融合扩散模型的超声成像算法研究

袁野; 黄民尚; 杨伟锋

doi:10.11999/JEIT251083

融合扩散模型的超声成像算法研究

doi: 10.11999/JEIT251083 cstr: 32379.14.JEIT251083

汕头大学广东省人工智能与现代超声工程技术研究中心汕头 515044

基金项目: 国家自然科学基金面上项目 (82071992)

详细信息

作者简介:
袁野：男，博士，副教授，研究方向为基于人工智能的医学图像分析

黄民尚：男，硕士研究生，研究方向为基于人工智能的医学图像分析

杨伟锋：男，硕士研究生，研究方向为基于人工智能的医学图像分析

通讯作者:
袁野　yuanye@stu.edu.cn

中图分类号: TN911.73; TP391
计量
- 文章访问数: 6
- HTML全文浏览量: 0
- PDF下载量: 10
- 被引次数: 0
出版历程
- 收稿日期: 2025-10-13
- 修回日期: 2026-03-03
- 录用日期: 2026-03-03
- 网络出版日期: 2026-03-15

Research on Ultrasound Imaging Algorithm Fused with Diffusion Model

*(Guangdong Provincial Engineering Research Center for Artificial Intelligence and Modern Ultrasonic Technology, Shantou University, Shantou 515044, China)

