaoyu Tan; Sijia Li; Fenghao Sun; Lu Gan; Liang Liu

cs.CVJan 8, 2026

SRUG: A Fusion-Driven Generator Network for Medical Image Translation

Authors: Xihe Qiu, Yang Dai, Xiaoyu Tan, Sijia Li, Fenghao Sun, Lu Gan, Liang Liu

Organizations: School of Electronic and Electrical Engineering, Shanghai University of Engineering Science, Shanghai, China · Department of Thoracic Surgery, Zhongshan Hospital of Fudan University, Shanghai 200032, China · Department of Medical Oncology, Cancer Center, and Fudan Zhangjiang Institute, Zhongshan Hospital of Fudan University, Shanghai 200032, China · Clinical Research Unit, Institute of Clinical Science, Zhongshan Hospital of Fudan University, Shanghai 200032, China

Abstract

MRI sequence synthesis aims to recover missing image contrast while preserving patient-specific anatomy. The choice of generation mechanism affects both optimization and the way source information reaches the synthesized image. In this study, we propose SRUG, a supervised standalone fusion-driven generator that learns a deterministic source-to-target mapping for paired MRI synthesis. Its direct reconstruction formulation removes generator-discriminator competition and requires neither variational latent sampling nor iterative diffusion denoising. To support structural fidelity within this formulation, SRUG uses a residual encoder adapted for image reconstruction, with a full-resolution convolutional stem, pooling-separated feature stages, and channel recalibration embedded in the residual branches. A nested multi-scale decoder (NMD) repeatedly fuses the resulting spatial features and predicts the target sequence through a single output head. L1 loss and Multi-Scale Structural Loss (MSS loss) jointly supervise the prediction, connecting the direct generation objective to anatomical detail recovery. Experiments on BraTS 2023 show competitive reconstruction fidelity and structural consistency across three MRI translation tasks. Architecture ablations assess the effects of channel recalibration, NMD, and structural supervision within the same generator framework. Additional evaluation on IXI supports applicability to another dataset and modality pair, while zero-shot testing on BraTS 2019 provides preliminary evidence of cross-dataset transferability. These results support direct supervised generation as a practical approach to paired MRI sequence synthesis, with the adapted encoding and reconstruction pathway providing the basis for structural preservation. The SRUG implementation is publicly available at https://github.com/RisingRich/SRUG.

Figures & tables

Fig. 1 : Schematic motivation for SRUG. Panels (a)–(d) illustrate concerns associated with adversarial optimization, variational reconstruction, diffusion-based translation, and zero-shot transfer, respectively. Panel (e) presents direct source-to-target reconstruction using the CRRB-based encoder, NMD, and paired L1 and MSS loss supervision. The listed advantages express design aims; synthesis quality and transferability are evaluated in the experiments.

Fig. 2 : SRUG workflow: (a) 2.5D preprocessing; (b) the encoder built from channel-recalibrated residual blocks (CRRBs), reconstruction with the nested multi-scale decoder (NMD), and loss supervision; and (c) encoder stage details. MSS loss denotes Multi-Scale Structural Loss, computed from MS-SSIM.

Task	Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MS-SSIM $\uparrow$	MSE $\downarrow$	NMSE $\downarrow$
BraTS 2023 T1 $\rightarrow$ T2	CycleGAN	18.9567 $\pm$ 0.2842	0.6912 $\pm$ 0.0093	0.1467 $\pm$ 0.0058	0.6293 $\pm$ 0.0286	861.8345 $\pm$ 52.1906	0.4430 $\pm$ 0.0278
	NICE-GAN	23.8071 $\pm$ 0.2556	0.8486 $\pm$ 0.0069	0.1146 $\pm$ 0.0017	0.8795 $\pm$ 0.0043	290.8364 $\pm$ 18.5864	0.1558 $\pm$ 0.0116
	Pix2Pix	25.1078 $\pm$ 0.0320	0.8711 $\pm$ 0.0007	0.0885 $\pm$ 0.0006	0.9018 $\pm$ 0.0007	212.5699 $\pm$ 2.0190	0.1133 $\pm$ 0.0014
	ResViT	24.8802 $\pm$ 0.1528	0.8660 $\pm$ 0.0017	0.1419 $\pm$ 0.0006	0.8992 $\pm$ 0.0027	218.8124 $\pm$ 7.9094	0.1157 $\pm$ 0.0037
	BBDM	25.5541 $\pm$ 0.0594	0.8272 $\pm$ 0.0034	0.0702 $\pm$ 0.0002	0.9240 $\pm$ 0.0007	200.5621 $\pm$ 3.6000	0.1151 $\pm$ 0.0033
	Our	27.3122 $\pm$ 0.0641	0.9161 $\pm$ 0.0006	0.0519 $\pm$ 0.0007	0.9436 $\pm$ 0.0004	136.5155 $\pm$ 2.1574	0.0737 $\pm$ 0.0013

Table 1: Quantitative results across MRI translation settings. MSE is measured on [0,255].

