A Machine Learning Alternative to Beamforming to Improve Ultrasound Image Quality for Interventional Access to the Kidney

波束成形的机器学习替代方案可提高肾脏介入治疗的超声图像质量

• 批准号: 10170765
• 负责人: Muyinatu A. Lediju Bell
• 金额: $ 23.5万
• 依托单位: JOHNS HOPKINS UNIVERSITY
• 依托单位国家: 美国
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-12-07 至 2022-04-30
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/10170765
• 关键词: Abdomen; Acoustics; Adult; Age; Anatomy; Area; Award; Breast biopsy; Bypass; Cancer Detection; Child; Clinical; Collaborations; Computer Vision Systems; Computers; Custom; Data; Evaluation; Excision; Family suidae; Funding; Geometry; Goals; Hospitals; Human; Image; Intervention; Interventional Ultrasonography; Kidney; