FOSCU: Feasibility of Synthetic MRI Generation via Duo-Diffusion Models for Enhancement of 3D U-Nets in Hepatic Segmentation

Youngung Han; Kyeonghun Kim; Seoyoung Ju; Yeonju Jean; Minkyung Cha; Seohyoung Park; Hyeonseok Jung; Nam-Joon Kim

cs.CVMar 31, 2026

FOSCU: Feasibility of Synthetic MRI Generation via Duo-Diffusion Models for Enhancement of 3D U-Nets in Hepatic Segmentation

Authors: Youngung Han, Kyeonghun Kim, Seoyoung Ju, Yeonju Jean, Minkyung Cha, Seohyoung Park, Hyeonseok Jung, Nam-Joon Kim, +3 more

Organizations: Seoul National University · OUTTA · Sangmyung University · Ewha Womans University · Chung-Ang University · Samsung Medical Center, Sungkyunkwan University School of Medicine · NVIDIA

Abstract

Medical image segmentation faces fundamental challenges including restricted access, costly annotation, and data shortage to clinical datasets through Picture Archiving and Communication Systems (PACS). These systemic barriers significantly impede the development of robust segmentation algorithms. To address these challenges, we propose FOSCU, which integrates Duo-Diffusion, a 3D latent diffusion model with ControlNet that simultaneously generates high-resolution, anatomically realistic synthetic MRI volumes and corresponding segmentation labels, and an enhanced 3D U-Net training pipeline. Duo-Diffusion employs segmentation-conditioned diffusion to ensure spatial consistency and precise anatomical detail in the generated data. Experimental evaluation on 720 abdominal MRI scans shows that models trained with combined real and synthetic data yield a mean Dice score gain of 0.67% over those using only real data, and achieve a 36.4% reduction in Fréchet Inception Distance (FID), reflecting enhanced image fidelity.

Figures & tables

Fig. 1: Examples of generated high-resolution synthetic abdominal MRI volumes with corresponding segmentation conditions overlaid. (a) Axial, (b) Coronal, and (c) Sagittal views showing segmentation masks inferred by a 3D U-Net trained on data generated with Duo-Diffusion. (d) Axial slices extracted from the generated 3D synthetic MRI volume.

Fig. 2: Overview of the proposed FOSCU framework. The workflow consists of three steps: (1) training the Duo-Diffusion architecture, which involves first training a 3D LDM to generate synthetic labels and then training a ControlNet to synthesize MRI volumes conditioned on paired labels, (2) performing inference with the pretrained Duo-Diffusion models to sequentially generate synthetic labels and corresponding synthetic volumes, and (3) leveraging the generated paired synthetic data to train a liver segmentation model capable of segmenting liver structures from synthetic MR images.

Fig. 3: The segmentation model in the proposed FOSCU framework is based on a standard 3D U-Net architecture. The network takes 3D synthetic abdominal MRI volumes as input and outputs liver segmentation masks. The numbers in each block indicate the feature map dimensions $H\times W\times D\times C$ , corresponding to height, width, depth, and channels.

Fig. 4: Representative examples of synthetic abdominal MRI generation. The left columns show non-tumor cases and the right columns show tumor (HCC) cases. Each example includes the ground truth label, the original volume, and the synthetic volume generated by Duo-Diffusion.

Method	FID ↓ (Ax.)	FID ↓ (Sag.)	FID ↓ (Cor.)	FID ↓ (Avg.)
3D-LDM (w/ VQVAE)	67.17	63.43	62.96	64.52
3D-LDM (w/ VAE)	40.69	37.34	37.83	38.62
Duo-Diffusion	29.27	27.62	28.04	28.31

TABLE I: Generative Model Performance Comparison

CNN Model	Segmentation	Real	Real + Synthesis	Improvement
U-Net	Liver-Only	0.9650	0.9662	+0.12 %
	Multi-Class	0.6968	0.7014	+0.66 %
ResUNet	Liver-Only	0.9633	0.9634	+0.01 %
	Multi-Class	0.6601	0.6652	+0.77 %
WideResUNet	Liver-Only	0.9657	0.9649	-0.08%
	Multi-Class	0.6961	0.7020	+0.85 %

TABLE II: Liver Segmentation Performance Results

Explore similar work

3D-LLDM: Label-Guided 3D Latent Diffusion Model for Improving High-Resolution Synthetic MR Imaging in Hepatic Structure Segmentation
Mar 25, 2026Kyeonghun Kim, Jaehyeok Bae, Youngung Han +10Medical Image Synthesis 3D Diffusion Models
OUTTA, Republic of Korea · Stanford University, USA · Seoul National University, Republic of Korea +6
Unsupervised Anatomical Feature Learning via Diffusion Models: Enhanced Medical Image Segmentation with Denoising Diffusion Probabilistic Models
Aug 26, 2026Akshat G, Divyansh Gupta, Shaleen Bhatnagar +2Medical Image Segmentation Medical Image Representation Learning
Manipal Institute of Technology Bengaluru, Manipal Academy of Higher Education, Manipal, India
Improving Medical Image Generative Models with Fréchet Distance Loss
Jul 14, 2026Andrew Marshall, Xuanang Xu, Xiaoran Zhang +3Synthetic Data Augmentation Distribution Matching
Department of Biomedical Engineering, Yale University · Department of Radiology & Biomedical Imaging, Yale University · Department of Electrical Engineering, Yale University

Mar 25, 2026J/K move · Enter open · S save

3D-LLDM: Label-Guided 3D Latent Diffusion Model for Improving High-Resolution Synthetic MR Imaging in Hepatic Structure Segmentation

Kyeonghun Kim, Jaehyeok Bae, Youngung Han, Joo Young Bae+9

OUTTA, Republic of Korea · Stanford University, USA · Seoul National University, Republic of Korea+6

Medical Image Synthesis 3D Diffusion Models Synthetic Data Augmentation Medical Image Segmentation

Open paper AskPDF arXiv

Deep learning and generative models are advancing rapidly, with synthetic data increasingly being integrated into training pipelines for downstream analysis tasks. However, in medical imaging, their adoption remains constrained by the scarcity of reliable annotated datasets. To address this limitation, we propose 3D-LLDM, a label-guided 3D latent diffusion model that generates high-quality synthetic magnetic resonance (MR) volumes with corresponding anatomical segmentation masks. Our approach uses hepatobiliary phase MR images enhanced with the Gd-EOB-DTPA contrast agent to derive structural masks for the liver, portal vein, hepatic vein, and hepatocellular carcinoma, which then guide volumetric synthesis through a ControlNet-based architecture. Trained on 720 real clinical hepatobiliary phase MR scans from Samsung Medical Center, 3D-LLDM achieves a Fréchet Inception Distance (FID) of 28.31, improving over GANs by 70.9% and over state-of-the-art diffusion baselines by 26.7%. When used for data augmentation, the synthetic volumes improve hepatocellular carcinoma segmentation by up to 11.153% Dice score across five CNN architectures.

Abstract

Explore similar work

3D-LLDM: Label-Guided 3D Latent Diffusion Model for Improving High-Resolution Synthetic MR Imaging in Hepatic Structure Segmentation

Unsupervised Anatomical Feature Learning via Diffusion Models: Enhanced Medical Image Segmentation with Denoising Diffusion Probabilistic Models

Improving Medical Image Generative Models with Fréchet Distance Loss