Diffusion Models in Medical Imaging: A Comprehensive Survey
-
摘要:
以扩散模型为代表的生成式人工智能近年来在医学成像领域取得了迅猛的进展。为了帮助更多学者全面的了解扩散模型这一先进技术，本文旨在提供扩散模型在医学成像领域的详细概述。首先以扩散模型的起源演变为主线介绍扩散建模框架的基础理论和基本概念；其次根据扩散模型的特点提供其在医学成像领域的系统分类，并涵盖不同成像模态如磁共振成像、计算机断层扫描、正电子发射计算机断层显像和光声成像等的广泛应用；最后讨论目前扩散模型的局限性并展望未来研究的潜在发展方向，为研究者后续的探索提供了一个直观的起始点。部分代码开源在网址：https://github.com/yqx7150/Diffusion-Models-for-Medical-Imaging。
Abstract:Generative artificial intelligence represented by diffusion models has significantly contributed to medical imaging reconstruction. To help researchers comprehensively understand the rich content of diffusion models, this review provides a detailed overview of diffusion models used in medical imaging reconstruction. The theoretical foundation and fundamental concepts underlying the diffusion modeling framework were first introduced, describing their origin and evolution. Second, a systematic characteristic-based taxonomy of diffusion models used in medical imaging reconstruction is provided, broadly covering their application to imaging modalities, including magnetic resonance imaging (MRI), computed tomography (CT), positron emission computed tomography (PET), and photoacoustic imaging (PAI). Finally, we discuss the limitations of current diffusion models and anticipate potential directions of future research, providing an intuitive starting point for subsequent exploratory research. Related codes are available at GitHub: https://github.com/yqx7150/Diffusion-Models-for-Medical-Imaging.
-
肋软骨是一种具有一定弹性、强度和柔韧性的透明软骨，因其来源安全可靠、组织量充足[1]，且具有移植后吸收率低、易塑形、自体肋软骨移植组织相容性好、无免疫排斥反应[2-3]等特点，是整形外科手术填充或支撑的理想材料，广泛应用于各种原因导致的鼻畸形矫正术、耳廓缺损外耳再造术等整形外科领域[1-4]，其中以采取第6～8肋软骨最为常用。肋软骨发育情况和钙化程度是术前评估的重要内容，术前了解患者肋软骨的形态至关重要。
相对于超声（ultrasound，US）和磁共振成像（magnetic resonance imaging，MRI），多层螺旋CT（multi-slice spiral computed tomography，MSCT）及后处理技术能够清晰准确地反映肋骨及肋软骨发育情况、真实形态、体积和钙化程度[5-7]，已广泛应用于临床[8-10]。
通常CT胶片上的影像均小于人体组织的实际大小，临床医生需要根据胶片上附带的比例尺进行换算才能获得组织的真实值。为了使临床医生的判断过程更加便捷和直观，我们将换算工作前移，使胶片上显示的肋软骨的形态特征和尺寸即为肋软骨的真实状态，并联合手术医生进行了实体测量加以验证。
1. 材料与方法
1.1 临床资料
影像资料来源于2021年8月至9月在中国医学科学院院整形外科医院就诊，进行鼻整形、外耳再造等手术治疗的患者。随机选取31例患者影像数据与手术测量数据相比较，其中男性20例，女性11例，年龄5.67～50.67岁，平均（15.6±9.78）岁；BMI 12.61～26.02 kg/m2，平均（18.7±3.04）kg/m2。
1.2 设备及材料
采用Philips Brilliance 64排MSCT机（Philips公司），Philips Extended Brilliance Workspace后处理工作站（Philips EBW，Philips公司，版本4.5.3），19英寸医用显示器，FUJI DRYPIX Plus激光相机（FUJIFILM公司），14英寸×17英寸（35 cm×43 cm）医用干式激光胶片（FUJIFILM公司）。量程为30 mm，分度值为0.05 mm的游标卡尺；量程为200 mm，分度值为0.5 mm的外科直尺。
1.3 研究方法
患者行全胸廓CT扫描后，在Philips EBW工作站上做图像后处理。以横轴面显示肋软骨厚度，以最大密度投影（maximum intensity projection，MIP）影像显示肋软骨长度和宽度[7]。打印胶片全部采用14×17规格，以“2×2分格呈4幅图像”和“1×1分格呈单幅图像”两种图像布局方式打印图像。
1.3.1 设计实物比例尺
首先测量出14×17医用胶片影像显示区域的总长为38.8 cm，此长度对应1×1分格的图像；之后将该显示区域以2×2分格后，测得胶片上每个分格的长度为15.2 cm；再测量出EBW工作站19英寸医用显示器屏幕上主影像显示区长度为25.8 cm。本研究中以上3个值均为固定值，不因患者体态改变而改变。
继而计算出EBW工作站显示屏上主影像显示区和14×17医用胶片影像显示区之间不同的比例关系，按照计算得出的比例关系制作了长度为8.49 cm的5 cm实物比例尺，截取中间6.65 cm部分作为10 cm实物比例尺。使用时，通过缩放影像使电子比例尺的5 cm或10 cm长度与对应的实物比例尺长度相吻合，这样打印的胶片影像中的电子比例尺能真实反映实物大小（图1（a）和图1（b））。
1.3.2 划分肋软骨升部及横部
鉴于肋软骨的形态特性，为便于描述，我们将第6～8肋软骨的长度按弧度每根划分为两部分，第1部分为肋软骨胸骨端至下方相邻的肋软骨相连的转折处，称为肋软骨升部；第2部分为此转折处至肋软骨肋骨端，称为肋软骨横部（图2）。
1.3.3 影像后处理及打印方式
影像后处理及打印过程分为3个步骤。
（1）二维影像的选取和打印：选取肋软骨胸骨端、肋软骨升部1/2处、肋软骨升部横部结合处、肋软骨肋骨端共4个点的二维图像，以横轴面为主显示影像，以冠状面、矢状面为辅助定位影像，矢冠轴三图同步，选用5 cm电子比例尺和实物比例尺相吻合，胶片以“2×2”分格打印（图3（a））。
![]()