摘要

摘要: 针对超声成像分辨率低及易受伪影干扰问题，该文提出基于扩散模型(DM)的U-DM超声成像质量优化方法。通过构建差值训练机制与解剖结构引导策略，结合改进型UNet网络架构实现多尺度特征融合，建立从含噪超声数据到高质量图像的映射关系，进而生成高质量超声图像。基于PICMUS数据集的实验结果表明，该文提出的U-DM方法在噪声抑制与结构保持方面显著优于Unet、UNetGAN等方法，能有效消除人工伪影并恢复解剖细节，其图像重建质量达到临床诊断要求。相较于生成对抗网络(GAN)，该文提出的融合扩散模型的超声成像方法展现出更稳定的训练特性和更优的泛化能力，克服了模式坍塌等固有问题，为突破超声成像质量瓶颈提供了新途径。
- 超声图像 /
- 扩散模型(DM) /
- UNet /
- 差值训练机制
Abstract: Objective Medical ultrasound imaging, which utilizes ultrasonic waves to probe human tissues and generates images via signal processing of the returning echoes, has become a vital clinical diagnostic tool due to its non-invasive, safe, and real-time nature. However, conventional ultrasound imaging is fundamentally limited by factors such as the finite width of ultrasonic pulses, variations in tissue acoustic impedance, and the complexity of echo signals, leading to pervasive challenges including insufficient spatial resolution, significant speckle noise, and off-axis artifacts. These limitations directly impair the detection of lesions and diagnostic accuracy. While traditional approaches focusing on hardware optimization and signal processing algorithms like adaptive beamforming have made incremental improvements, they are often constrained by physical laws, computational complexity, and reliance on manual parameter tuning. Recent deep learning-based methods, particularly those using generative adversarial networks (GANs), offer promising results but suffer from training instability and poor interpretability. The emerging diffusion model, a state-of-the-art generative paradigm, has demonstrated superior robustness and generalization in computed tomography (CT) and magnetic resonance imaging (MRI) reconstruction, yet its application in ultrasound imaging remains largely unexplored. This study aims to fill this critical gap by developing a novel diffusion model-based framework for high-quality ultrasound image formation, seeking to overcome the inherent limitations of existing methods and provide a stable, efficient, and interpretable solution for enhancing ultrasound image quality. Methods This research proposes a novel ultrasound imaging method based on a denoising diffusion probabilistic model (DDPM). The core of our approach is a multi-scale diffusion network architecture designed to progressively refine a low-quality ultrasound image (e.g., one formed by a simple Delay-and-Sum, DAS, beamformer) into a high-quality image. The process consists of a forward and a reverse process. In the forward process, Gaussian noise is gradually added to a high-quality ground-truth image over a series of timesteps. The reverse process is trained to learn the conditional denoising function. Our custom-designed denoising network takes a low-resolution DAS image as a conditional input and fuses it with the noisy image at each denoising step through residual connections and feature-wise transformations at multiple scales. This deep fusion mechanism allows the network to effectively incorporate the underlying anatomical structure from the low-quality input while iteratively removing noise and artifacts through the diffusion process. The model was trained using a dataset of paired low-quality and high-quality ultrasound images, with the high-quality images serving as the training target. The training objective was to maximize the variational lower bound on the likelihood, effectively teaching the network to reverse the noising process. The performance of the proposed method was quantitatively evaluated against traditional DAS, minimum variance (MV) beamforming, and a leading GAN-based super-resolution method using metrics including Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Results and Discussions The proposed diffusion model demonstrated superior performance in enhancing ultrasound image quality. Quantitatively, our method achieved a mean PSNR of 35.2 dB and an SSIM of 0.933, representing a significant improvement of 4.5 dB in PSNR while maintaining exceptional structural fidelity compared to conventional beamforming approaches. The method also consistently outperformed adaptive minimum variance beamforming and GAN-based approaches across all evaluation metrics, including contrast-to-noise ratio. Visual assessment confirms these quantitative findings. The generated images exhibit markedly reduced speckle noise and significantly enhanced boundary clarity of anatomical structures. Critically, these improvements were achieved without introducing the blurring or artificial textures commonly observed in other deep learning-based methods. The multi-scale architecture with conditional feature injection effectively preserved structural integrity, as evidenced by the clear and continuous edges in the output.The progressive denoising nature of our approach provides inherent interpretability to the image refinement process. Unlike the opaque single-step generation of other deep learning models, our method offers transparent, step-wise enhancement from initial input to final output. Furthermore, the training process remained stable and convergent, avoiding the instability issues that frequently plague adversarial training methods. Ablation studies confirmed the critical importance of the deep fusion mechanism, while resolution analysis verified substantial improvements in both lateral and axial resolution compared to all baseline methods. Conclusions This study successfully developed and validated a novel ultrasound imaging method based on a diffusion model. The proposed framework effectively addresses key limitations in conventional and existing deep learning-based approaches. It bypasses the complex matrix computations and manual parameter tuning required by adaptive beamformers and offers a more stable training paradigm compared to GANs. The results conclusively demonstrate that the method can significantly enhance image quality by substantially improving the PSNR and maintaining excellent structural similarity, leading to images with suppressed noise, reduced artifacts, and improved resolution. The multi-scale diffusion process ensures the preservation of anatomical structures while providing a degree of interpretability to the image generation process. This work establishes diffusion models as a powerful and promising new paradigm for advanced ultrasound imaging, offering a robust and high-performance technical pathway to break through the current bottlenecks in ultrasound image quality, with potential for broad clinical impact.
- Ultrasound imaging /
- Diffusion model (DM) /
- UNet /
- Difference training mechanisms

HTML全文

图 1 U-DM模型流程图

下载: 全尺寸图片幻灯片

图 2 PICMUS数据集超声成像示例，单角度成像(上)和75角度相干平面波复合成像(下)

下载: 全尺寸图片幻灯片

图 3 U-DM与1PW、75PWs的超声成像局部细节对比图

下载: 全尺寸图片幻灯片

图 4 U-DM模型与1PW、75PWs的点散射体41mm处波形图对比

下载: 全尺寸图片幻灯片

图 5 消融实验下颈动脉纵截面与模拟囊肿对比图像

下载: 全尺寸图片幻灯片

图 6 U-DM与各模型的颈动脉成像对比(横截面(下)纵截面(上))

下载: 全尺寸图片幻灯片

图 7 本文U-DM模型与各模型的囊肿成像对比(真实囊肿(上)和模拟囊肿(下))

下载: 全尺寸图片幻灯片

图 8 U-DM与各模型的点散射体成像对比(真实点散射体(上)和模拟点散射体(下))

下载: 全尺寸图片幻灯片

图 9 点散射体图像41 mm处的波形图对比。注：模拟数据半峰全宽(FWHM，单位：mm)：UNet 0.82、UNet+Attention 0.85、UNetGAN 0.91、AUGAN 0.88、U-DM 0.89；真实数据 FWHM(单位：mm)：UNet 0.95、UNet+Attention 0.98、UNetGAN 1.02、AUGAN 0.99、U-DM 0.98

