Style-Driven Data Synthesis and Degradation-Aware Enhancement for Ultrasound Image Restoration

Yu-Kai Wang; Chun-Xin Tan; Manh-Hung Nguyen; Ching-Chun Huang

cs.CVSep 28, 2026

Style-Driven Data Synthesis and Degradation-Aware Enhancement for Ultrasound Image Restoration

Authors: Yu-Kai Wang, Chun-Xin Tan, Manh-Hung Nguyen, Ching-Chun Huang

Organizations: Department of Computer Science, National Yang Ming Chiao Tung University, Hsinchu, Taiwan · Ho Chi Minh City University of Technology and Engineering, Ho Chi Minh City, Vietnam

Abstract

Low-cost handheld ultrasound devices can be widely deployed compared to professional hospital ultrasound machines. However, their images suffer from compound degradation that can mislead clinical judgment. Motivated by this observation, mapping handheld low-quality (LQ) to hospital high-quality (HQ) images has been considered a valuable research question. Conventionally, the mapping requires pixel-aligned LQ-HQ pairs. This requirement is unsatisfactory in practical scenarios because real scans at different times are never pixel-aligned. This paper addresses the challenge with a two-stage framework. The first stage generates pixel-aligned LQ-HQ datasets, and the second stage trains an enhancement model that improves LQ images. The first stage trains a cycle-consistent style-transfer model on unaligned real LQ-HQ pairs to learn a HQ-to-LQ model. Then, the model transforms real HQ images into pixel-aligned LQ images. Based on the dataset generated by the first stage, the second stage uses the Dual Degradation-Guided (DDG) Low-Rank Adaptation (LoRA) method to fine-tune an LQ-to-HQ model based on aligned pairs. In this stage, the model is based on the well known PiSA-SR framework but inserts a degradation-conditioned correction matrix. Experimental results on the USenhance2023 dataset show that the FID metric is improved by 16.7% over the strongest baseline while other metrics indicate that our enhanced outputs are well aligned with the real HQ distribution. The source code of our method is available at https://github.com/Jason0411202/DDG_LoRA.

Figures & tables

Figure 1 : Style-driven data synthesis pipeline. Stage 1 splits USenhance2023 into a 4:1 train/test partition (1), trains a style transfer model $G$ on unaligned real LQ–HQ pairs (2), applies $G$ to each real HQ image $y_{i}$ to obtain a pixel-aligned LQ counterpart $\tilde{x}_{i}$ (3), and assembles the USenhance2023-Aligned dataset (4). Stage 2 trains the DDG-LoRA enhancer on this dataset (5).

Figure 2 : DDG-LoRA architecture. The pixel stage trains $\Delta\theta_{\text{pix}}$ , then the semantic stage freezes it and trains $\Delta\theta_{\text{sem}}$ . Within each stage, DDG-LoRA replaces the static LoRA update with a degradation-conditioned variant driven by a descriptor $\boldsymbol{d}$ that the DE network extracts from the LQ image $x_{L}$ . The semantic stage further drops the CSD loss $\mathcal{L}_{\text{CSD}}$ .

Training LQ Source	FID $\downarrow$	NIQE $\downarrow$	PI $\downarrow$	Tenengrad $\uparrow$	Entropy $\uparrow$
realesrgan_deg [ 14 ]	127.01	6.038	5.598	0.0608	6.233
physics_guided_deg [ 8 ]	127.17	5.485	5.198	0.0618	6.320
usbsr_deg [ 7 ]	156.07	6.824	6.635	0.0405	6.287
ultrasound_deg	185.75	7.909	7.377	0.0336	6.163
USenhance2023-Aligned (ours)	99.00	5.619	4.764	0.1043	6.861

Table 1 : Comparison of DDG-LoRA (ours) trained with different LQ data sources, evaluated on the real-world USenhance2023 test set. Red = best, Blue = second best.

Panel of Figure 3 — Figure 3 : Visual comparison of all methods in Table 2 on a real-world USenhance2023 sample. The LQ input is shown in (a), the enhanced output of each method in (b)–(i), and the GT reference in (j). Note that because both the LQ and the GT are real-world scans, they are not pixel-aligned.