Fig. 3 : Normalized radar comparison across seven MRI translation settings and six metrics.

Dataset	Task	Train patients	Test patients	Split strategy
BraTS 2023	T1 $\rightarrow$ T2 T1 $\rightarrow$ FLAIR T2 $\rightarrow$ FLAIR	237	56	patient-wise
IXI	PD $\rightarrow$ T2	462	116	patient-wise
BraTS 2019	T1 $\rightarrow$ T2 T1 $\rightarrow$ FLAIR T2 $\rightarrow$ FLAIR	–	335	external testing

Table 2: Patient-level train/test splits for MRI translation experiments.

Fig. 4 : Metric distributions across MRI translation settings.

Fig. 5 : Pixel-intensity correlation with target images; darker colors indicate stronger correlation.

Task	Metric	Strongest baseline	Baseline value	Our	$\Delta$ (Our-base)	Direction	Wilcoxon $p$	Bonferroni $p$
T1 $\rightarrow$ T2	PSNR $\uparrow$	BBDM	25.5541	27.3122	1.7581	Our higher	$8.459\times 10^{-11}$	$5.075\times 10^{-10}$
	SSIM $\uparrow$	Pix2Pix	0.8711	0.9161	0.0450	Our higher	$3.774\times 10^{-11}$	$2.264\times 10^{-10}$
	LPIPS $\downarrow$	BBDM	0.0702	0.0519	-0.0183	Our lower	$3.774\times 10^{-11}$	$2.264\times 10^{-10}$
	MS-SSIM $\uparrow$	BBDM	0.9240	0.9436	0.0196	Our higher	$3.984\times 10^{-11}$	$2.391\times 10^{-10}$
	MSE $\downarrow$	BBDM	200.5621	136.5155	-64.0466	Our lower	$3.984\times 10^{-11}$	$2.391\times 10^{-10}$
	NMSE $\downarrow$	Pix2Pix	0.1133	0.0737	-0.0396	Our lower	$3.984\times 10^{-11}$	$2.391\times 10^{-10}$

Table 3: Wilcoxon significance tests for BraTS 2023 translation tasks.

Panel of Figure 9 — Fig. 6 : T1 $\rightarrow$ T2 ablation of channel recalibration, NMD, and MSS loss in SRUG: (a) normalized grouped bars; (b) normalized heatmap; (c) component-removal profiles. Configurations are defined in Table 4 .

Panel of Figure 10 — Fig. 7 : Full-image error heatmaps; brighter regions indicate larger errors.

Fig. 8 : ROI-level error heatmaps in selected anatomical regions; brighter regions indicate larger errors.

Generator Configuration				Image Quality Assessment
Encoder w/o SE	CRRB encoder	Plain decoder	NMD	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MS-SSIM $\uparrow$	MSE $\downarrow$	NMSE $\downarrow$
Replacement experiments: module substitution under the same generator pipeline.
$\checkmark$		$\checkmark$		26.9974	0.9050	0.0624	0.9402	142.7445	0.0769
	$\checkmark$	$\checkmark$		27.1447	0.9020	0.0733	0.9411	136.3203	0.0736
$\checkmark$			$\checkmark$	27.1491	0.9131	0.0533	0.9410	140.7279	0.0766
	$\checkmark$		$\checkmark$	27.3122	0.9161	0.0519	0.9436	136.5155	0.0737

Table 4: T1

\rightarrow

T2 ablation of channel recalibration, NMD, and MSS loss in SRUG. Upper block: encoder–decoder variants; lower block: component removal. MSE is measured on [0,255].

Fig. 9 : Effect of channel recalibration on encoder feature maps at levels $x_{1,0}$ – $x_{4,0}$ . Top: encoder variant without channel recalibration. Bottom: encoder built from CRRBs.

Fig. 10 : Effect of encoder channel recalibration on NMD feature maps at nodes $x_{0,1}$ – $x_{0,4}$ . Top: features from the encoder variant without channel recalibration. Bottom: features from the encoder built from CRRBs.