下载: 全尺寸图片幻灯片

表 1 基于颈动脉数据的PSNR和SSIM指标对比结果

超声图像数据	1PW	U-DM
颈动脉横截面PSNR(dB)	21.343	35.423
颈动脉纵截面PSNR(dB)	19.884	34.944
颈动脉横截面 SSIM	0.730	0.931
颈动脉纵截面 SSIM	0.709	0.927

下载: 导出CSV

表 2 囊肿体膜数据的CR、CNR指标结果

超声图像数据	1PW	U-DM
真实囊肿体膜CR(dB)	14.293	30.849
模拟囊肿体膜CR(dB)	16.823	36.502
真实囊肿体膜CNR	1.797	3.533
模拟囊肿体膜CNR	1.093	1.679

下载: 导出CSV

表 3 消融实验结果

模型配置	PSNR(dB)	SSIM	CR(dB)	CNR
完整 U-DM 模型	34.948	0.927	27.032	3.335
去除差值训练	26.295	0.773	24.468	2.646
去除解剖引导	23.469	0.675	22.259	2.523

下载: 导出CSV

表 4 U-DM方法与其他网络结构在颈动脉图像的PSNR、SSIM指标上的对比

方法	UNet	UNet+Attention	UNetGAN	AUGAN	U-DM
颈动脉横截面PSNR(dB)	30.514	32.239	33.943	34.128	35.423
颈动脉纵截面PSNR(dB)	29.668	31.125	33.509	33.486	34.944
颈动脉横截面 SSIM	0.872	0.888	0.912	0.921	0.931
颈动脉纵截面 SSIM	0.861	0.873	0.903	0.905	0.927

下载: 导出CSV

表 5 囊肿图像的CR、CNR指标对比

方法	UNet	UNet+Attention	UNetGAN	AUGAN	U-DM
真实囊肿体膜 CR(dB)	28.820	30.445	31.061	31.496	30.849
模拟囊肿体膜 CR(dB)	35.318	36.034	44.507	44.873	36.502
真实囊肿体膜 CNR	2.721	3.110	3.429	3.481	3.533
模拟囊肿体膜 CNR	1.771	1.798	1.992	2.087	1.897

下载: 导出CSV

表 6 不同模型在 PICMUS 各场景下的性能变异系数

方法	PSNR变异系数 (CV%)	SSIM变异系数(CV%)	CR变异系数(CV%)	CNR变异系数(CV%)	综合变异系数(CV%)
AUGAN	11.0%	5.3%	10.1%	17.0%	10.9%
UNetGAN	10.6%	4.0%	8.2%	18.6%	10.4%
UNet+Attention	2.6%	1.4%	6.0%	5.5%	3.9%
U-DM	3.8%	1.3%	7.1%	2.3%	3.6%
UNet	0.7%	0.2%	6.1%	5.8%	3.2%

下载: 导出CSV

参考文献(26)