Method	FID $\downarrow$	NIQE $\downarrow$	PI $\downarrow$	Tenengrad $\uparrow$	Entropy $\uparrow$
EDSR [ 10 ]	187.57	9.058	7.541	0.0573	6.360
CycleSR [ 2 ]	192.36	8.293	7.139	0.0489	6.512
DARSR [ 19 ]	235.03	8.883	7.466	0.0680	6.341
PDM [ 11 ]	183.92	8.236	7.416	0.0317	6.415
USBSR [ 7 ]	168.33	8.416	7.609	0.0426	6.530
OSEDiff [ 16 ]	146.80	6.305	5.837	0.0531	6.273

Table 2 : Overall comparison of different enhancement methods on the real-world USenhance2023 test set. Red = best, Blue = second best.

Configuration	FID $\downarrow$	NIQE $\downarrow$	PI $\downarrow$	Tenengrad $\uparrow$	Entropy $\uparrow$
w/o DDG-LoRA	98.98	5.589	4.815	0.1006	6.808
w/ DDG-LoRA (ours)	99.00	5.619	4.764	0.1043	6.861

Table 3 : Ablation study on the degradation conditioning module in DDG-LoRA. Bold red indicates the better result for each metric.

Explore similar work

HIR-ALIGN: Enhancing Hyperspectral Image Restoration via Diffusion-Based Data Generation
May 13, 2026Li Pang, Heng Zhao, Yijia Zhang +2Image Restoration Hyperspectral Image
A Paired Point-of-Care Ultrasound Dataset for Image Quality Enhancement and Benchmarking via a cGAN Baseline
May 8, 2026Lennard M. van Karnenbeek, Hilde G. A. van der Pol, Mark Wijkhuizen +5Clinical Ultrasound Dataset Ultrasound
Department of Nanobiophysics, Faculty of Science and Technology, University of Twente, Drienerlolaan 5, 7522 NB Enschede, The Netherlands · Image-Guided Surgery, Department of Surgery, Netherlands Cancer Institute, Plesmanlaan 121, 1066 CX Amsterdam, The Netherlands · Department of Radiology, Netherlands Cancer Institute, Amsterdam, The Netherlands +1
Language-Assisted Super-Resolution from Real-World Low-Resolution Patches
Jun 30, 2026Joonkyu Park, Kyoung Mu LeeReal-World Image Super-Resolution Generative Super-Resolution
Dept. of ECE&ASRI, 2IPAI, Seoul National University

May 13, 2026, cs.CVJ/K move · Enter open · S save

HIR-ALIGN: Enhancing Hyperspectral Image Restoration via Diffusion-Based Data Generation

Li Pang, Heng Zhao, Yijia Zhang, Deyu Meng+1

Image Restoration Hyperspectral Image Restoration Signal Reconstruction

Open paper AskPDF arXiv

Hyperspectral image (HSI) restoration is crucial for reliable analysis, as real-world HSIs suffer from noise, blur, and resolution loss. However, existing models trained on source data often fail on target domains lacking clean references, a common real-world scenario. To address this, we present HIR-ALIGN, a plug-and-play target-adaptive augmentation framework that enhances HSI restoration by augmenting limited training images with synthetic data matching the target distribution, without extra clean target-domain HSI data. It has three stages: (i) proxy generation, where off-the-shelf restoration models are applied to degraded target observations to produce semantics-preserving proxy HSIs that approximate clean target-domain images; (ii) distribution-adaptive synthesis, where a blur-robust unCLIP diffusion model generates target-aligned RGBs from proxy RGBs with prompt conditioning and embedding-space noise initialization. The warp-based spectral transfer module then synthesizes HSIs by aligning each generated RGB with its proxy RGB, estimating soft patch-wise transport weights, and applying these weights and learnable local interpolation kernels to the proxy HSI; and (iii) aligned supervised finetuning, where restoration networks pretrained on the source distribution are finetuned with proxy HSIs and synthesized target-aligned HSIs, then deployed on degraded target images. We also provide theoretical analysis showing that, under stated assumptions, the proposed augmentation-based finetuning obtains a tighter target-domain restoration-risk upper bound by jointly improving target-distribution coverage and controlling spectral bias. Experiments on simulated and real datasets across denoising, super-resolution, and other restoration tasks demonstrate that HIR-ALIGN is superior to proxy-only target-adaptation baselines and outperforms representative unsupervised methods in most cases.

Abstract

Explore similar work

HIR-ALIGN: Enhancing Hyperspectral Image Restoration via Diffusion-Based Data Generation

A Paired Point-of-Care Ultrasound Dataset for Image Quality Enhancement and Benchmarking via a cGAN Baseline

Language-Assisted Super-Resolution from Real-World Low-Resolution Patches