图 3 利用5 cm和10 cm实物比例尺在胶片上打印肋软骨真实大小影像Figure 3. Print the true size image of costal cartilage on filsm with 5 cm and 10 cm physical scales（2）三维影像的选取和打印：运用MIP影像进行三维重组，获取显示全部前肋弓（含肋软骨）的三维影像，选用10 cm电子比例尺和实物比例尺相吻合，调整图像大小以1×1分格打印，此胶片上可观察和测量全部肋软骨升部的长度和宽度（图3（b））。
保持三维影像上的人体正中线垂直于水平面，调整影像大小至5 cm电子比例尺和实物比例尺相吻合，然后将影像分别向人体正中线方向旋转，使左右两侧肋软骨横部整体旋转到正对视者的位置，此位置用作肋软骨横部长度和宽度的测量，以2×2分格打印（图3（c））。
上述3个步骤形成了胶片在不同分格状态下影像的打印。
1.4 统计学方法
收集的数据采用运用 SAS软件进行统计分析。计算患者第6肋软骨6个相同位置胶片上测量值和人体实际测量值的均数和标准差，采用t检验的方法比较胶片上测量值和人体实际测量值的差异，P＜0.05为差异有统计学意义。
2. 结果
2.1 胶片验证
测量已打印胶片上显示的比例尺，胶片上所示的10 cm和5 cm比例尺标注值与直尺的真实值相等（图4（a）和图4（b）），表明胶片为等比例打印。
2.2 胶片测量值和手术中实体测量值的统计学比较
对31例患者中的14例患者进行手术中取材验证，选取患者右侧第6肋软骨升部、横部的长度，升部横部连接处宽度、厚度，肋软骨胸骨端、肋骨端的厚度等6个测量目标，采用SAS软件比较上述目标的胶片测量值（图5）和实体测量值（图6）之间的差异（表1）。经t检验分析，两种方法6个位置的测量值差异无统计学意义，表明胶片上测量值等同于人体真实值。
![]()
图 5 在胶片上测量右侧第6肋软骨胸骨端厚度（矢冠轴3图以十字线同步定位）Figure 5. The measure the thickness of the sternal end of the sixth costal cartilage on the film![]()
图 6 右侧第6肋软骨胸骨端术中实体测量厚度Figure 6. The measured thickness of the sternal end of the sixth costal cartilage on the right during operation表 1 肋软骨6个相同位置胶片测量值和人体测量值差异统计Table 1. Difference statistics of film measurement value and human measurement value at five same positions of costal cartilage分类 胶片测量值（均数±标准差） 统计检验 实体 胶片 t P 肋软骨升部长度/cm 77.46±13.46 77.84±13.45 -0.07 0.94 肋软骨横部长度/cm 33.44±5.49 33.24±5.85 0.09 0.93 升部横部连接处宽度/cm 15.83±1.94 15.80±2.24 0.03 0.98 升部横部连接处厚度/cm 5.95±1.77 5.96±1.77 -0.01 0.99 肋软骨胸骨端厚度/cm 12.05±2.50 12.02±2.49 0.03 0.98 肋软骨肋骨端厚度/cm 8.48±1.11 8.49±1.02 -0.04 0.97 3. 讨论
在医学影像领域，胶片上的影像等比例显示一般只出现在普通X线摄影中，传统的屏-片系统[11]胶片经X线曝光后通过显影、定影等一系列步骤在胶片上形成可视的有黑白密度差的影像，这种影像在不考虑X线曝光时斜射线效应[12]的前提下，可以视为与人体组织实际大小相等。CT是医学影像进入计算机数字时代的代表，其成像方式是X线束沿人体长轴对检查部位进行一定厚度的逐层扫描或螺旋扫描，光信号转变为电信号，经计算机的数模转换，每个数字转为黑白不等的灰阶，并按矩阵排列即构成CT图像。所产生的影像数量庞大，可生成成百上千幅图像，由于胶片要做到可视性和承载量两者的统一，因此实际工作中从未实现过在胶片上打印出和人体某一组织真实大小相同的CT影像。
肋软骨的术前筛选、评估，术中的采取是鼻整形、外耳再造等整形外科手术治疗过程中的重要环节[2]。王永振等[6]使用多层螺旋CT三维重组技术在MIP图像中测量肋软骨长、宽、厚度，对肋软骨质量进行评估；汤婷等[13]多层螺旋CT容积再现技术在肋软骨切取中的应用，均证实CT三维重组影像测得肋软骨数据与术中测量结果基本一致。而对于外科医生来讲，能够直观且便捷地获得与人体组织实际大小相同的影像是手术设计和实际操作中的一个迫切愿望。
虽然现在有多种计算机测量软件且使用方便，但是在术中，由于术者操作电脑的不便，在手术室基本还是使用观片灯对胶片进行阅读和测量。因此实现影像在胶片上的真实大小打印，使术者不需经过比例换算便可获得真实尺寸，实用价值毋庸置疑，同时又避免比例换算过程中存在的影响精度的系统性误差以及发生选择性错误等缺点。
实现胶片的真实大小打印，可以使术者观察更直接、测量更便捷、测量值更准确、术中操作更从容。同时胶片打印是基于DICOM 3.0[14]协议下的打印模式，具有严谨的科学性。所获得的长、宽、厚的真实值及全部前肋弓真实形态可以帮助医生准确定位肋软骨，了解肋软骨本身的发育情况，钙化程度及其与周围肋软骨、肋骨以及胸骨联合情况以及和脏器的毗邻关系，有利于医生筛选出适宜手术的患者，提高术中肋软骨采取的质量和效率，缩短时间、减小切口、减少软骨及周围软组织损伤，减少气血胸及肋软骨折断的发生机率，同时避免肋软骨及其采取术区过长时间暴露，降低感染发生率[15]。
本方法在进行胶片测量和实体验证测量时发现所选的的点测量值上有微小的偏差，造成的原因可能是实体测量点和胶片选点的位置稍有偏移，也不排除影像后处理过程中在生成胶片影像时电子比例尺与实物比例尺10 cm或5 cm吻合度是略有偏差造成的，不过这两种情况的误差都非常小，实际手术中基本可以忽略不计。
此外，由于肋软骨和相邻的皮下脂肪及脏器的密度比较接近，在显示肋软骨厚度方面，横轴位影像显示的肋软骨断面会因肋软骨被相邻组织紧贴而区分不清，实际操作中可以采取以矢状面替代横轴面为主显示影像的方法，解决上述问题。
从解剖形态上分，肋软骨属形态不规则的透明软骨组织，从第5肋软骨开始自人体躯干的正中向两侧后下方走形，人体正面观呈左右基本对称的“八”字形，人体第5～8肋软骨的胸骨端和肋骨端处在不同的冠状面上，而三维影像在胶片上是以二维的形式显示，其测量值尤其是肋软骨总长度这一指标确实会存在一定误差[13]。因此在打印胶片时全部肋软骨除以1×1分格正面显示以外，还需将胸廓整体向左右方向分别做水平旋转，把肋软骨横部转至于视者正对位置，以2×2分格打印，这样获得的肋软骨横部测量值更为准确。之所以选择MIP影像显示肋软骨长度和宽度，主要是各种CT影像后处理技术中，MIP相对于VR等技术显示肋软骨边界更为清晰，VR图像立体性强，与人体肋骨及肋软骨解剖结果更相近[16]，更适合多角度观察肋软骨形态及走行[7]以及软骨量及钙化情况[17]。
本研究针对临床医生希望获得一种能够直观快速而精准地量化目标肋软骨组织的影像学技术手段这一临床需求，探索并总结形成了一种可以在胶片上直接提供真实准确肋软骨影像及形态学参数进而反映肋软骨组织量的技术方法。通过技术手段创新性地获得了胶片中与真实尺寸相等的10 cm和5 cm比例尺，实现了胶片的等比例打印，并经过对比验证证明胶片上测量值等同于人体真实值。该技术获得的胶片可以为术前设计和术中测量提供有力的帮助。