Explore similar work

Jun 23, 2026

Prob-BBDM: a Probabilistic Brownian Bridge Diffusion Model for MRI sequence image-to-image translation

AI-driven image-to-image synthesis is rapidly advancing, with growing applications in medical imaging. Multi-modal image analysis plays a crucial role in optimizing examination quality, yet acquiring multiple imaging modalities in clinical settings remains resource-intensive and time-consuming, especially for 3D imaging. To address this challenge, we propose a novel image-to-image translation model based on Brownian Bridge Diffusion Models (BBDM), which synthesizes magnetic resonance imaging (MRI) sequences from 2D axial slices. Our approach integrates a variational encoder-guided diffusion mechanism, leveraging probabilistic image distributions to enhance synthesis quality. Evaluated on the BraTS 2021 dataset, our Probabilistic-BBDM (Prob-BBDM) achieves superior performance across multiple translation tasks, reaching up to 88.46% SSIM and 26.09 dB PSNR, with consistent improvements over baselines. Notably, our diffusion process requires only 4 steps, making it computationally efficient while maintaining high-quality synthesis. To further validate generalizability, we test Prob-BBDM on an external third-party dataset, demonstrating consistent performance across domains. Additionally, we assess the clinical utility of the synthesized slices by using them as input to a pre-trained segmentation model. Tumor segmentation yields a Dice score of 88.71% and an HD95 of 3.49 mm, confirming that the synthesized slices preserve critical diagnostic information. These results highlight the potential of Prob-BBDM for high-quality, efficient, and generalizable MRI synthesis, offering a promising step toward improved medical image translation.

Martin Valls, Pascal Bourdon, Christine Fernandez-Maloigne +2

UFR SFA · University of Poitiers, CNRS, XLIM, France · I3M common laboratory CNRS-Siemens Healthinners, Poitiers University Hospital and University of Poitiers, France +4

Image-to-Image Translation Magnetic Resonance Imaging Medical Image Synthesis Diffusion Models

Ask PDF

Mar 13, 2026

Multiscale Structure-Guided Latent Diffusion for Multimodal MRI Translation

Although diffusion models have achieved remarkable progress in multi-modal magnetic resonance imaging (MRI) translation tasks, existing methods still tend to suffer from anatomical inconsistencies or degraded texture details when handling arbitrary missing-modality scenarios. To address these issues, we propose a latent diffusion-based multi-modal MRI translation framework, termed MSG-LDM. By leveraging the available modalities, the proposed method infers complete structural information, which preserves reliable boundary details. Specifically, we introduce a style--structure disentanglement mechanism in the latent space, which explicitly separates modality-specific style features from shared structural representations, and jointly models low-frequency anatomical layouts and high-frequency boundary details in a multi-scale feature space. During the structure disentanglement stage, high-frequency structural information is explicitly incorporated to enhance feature representations, guiding the model to focus on fine-grained structural cues while learning modality-invariant low-frequency anatomical representations. Furthermore, to reduce interference from modality-specific styles and improve the stability of structure representations, we design a style consistency loss and a structure-aware loss. Extensive experiments on the BraTS2020 and WMH datasets demonstrate that the proposed method outperforms existing MRI synthesis approaches, particularly in reconstructing complete structures. The source code is publicly available at https://github.com/ziyi-start/MSG-LDM.

Jianqiang Lin, Zhiqiang Shen, Peng Cao +3

Northeastern University, Shenyang, China · Key Laboratory of Intelligent Computing in Medical Image, Shenyang, China · Computer Science and Engineering, Northeastern University, Shenyang, China +5

Medical Image Synthesis Latent Diffusion Models Multimodal Disentangled Representation Learning

Ask PDF

Aug 5, 2026

VoxStruct3D: Structure-Leading Flow Matching for Voxel-Space 3D MRI Synthesis

High-fidelity 3D MRI synthesis requires both globally coherent anatomy and fine-grained voxel-level detail. Although latent diffusion makes volumetric generation tractable, its image autoencoder introduces a reconstruction bottleneck that can limit the fine detail recoverable in the final volume. We present VoxStruct3D, a voxel-space flow-matching framework that directly models full-resolution MRI volumes using a clean-data prediction objective. Its Volumetric Voxel Generator (VVG) combines factorized 3D patch embedding with overlapping upsampling, time-modulated residual refinement, and skip fusion, enabling neighboring tokens to jointly reconstruct shared voxel regions and suppress patch-boundary artifacts. To complement direct voxel-space modeling with an explicit anatomical prior, we further introduce a Structure-First, Image-Follows (SFIF) strategy. A frozen pretrained 3D medical encoder and a StructVAE extract compact structure tokens that preserve dominant anatomy, while a structure-leading schedule keeps their trajectory ahead of the image trajectory. Patch-Aligned RoPE spatially aligns the unequal token grids, and asymmetric attention enforces one-way guidance from structure to image. Experiments on pathological and healthy T1-weighted brain MRI datasets show that VoxStruct3D achieves the strongest overall performance across feature-distribution alignment, sample diversity, and perceptual quality, producing anatomically coherent and visually realistic volumes.

Fang Li, Yang Gao, Shihao Zou +5

State Key Laboratory of Virtual Reality Technology and Systems, Beihang University · Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Faculty of Computer Science and Control Engineering, Shenzhen University of Advanced Technology

Flow Matching 3D Medical Imaging Medical Image Synthesis

Ask PDF