[1]	王浩, 李娟, 张远, 等. 医学超声成像技术的发展现状与挑战[J]. 中国医疗器械杂志, 2022, 46(5): 521–526. doi: 10.3969/j.issn.1671-7104.2022.05.009.(查阅网上资料,未找到本条文献信息且doi与文献内容不符,请确认). WANG Hao, LI Juan, ZHANG Yuan, et al. Development status and challenges of medical ultrasound imaging technology[J]. Chinese Journal of Medical Devices, 2022, 46(5): 521–526. doi: 10.3969/j.issn.1671-7104.2022.05.009.
[2]	ZHANG Yuhan, CHEN Lina, WANG Xiaoming, et al. Ultrasound image denoising based on wavelet transform and adaptive thresholding[J]. IEEE Transactions on Medical Imaging, 2021, 40(8): 2054–2065. doi: 10.1109/TMI.2021.3071234.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[3]	刘敏, 赵强, 孙伟. 超声成像中离轴伪影的产生机制与抑制方法[J]. 声学技术, 2023, 42(1): 89–95. doi: 10.16300/j.cnki.1000-3630.2023.01.015.(查阅网上资料,未找到本条文献信息且doi与文献内容不符,请确认). LIU Min, ZHAO Qiang, and SUN Wei. Generation mechanism and suppression method of off-axis artifacts in ultrasound imaging[J]. Technical Acoustics, 2023, 42(1): 89–95. doi: 10.16300/j.cnki.1000-3630.2023.01.015.
[4]	CHEN Minjie, LI Yanping, ZHANG Hongwei, et al. Impact of ultrasound image quality on early liver lesion detection: A clinical study[J]. European Radiology, 2022, 32(6): 4123–4132. doi: 10.1007/s00330-021-08546-x.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[5]	WANG Zhiqiang, LIU Jingwen, ZHANG Lei, et al. Design of a 128-element phased array probe for high-resolution ultrasound imaging[J]. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 2021, 68(4): 1201–1210. doi: 10.1109/TUFFC.2021.3052345.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[6]	MORA P, CHEKROUN M, RAETZ S, et al. Nonlinear generation of a zero group velocity mode in an elastic plate by non-collinear mixing[J]. Ultrasonics, 2022, 119: 106589. doi: 10.1016/j.ultras.2021.106589.
[7]	张克潜, 李德杰. 微波与光电子学中的电磁理论[M]. 2版. 北京: 电子工业出版社, 2001: 210–215. ZHANG Keqian and LI Dejie. Electromagnetic Theory for Microwaves and Optoelectronics[M]. 2nd ed. Beijing: Publishing House of Electronics Industry, 2001: 210–215.
[8]	JENSEN J A. Fundamentals of Medical Ultrasound[M]. 2nd ed. New York: Cambridge University Press, 2018: 156–162. (查阅网上资料, 未找到本条文献信息, 请确认).
[9]	CAPON J. High-resolution frequency-wavenumber spectrum analysis[J]. Proceedings of the IEEE, 1969, 57(8): 1408–1418. doi: 10.1109/PROC.1969.7278.
[10]	JIANG Tao, ZHANG Yufei, WANG Ling, et al. Robust adaptive beamforming for ultrasound imaging under broadband signal conditions[J]. IEEE Journal of Biomedical and Health Informatics, 2021, 25(9): 3345–3354. doi: 10.1109/JBHI.2021.3087643.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[11]	HO D J, MONTSERRAT D M, FU Chichen, et al. Sphere estimation network: Three-dimensional nuclei detection of fluorescence microscopy images[J]. Journal of Medical Imaging, 2020, 7(4): 044003. doi: 10.1117/1.JMI.7.4.044003.
[12]	CHEN Xin, WANG Hui, LIU Zhi, et al. Diagonal loading adaptive beamforming for ultrasound image quality enhancement[J]. Ultrasonic Imaging, 2022, 44(2): 89–102. doi: 10.1177/01617346211068234.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[13]	赵鑫, 马晓伟, 刘杰. 自适应波束成形在超声成像中的应用进展[J]. 中国医学物理学杂志, 2023, 40(3): 301–307. doi: 10.3969/j.issn.1005-202X.2023.03.005.(查阅网上资料,未找到本条文献信息且doi与文献内容不符,请确认). ZHAO Xin, MA Xiaowei, and LIU Jie. Application progress of adaptive beamforming in ultrasound imaging[J]. Chinese Journal of Medical Physics, 2023, 40(3): 301–307. doi: 10.3969/j.issn.1005-202X.2023.03.005.
[14]	LIU Yong, ZHANG Hui, WANG Zhijie, et al. End-to-end ultrasound image super-resolution using a modified U-Net[J]. IEEE Transactions on Medical Imaging, 2022, 41(5): 1123–1134. doi: 10.1109/TMI.2022.3145678.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[15]	BILLOT B, GREVE D N, PUONTI O, et al. SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining[J]. Medical Image Analysis, 2023, 86: 102789. doi: 10.1016/j.media.2023.102789.
[16]	ZHANG Cheng, JIANG Tao, LIU Jie, et al. Transformer-based ultrasound beamforming: A real-time implementation[J]. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 2023, 70(2): 345–354. doi: 10.1109/TUFFC.2022.3228976.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[17]	李建峰, 张宇, 王莉, 等. 基于深度学习的超声图像伪影抑制算法[J]. 电子与信息学报, 2024, 46(8): 2567–2575. doi: 10.11999/JEIT240512.(查阅网上资料,未找到本条文献信息且doi打不开,请确认). LI Jianfeng, ZHANG Yu, WANG Li, et al. Ultrasound image artifact suppression algorithm based on deep learning[J]. Journal of Electronics & Information Technology, 2024, 46(8): 2567–2575. doi: 10.11999/JEIT240512.
[18]	孙萌, 周明, 吴涛. 深度学习超声成像的可解释性研究进展[J]. 生物医学工程学杂志, 2023, 40(2): 389–396. doi: 10.7507/1001-5515.202208052.(查阅网上资料,未找到本条文献信息且doi打不开,请确认). SUN Meng, ZHOU Ming, and WU Tao. Research progress on interpretability of deep learning-based ultrasound imaging[J]. Journal of Biomedical Engineering, 2023, 40(2): 389–396. doi: 10.7507/1001-5515.202208052.
[19]	WANG Hong, LI Jun, ZHANG Ying, et al. GAN-based ultrasound image enhancement for breast lesion detection[J]. Computerized Medical Imaging and Graphics, 2022, 98: 102185. doi: 10.1016/j.compmedimag.2022.102185.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[20]	GOODFELLOW I J, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[C]. Proceedings of the 28th International Conference on Neural Information Processing Systems - Volume 2, Montreal, Canada, 2014: 2672–2680.
[21]	LIU Zhen, CHEN Guang, WANG Lin, et al. Stabilizing GAN training for medical image generation: A survey[J]. IEEE Reviews in Biomedical Engineering, 2023, 16: 234–250. doi: 10.1109/RBME.2022.3227890.(查阅网上资料,未找到本条文献信息且doi打不开,请确认).
[22]	KIM T, LEE J, PARK S, et al. Diffusion model-based MRI reconstruction with reduced scan time[J]. Magnetic Resonance in Medicine, 2022, 88(4): 1823–1835. doi: 10.1002/mrm.29203.(查阅网上资料,未找到本条文献信息且doi与文献内容不符,请确认).
[23]	LIU Jiawei, ZHANG Hongliang, WANG Zhijie, et al. CT image artifact reduction using a diffusion model[J]. Medical Physics, 2023, 50(3): 1689–1700. doi: 10.1002/mp.16245.(查阅网上资料,未找到本条文献信息且doi与文献内容不符,请确认).
[24]	赵阳, 陈曦, 刘浩. 扩散模型在医学影像重建中的改进与应用[J]. 电子与信息学报, 2025, 47(2): 589–597. doi: 10.11999/JEIT250108.(查阅网上资料,未找到本条文献信息且doi打不开,请确认). ZHAO Yang, CHEN Xi, and LIU Hao. Improvement and application of diffusion model in medical image reconstruction[J]. Journal of Electronics & Information Technology, 2025, 47(2): 589–597. doi: 10.11999/JEIT250108.
[25]	陈阳, 李娜, 张伟. 扩散模型在医学影像领域的应用综述[J]. 自动化学报, 2023, 49(5): 921–938. doi: 10.16383/j.aas.c220638.(查阅网上资料,未找到本条文献信息且doi与文献内容不符,请确认). CHEN Yang, LI Na, and ZHANG Wei. A review of diffusion models in medical imaging[J]. Acta Automatica Sinica, 2023, 49(5): 921–938. doi: 10.16383/j.aas.c220638.
[26]	杨磊, 宋昊, 申瑞阳, 等. 强稀疏低副瓣近场聚焦稀疏阵列三维成像[J]. 电子与信息学报, 2024, 46(12): 4471–4482. doi: 10.11999/JEIT231278. YANG Lei, SONG Hao, SHEN Ruiyang, et al. High sparsity and low sidelobe near-field focused sparse array for three-dimensional imagery[J]. Journal of Electronics & Information Technology, 2024, 46(12): 4471–4482. doi: 10.11999/JEIT231278.