因条件所限，本研究只从PHILIPS 64排CT入手，未涉及其他品牌的CT机，在后续工作中会尽可能地在更多品牌的设备上进行该方法的实践和验证，以将此方法做更广泛的推广。
-
![]()
图 1 （a）扩散模型起源演变的时间轴；（b）按照关键词（diffusion model | medical imaging）、（score-based model | medical imaging）、（diffusion | medical | probabilistic model）在Google Scholar和Arxiv Sanity Preserver进行近5年相关论文检索，筛选并删除相同的结果最终统计扩散模型应用于不同成像模态分类的论文比例
Figure 1. (a) Timeline of the origin and evolution of the diffusion model; (b) Relevant papers published within the last five years found via Google Scholar and Arxiv Sanity Preserver searches using the keywords (diffusion model | medical imaging), (score-based model | medical imaging), and (diffusion | medical | probabilistic model). Duplicate results were identified and deleted before calculating the proportion of papers applying a diffusion model to different imaging modality classifications
![]()
图 2 扩散模型应用于不同成像模态的论文分类
Figure 2. Classification of papers describing the application of diffusion models to different imaging modalities
![]()
图 3 扩散模型分别在图像域和K空间域的基本框架。其中前向过程从初始数据开始逐步添加噪声，使其最终接近纯噪声分布；逆向过程通过学习逐步去噪，从随机噪声中还原为高质量的目标数据
Figure 3. Basic framework of the diffusion model in the image and K-space domain. The forward process gradually adds noise from the initial data to approach a pure noise distribution; the reverse process gradually removes noise through learning and recovers high-quality target data from random noise
![]()
图 4 单线圈脑部数据集在伪径向采样模式下加速因子为5时不同方法的重建结果比较。绿色框和红色框分别表示感兴趣区域及其残差图
Figure 4. Comparison of reconstruction results using different methods on a single-coil brain dataset in pseudo-radial sampling mode with an acceleration factor of 5. The green and red boxes indicate the region of interest and its residual map, respectively
![]()
图 5 不同方法生成的CT图像的数值比较
Figure 5. Numerical comparison of CT images generated by different methods
![]()
图 6 不同算法在泊松采样模式下加速因子为4的重建结果比较
Figure 6. Comparison of reconstruction results using different algorithms with an acceleration factor of 4 in Poisson sampling mode
![]()
图 7 随机采样模式下加速因子为5和均匀采样模式下加速因子为4的fastMRI膝关节和SIAT脑部重建结果
Figure 7. Reconstruction results of fastMRI knee joint and SIAT brain imaging with an acceleration factor of 5 in random sampling mode and an acceleration factor of 4 in uniform sampling mode
![]()
图 8 不同算法下人体CT扫描有限角度（60°和90°）重建结果（右下角数值为PSNR值）
Figure 8. Reconstruction results of limited angles (60° and 90°) of human CT scan data using different algorithms (the value in each lower right corner is the PSNR value)
![]()
图 9 5 e3剂量下CT的重建结果及残差图
Figure 9. Reconstruction results and residual CT images at a dose of 5 e3
![]()
图 10 稀疏角度为32°时重建血管的过程（数字分别代表PSNR值和SSIM值）
Figure 10. Process of blood vessel reconstruction when the sparse angle is 32° (The numbers represent PSNR and SSIM values, respectively)
表 1 五大生成模型全方位对比
Table 1 Comprehensive comparison of five generative models
特点/模型 VAE GAN EBM Flow Diffusion Model 基本原理 由编码器将数据映射到潜在空间概率分布，再由解码器生成数据 生成器和判别器对抗训练，生成器生成数据，判别器区分真假 定义一个可微的能量函数，将数据点的概率分布与其能量值联系 通过可逆变换将简单分布映射到复杂数据分布。 从噪声分布起，逐步去噪恢复目标数据 优点 1.生成质量高
2.训练稳定
3.潜在空间连续1.生成质量高
2.多样性好
3.应用广泛1.灵活性高
2.隐式生成
3.表示能力强1.高效样本生成和密度估计
2.可解释性强1.生成质量高
2.强大的建模能力
3.广泛的应用场景缺点 1.生成样本模糊
2.计算复杂度高
3.难捕捉复杂分布1.训练困难
2.对数据敏感
3.计算资源消耗大1.配分函数难计算
2.训练不稳定
3.采样效率低1.设计合适的变换
2.模块具有挑战性
3.计算资源需求高1.训练过程复杂
2.对噪声模型依赖性
3.生成速度较慢训练稳定性 稳定 不稳定 稳定 稳定 稳定 生成质量 较高 高 较高 高 高 计算资源需求 中等 高 中等 高 高 模型复杂度 中等 高 中等 高 高 可解释性 较好 较差 较好 好 较差 灵活性 较低 高 高 高 高 对数据分布假设 较强 较弱 较强 较强 较弱 数据质量敏感性 较低 高 较低 较低 较低 生成速度 中等 中等 中等 中等 较慢 
下载: 导出CSV
表 2 不同重建算法结果定量比较
Table 2 Quantitative comparison between the results of different reconstruction algorithms
ESPIRiT LINDBERG EBMRec SAKE WKGM SVD-WKGM T1 GE brain 2D random R=4 39.08/0.933 38.98/0.961 40.17/0.968 41.54/0.952 40.67/0.969 43.85/0.970 2D random R=6 36.01/0.921 35.16/0.958 36.55/0.952 38.09/0.932 37.14/0.957 39.94/0.960 T2 transverse brain 2D Poisson R=4 31.74/0.819 32.87/0.901 33.19/0.915 33.91/0.896 33.35/0.907 34.58/0.917 2D Poisson R=10 28.95/0.798 26.17/0.822 29.59/0.839 29.75/0.823 29.17/0.823 31.69/0.841
下载: 导出CSV
-
[1] BAO J M, CHEN D, WEN F, et al. CVAE-GAN: Fine-grained image generation through asymmetric training[C]//Proceedings of the IEEE international conference on computer vision, Venice, Italy, 2017: 2745-2754. DOI: 10.1109/iccv.2017.299.
[2] RAZAVI A, Van den OORD A, VINYALS O. Generating diverse high-fidelity images with VQ-VAE-2[J]. Advances in neural information processing systems, 2019, 32. DOI: 10.48550/arXiv.1906.00446.
[3] KONG Z F, PING W, HUANG J J, et al. Diffwave: A versatile diffusion model for audio synthesis[J]. arXiv Preprint arXiv: 2009.09761, 2020.
[4] OORD A, DIELEMAN S, ZEN H, et al. Wavenet: A generative model for raw audio[J]. arXiv Preprint arXiv: 1609.03499, 2016.
[5] LI X, THICKSTUN J, GULRAJANI I, et al. Diffusion-lm improves controllable text generation[J]. Advances in Neural Information Processing Systems, 2022, 35: 4328-4343. DOI: 10.48550/arXiv.2205.14217.
[6] YANG G D, HUANG X, HAO Z K, et al. Pointflow: 3D point cloud generation with continuous normalizing flows[C]//Proceedings of the IEEE/CVF international conference on computer vision, Seoul, Korea (South), 2019: 4541-4550. DOI: 10.48550/arXiv.1906.12320.
[7] BOND-TAYLOR S, LEACH A, LONG Y, et al. Deep generative modelling: A comparative review of VAEs, GANs, normalizing flows, energy-based and autoregressive models[J]. IEEE transactions on pattern analysis and machine intelligence, 2021. DOI: 10.48550/arXiv.2103.04922.
[8] GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[J]. Advances in neural information processing systems, 2014, 27. DOI: 10.48550/arXiv.1406.2661.
[9] JIMENEZ REZENDE D, MOHAMED S, WIERSTRA D. Stochastic backpropagation and approximate inference in deep generative models[J]. arXiv Preprint arXiv: 1401.4082, 2014.
[10] DINH L, SOHL-DICKSTEIN J, BENGIO S. Density estimation using real nvp[J]. arXiv Preprint arXiv: 1605.08803, 2016.
[11] SONG Y, SOHL-DICKSTEIN J, KINGMA D P, et al. Score-based generative modeling through stochastic differential equations[J]. arXiv Preprint arXiv: 2011.13456, 2020.
[12] KARRAS T, AITTALA M, AILA T, et al. Elucidating the design space of diffusion-based generative models[J]. Advances in Neural Information Processing Systems, 2022, 35: 26565-26577. DOI: 10.48550/arXiv.2206.00364.
[13] SOHL-DICKSTEIN J, WEISS E, MAHESWARANATHAN N, et al. Deep unsupervised learning using nonequilibrium thermodynamics[C]//International Conference on Machine Learning, Lille, France, 2015: 2256-2265. DOI: 10.48550/arXiv.1503.03585.
[14] SONG Y, DURKAN C, MURRAY I, et al. Maximum likelihood training of score-based diffusion models[J]. Advances in Neural Information Processing Systems, 2021, 34: 1415-1428. DOI: 10.48550/arXiv.2101.09258.
[15] SINHA A, SONG J M, MENG C L, et al. D2c: Diffusion-decoding models for few-shot conditional generation[J]. Advances in Neural Information Processing Systems, 2021, 34: 12533-12548. DOI: 10.48550/arXiv.2106.06819.
[16] VAHDAT A, KREIS K, KAUTZ J. Score-based generative modeling in latent space[J]. Advances in Neural Information Processing Systems, 2021, 34: 11287-11302. DOI: 10.48550/arXiv.2106.05931.
[17] SAHARIA C, HO J, CHAN W, et al. Image super-resolution via iterative refinement[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 45(4): 4713-4726. DOI: 10.48550/arXiv.2104.07636.
[18] PANDEY K, MUKHERJEE A, RAI P, et al. Diffusevae: Efficient, controllable and high-fidelity generation from low-dimensional latents[J]. arXiv Preprint arXiv: 2201.00308, 2022.
[19] BAO F, LI C X, ZHU J, et al. Analytic-dpm: An analytic estimate of the optimal reverse variance in diffusion probabilistic models[J]. arXiv preprint arXiv: 2201.06503, 2022.
[20] DOCKHORN T, VAHDAT A, KREIS K. Score-based generative modeling with critically-damped Langevin diffusion[J]. arXiv preprint arXiv: 2112.07068, 2021.
[21] LIU N, LI S, DU Y L, et al. Compositional visual generation with composable diffusion models[C]//European Conference on Computer Vision, Tel-Aviv, Israel, 2022: 423-439. DOI: 10.48550/arXiv.2206.01714.
[22] JIANG Y M, YANG S, QIU H N, et al. Text2human: Text-driven controllable human image generation[J]. ACM Transactions on Graphics (TOG), 2022, 41(4): 1-11. DOI: 10.48550/arXiv.2205.15996.
[23] BATZOLIS G, STANCZUK J, SCHONLIEB C B, et al. Conditional image generation with score-based diffusion models[J]. arXiv preprint arXiv: 2111.13606, 2021.
[24] DANIELS M, MAUNU T, HAND P. Score-based generative neural networks for large-scale optimal transport[J]. Advances in neural information processing systems, 2021, 34: 12955-12965. DOI: 10.48550/arXiv.2110.03237.
[25] CHUNG H, SIM B, YE J C. Come-closer-diffuse-faster: Accelerating conditional diffusion models for inverse problems through stochastic contraction[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 12413-12422. DOI: 10.48550/arXiv.2112.05146.
[26] KAWAR B, ELAD M, ERMON S, et al. Denoising diffusion restoration models[J]. Advances in Neural Information Processing Systems, 2022, 35: 23593-23606. DOI: 10.48550/arXiv.2201.11793.
[27] ESSER P, ROMBACH R, BLATTMANN A, et al. Imagebart: Bidirectional context with multinomial diffusion for autoregressive image synthesis[J]. Advances in Neural Information Processing Systems, 2021, 34: 3518-3532. DOI: 10.48550/arXiv.2108.08827.
[28] LUGMAYR A, DANELLJAN M, ROMERO A, et al. Repaint: Inpainting using denoising diffusion probabilistic models[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 11461-11471. DOI: 10.48550/arXiv.2201.09865.
[29] JING B, CORSO G, BERLINGHIERI R, et al. Subspace diffusion generative models[C]//European Conference on Computer Vision, Tel-Aviv, Israel, 2022: 274-289. DOI: 10.1007/978-3-031-20050-2_17.
[30] BARANCHUK D, RUBACHEV I, VOYNOV A, et al. Label-efficient semantic segmentation with diffusion models[J]. arXiv preprint arXiv: 2112.03126, 2021.
[31] GRAIKOS A, MALKIN N, JOJIC N, et al. Diffusion models as plug-and-play priors[J]. Advances in Neural Information Processing Systems, 2022, 35: 14715-14728. DOI: 10.48550/arXiv.2206.09012.
[32] WOLLEB J, SANDKÜHLER R, BIEDER F, et al. Diffusion models for implicit image segmentation ensembles[C]//International Conference on Medical Imaging with Deep Learning, New Orleans, USA, 2022: 1336-1348. DOI: 10.48550/arXiv.2112.03145.
[33] AMIT T, SHAHARBANY T, NACHMANI E, et al. Segdiff: Image segmentation with diffusion probabilistic models[J]. arXiv preprint arXiv: 2112.00390, 2021.
[34] ZIMMERMANN R S, SCHOTT L, SONG Y, et al. Score-based generative classifiers[J]. arXiv preprint arXiv: 2110.00473, 2021.
[35] PINAYA W H L, GRAHAM M S, GRAY R, et al. Fast unsupervised brain anomaly detection and segmentation with diffusion models[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention, New Orleans, USA, 2022: 705-714. DOI: 10.48550/arXiv.2206.03461.
[36] WOLLEB J, BIEDER F, SANDKUHLER R, et al. Diffusion models for medical anomaly detection[C]//International Conference on Medical image computing and computer-assisted intervention, New Orleans, USA, 2022: 35-45. DOI: 10.48550/arXiv.2203.04306.
[37] WYATT J, LEACH A, SCHMON S M, et al. Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 650-656. DOI: 10.1109/CVPRW56347.2022.00080.
[38] JALAL A, ARVINTE M, DARAS G, et al. Robust compressed sensing MRI with deep generative priors[J]. Advances in Neural Information Processing Systems, 2021, 34: 14938-14954. DOI: 10.48550/arXiv.2108.01368.
[39] CHUNG H, YE J C. Score-based diffusion models for accelerated MRI[J]. Medical image analysis, 2022, 80: 102479. DOI: 10.48550/arXiv.2110.05243.
[40] CAO H Q, TAN C, GAO Z Y, et al. A survey on generative diffusion model[J]. arXiv preprint arXiv: 2209.02646, 2022.
[41] YANG L, ZHANG Z L, SONG Y, et al. Diffusion models: A comprehensive survey of methods and applications[J]. ACM Computing Surveys, 2023, 56(4): 1-39. DOI: 10.48550/arXiv.2209.00796.
[42] KAZEROUNI A, AGHDAM E K, HEIDARI M, et al. Diffusion models in medical imaging: A comprehensive survey[J]. Medical Image Analysis, 2023: 102846. DOI: 10.1016/jmedia.2023.102846.
[43] VINCENT P, LAROCHELLE H, BENGIO Y, et al. Extracting and composing robust features with denoising autoencoders[C]//Proceedings of the 25th international conference on Machine learning, Helsinki, Finland, 2008: 1096-1103. DOI: 10.1145/1390156.1390294.
[44] VINCENT P, LAROCHELLE H, LAJOIE I, et al. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion[J]. Journal of Machine Learning Research, 2010, 11(12). DOI: 10.5555/1756006.1953039.
[45] VINCENT P. A connection between score matching and denoising autoencoders[J]. Neural computation, 2011, 23(7): 1661-1674. DOI: 10.1162/NECO_a_00142.
[46] SONG Y, ERMON S. Generative modeling by estimating gradients of the data distribution[J]. Advances in Neural Information Processing Systems, 2019, 32. DOI: 10.48550/arXiv.1907.05600.
[47] SONG Y, GARG S, SHI J, et al. Sliced score matching: A scalable approach to density and score estimation[C]//Proceedings of The 35th Uncertainty in Artificial Intelligence Conference, Tel Aviv, Israel, 2020: 574-584. DOI: 10.48550/arXiv.1905.07088.
[48] HO J, JAIN A, ABBEEL P. Denoising diffusion probabilistic models[J]. Advances in Neural Information Processing Systems, 2020, 33: 6840-6851. DOI: 10.48550/arXiv.2006.11239.
[49] ERHAN D, COURVILLE A, BENGIO Y, et al. Why does unsupervised pre-training help deep learning?[C]//Proceedings of the 13th International Conference on Artificial Intelligence and Statistics, Pittsburgh, USA, 2010: 201-208.
[50] Le CUN Y, FOGELMAN-SOULIE F. Modeles connexionnistes de l’apprentissage[J]. Intellectica, 1987, 2(1): 114-143. DOI: 10.3406/intel.1987.1804.
[51] GALLINARI P, LECUN Y, THIRIA S, et al. Distributed associative memories: A comparison[M]. La Villette, Paris: Proceedings of Cognitiva 87, 1987.
[52] SEUNG H S. Learning continuous attractors in recurrent networks[J]. Advances in Neural Information Processing Systems, 1997, 10. DOI: 10.5555/3008904.3008997.
[53] TEZCAN K C, BAUMGARTNER C F, LUECHINGER R, et al. MR image reconstruction using deep density priors[J]. IEEE Transactions on Medical Imaging, 2018, 38(7): 1633-1642. DOI: 10.1109/TMI.2018.2887072.
[54] LIU Q G, YANG Q X, CHENG H T, et al. Highly under-sampled magnetic resonance imaging reconstruction using autoencoding priors[J]. Magnetic Resonance in Medicine, 2020, 83(1): 322-336. DOI: 10.1002/mrm.27921.
[55] WANG S Y, LV J J, HE Z N, et al. Denoising auto-encoding priors in undecimated wavelet domain for MR image reconstruction[J]. Neurocomputing, 2021, 437: 325-338. DOI: 10.1016/j.neucom.2020.09.086.
[56] LIU X S, ZHANG M H, LIU Q G, et al. Multi-contrast MR reconstruction with enhanced denoising autoencoder prior learning[C]//2020 IEEE 17th International Symposium on Biomedical Imaging, Iowa City, USA, 2020: 1-5. DOI: 10.1109/ISBI45749.2020.9098334.
[57] ZHANG M H, LI M T, ZHOU J J, et al. High-dimensional embedding network derived prior for compressive sensing MRI reconstruction[J]. Medical Image Analysis, 2020, 64: 101717. DOI: 10.1016/j.media.2020.101717.
[58] BLOCK A, MROUEH Y, RAKHLIN A. Generative modeling with denoising auto-encoders and Langevin sampling[J]. arXiv Preprint arXiv: 2002.00107, 2020.
[59] PARISI G. Correlation functions and computer simulations[J]. Nuclear Physics B, 1981, 180(3): 378-384. DOI: 10.1016/0550-3213(81)90056-0.
[60] GRENANDER U, MILLER M I. Representations of knowledge in complex systems[J]. Journal of the Royal Statistical Society: Series B (Methodological), 1994, 56(4): 549-581. DOI: 10.1111/j.2517-6161.1994.tb02000.x.
[61] QUAN C, ZHOU J J, ZHU Y Z, et al. Homotopic gradients of generative density priors for MR image reconstruction[J]. IEEE Transactions on Medical Imaging, 2021, 40(12): 3265-3278. DOI: 10.48550/arXiv.2008.06284.
[62] ZHU W Q, GUAN B, WANG S S, et al. Universal generative modeling for calibration-free parallel MR imaging[C]//2022 IEEE 19th International Symposium on Biomedical Imaging, Kolkata, India, 2022: 1-5. DOI: 10.1109/ISBI52829.2022.9761446.
[63] HE Z N, ZHANG Y K, GUAN Y, et al. Iterative reconstruction for low-dose CT using deep gradient priors of generative model[J]. IEEE Transactions on Radiation and Plasma Medical Sciences, 2022, 6(7): 741-754. DOI: 10.48550/arXiv.2009.12760.
[64] LYU Q, WANG G. Conversion between CT and MRI images using diffusion and score-matching models[J]. arXiv Preprint arXiv: 2209.12104, 2022.
[65] CHUNG H, LEE E S, YE J C. MR image denoising and super-resolution using regularized reverse diffusion[J]. IEEE Transactions on Medical Imaging, 2022, 42(4): 922-934. DOI: 10.48550/arXiv.2203.12621.
[66] MENG X X, GU Y N, PAN Y S, et al. A novel unified conditional score-based generative framework for multi-modal medical image completion[J]. arXiv Preprint arXiv: 2207.03430, 2022.
[67] LUO G, BLUMENTHAL M, HEIDE M, et al. Bayesian MRI reconstruction with joint uncertainty estimation using diffusion models[J]. Magnetic Resonance in Medicine, 2023, 90(1): 295-311. DOI: 10.1002/mrm.29624.
[68] PENG C, GUO P F, ZHOU S K, et al. Towards performant and reliable undersampled MR reconstruction via diffusion model sampling[C]// International Conference on Medical Image Computing and Computer-Assisted Intervention, Singapore, Singapore, 2022: 623-633. DOI: 10.1007/978-3-031-16446-0_59.
[69] GUNGOR A, DAR S U H, OZTURK S, et al. Adaptive diffusion priors for accelerated MRI reconstruction[J]. Medical Image Analysis, 2023: 102872. DOI: 10.1016/j.media.2023.102872.
[70] XIE Y T, LI Q Z. Measurement-conditioned denoising diffusion probabilistic model for under-sampled medical image reconstruction[J]. arXiv Preprint arXiv: 2203.03623, 2022.
[71] OZTURKLER B, LIU C, ECKART B, et al. SMRD: Sure-based robust MRI reconstruction with diffusion models[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention, Vancouver, Canada, 2023: 199-209. DOI: 10.1007/978-3-031-43898-1_20.
[72] MIRZA M U, DALMAZ O, BEDEL H A, et al. Learning fourier-constrained diffusion bridges for MRI reconstruction[J]. arXiv Preprint arXiv: 2308.01096, 2023.
[73] BIAN W Y, JANG A, LIU F. Diffusion modeling with domain-conditioned prior guidance for accelerated MRI and qMRI reconstruction[J]. arXiv Preprint arXiv: 2309.00783, 2023.
[74] KORKMAZ Y, CUKUR T, PATEL V M. Self-supervised MRI reconstruction with unrolled diffusion models[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention, Vancouver, Canada, 2023: 491-501. DOI: 10.1007/978-3-031-43999-5_47.
[75] CHEN L X, TIAN X Y, WU J J, et al. JSMoCo: Joint coil sensitivity and motion correction in parallel MRI with a self-calibrating score-based diffusion model[J]. arXiv Preprint arXiv: 2310.09625, 2023.
[76] RAVULA S, LEVAC B, JALAL A, et al. Optimizing sampling patterns for compressed sensing MRI with diffusion generative models[J]. arXiv Preprint arXiv: 2306.03284, 2023.
[77] SONG Y, SHEN L Y, XING L, et al. Solving inverse problems in medical imaging with score-based generative models[J]. arXiv Preprint arXiv: 2111.08005, 2021.
[78] CHUNG H, SIM B, RYU D, et al. Improving diffusion models for inverse problems using manifold constraints[J]. arXiv Preprint arXiv: 2206.00941, 2022.
[79] TU Z J, JIANG C, GUAN Y, et al. K-space and image domain collaborative energy-based model for parallel MRI reconstruction[J]. Magnetic Resonance Imaging, 2023, 99: 110-122. DOI: 10.1016/j.mri.2023.02.004.
[80] PENG H, JIANG C, CHENG J, et al. One-shot generative prior in hankel-k-space for parallel imaging reconstruction[J]. IEEE Transactions on Medical Imaging, 2023. DOI: 10.1109/TMI.2023.3288219.
[81] YU C M, GUAN Y, KE Z W, et al. Universal generative modeling in dual domains for dynamic MRI[J]. NMR in Biomedicine, 2023, 36(12): e5011. DOI: 10.1002/nbm.5011.
[82] GUAN Y, YU C M, LU S Y, et al. Correlated and multi-frequency diffusion modeling for highly under-sampled MRI reconstruction[J]. arXiv Preprint arXiv: 2309.00853, 2023.
[83] ZHANG W, XIAO Z W, TAO H, et al. Low-rank tensor assisted K-space generative model for parallel imaging reconstruction[J]. Magnetic Resonance Imaging, 2023, 103: 198-207. DOI: 10.1016/j.mri.2023.07.004.
[84] CUI Z X, CAO CT, LIU S N, et al. Self-score: Self-supervised learning on score-based models for MRI reconstruction[J]. arXiv Preprint arXiv: 2209.00835, 2022.
[85] CUI Z X, LIU C C, FAN X H, et al. Physics-informed deep MRI: K-space interpolation meets heat diffusion[J]. IEEE Transactions on Medical Imaging, 2024, 43(10): 3503-3520. DOI: 10.1109/TMI.2024.3462988.
[86] CAO C T, CUI Z X, LIU S N, et al. High-frequency space diffusion models for accelerated MRI[J]. arXiv Preprint arXiv: 2208.05481, 2022.
[87] GUAN Y, YU C M, CUI Z X, et al. Correlated and multi-frequency diffusion modeling for highly under-sampled MRI reconstruction[J]. IEEE Transactions on Medical Imaging, 2024, 43(10): 3490-3502. DOI: 10.1109/TMI.2024.3381610.
[88] HUANG B, LU S Y, LIU Q G, et al. One-sample diffusion modeling in projection domain for low-dose CT imaging[J]. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. DOI: 10.1109/TRPMS.2024.3392248.
[89] WU W W, WANG Y Y. Data-iterative optimization score model for stable ultra-sparse-view CT reconstruction[J]. arXiv Preprint arXiv: 2308.14437, 2023.
[90] ZHANG W H, HUANG B, LIU Q G, et al. Low-rank angular prior guided multi-diffusion model for few-shot low-dose CT reconstruction[J]. IEEE Transactions on Radiation and Plasma Medical Sciences, 2024. DOI: 10.1109/TCI.2024.3503366.
[91] LIU J M, ANIRUDH R, THIAGARAJAN J J, et al. DOLCE: A model-based probabilistic diffusion framework for limited-angle CT reconstruction[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision, Paris, France, 2023: 10498-10508. DOI: 10.48550/arXiv.2211.12340.
[92] XIA W J, SHI Y Y, NIU C, et al. Diffusion prior regularized iterative reconstruction for low-dose CT[J]. arXiv Preprint arXiv: 2310.06949, 2023.
[93] XIA W J, CONG W X, WANG G. Patch-based denoising diffusion probabilistic model for sparse-view CT reconstruction[J]. arXiv Preprint arXiv: 2211.10388, 2022.
[94] XIA W J, NIU C, CONG W X, et al. Sub-volume-based denoising diffusion probabilistic model for cone-beam CT reconstruction from incomplete data[J]. arXiv e-Prints, 2023: arXiv: 2303.12861.
[95] GAO Q, LI Z L, ZHANG J P, et al. CoreDiff: Contextual error-modulated generalized diffusion model for low-dose CT denoising and generalization[J]. arXiv Preprint arXiv: 2304.01814, 2023.
[96] WU W W, WANG Y Y, LIU Q G, et al. Wavelet-improved score-based generative model for medical imaging[J]. IEEE Transactions on Medical Imaging, 2023. DOI: 10.1109/TMI.2023.3325824.
[97] XU K, LU S Y, HUANG B, et al. Stage-by-stage wavelet optimization refinement diffusion model for sparse-view CT reconstruction[J]. arXiv Preprint arXiv: 2308.15942, 2023.
[98] ZHANG J J, MAO H Y, WANG X R, et al. Wavelet-inspired multi-channel score-based model for limited-angle CT reconstruction[J]. IEEE Transactions on Medical Imaging, 2024, 43(10): 3436-3448. DOI: 10.1109/TMI.2024.3367167.
[99] GUAN B, YANG C L, ZHANG L, et al. Generative modeling in sinogram domain for sparse-view CT reconstruction[J]. IEEE Transactions on Radiation and Plasma Medical Sciences, 2023. DOI: 10.1109/TRPMS.2023.3309474.
[100] LIU X, XIE Y Q, DIAO S H, et al. Unsupervised CT metal artifact reduction by plugging diffusion priors in dual domains[J]. IEEE Transactions on Medical Imaging, 2024, 43(10): 3533-3545. DOI: 10.1109/TMI.2024.3351201.
[101] WANG Y Y, LI Z R, WU W W. Time-reversion fast-sampling score-based model for limited-angle CT reconstruction[J]. IEEE Transactions on Medical Imaging, 2024, 43(10): 3449-3460. DOI: 10.1109/TMI.2024.3418838.
[102] WU W W, PAN J Y, WANG Y Y, et al. Multi-channel optimization generative model for stable ultra-sparse-view CT reconstruction[J]. IEEE Transactions on Medical Imaging, 2024, 43(10): 3416-3475. DOI: 10.1109/TMI.2024.3376414.
[103] HAN Z Y, WANG Y H, ZHOU L P, et al. Contrastive diffusion model with auxiliary guidance for coarse-to-fine PET reconstruction[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention, Vancouver, Canada, 2023: 239-249. DOI: 10.1007/978-3-031-43999-5_23.
[104] JIANG C W, PAN Y S, LIU M X, et al. PET-diffusion: Unsupervised PET enhancement based on the latent diffusion model[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention, Vancouver, Canada, 2023: 3-12. DOI: 10.1007/978-3-031-43907-0_1.
[105] XIE T F, CAO C T, CUI Z X, et al. Brain pet synthesis from MRI using joint probability distribution of diffusion model at ultrahigh fields[J]. arXiv Preprint arXiv: 2211.08901, 2022.
[106] XIE T F, CAO C R, CUI Z X, et al. Synthesizing PET images from high-field and ultra-high-field MR images using joint diffusion attention model[J]. arXiv Preprint arXiv: 2305.03901, 2023.
[107] XIE T F, CUI Z X, LUO C, et al. Joint diffusion: Mutual consistency-driven diffusion model for PET-MRI co-reconstruction[J]. arXiv Preprint arXiv: 2311.14473, 2023.
[108] PAN S Y, ABOUEI E, PENG J B, et al. Full-dose PET synthesis from low-dose PET using high-efficiency diffusion denoising probabilistic model[J]. arXiv Preprint arXiv: 2308.13072, 2023.
[109] SONG X L, WANG G J, ZHONG W H, et al. Sparse-view reconstruction for photoacoustic tomography com-bining diffusion model with model-based iteration[J]. Photoacoustics, 2023, 33: 100558. DOI: 10.1016/j.pacs.2023.100558.
[110] TONG S Q, LAN H R, NIE L M, et al. Score-based generative models for photoacoustic image reconstruction with rotation consistency constraints[J]. arXiv Preprint arXiv: 2306.13843, 2023.
[111] DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[J]. arXiv Preprint arXiv: 2010.11929, 2020.
[112] WANG Z H, YANG Y Y, CHEN Y Z, et al. Mutual information guided diffusion for zero-shot cross-modality medical image translation[J]. IEEE Transactions on Medical Imaging, 2024, 43(8): 2825-2838. DOI: 10.1109/TMI.2024.3382043.
[113] CHU J Y, DU C H, LIN X Y, et al. Highly accelerated MRI via implicit neural representation guided posterior sampling of diffusion models[J]. Medical Image Analysis, 2025, 100.
计量
- 文章访问数: 1241
- HTML全文浏览量: 122
- PDF下载量: 204
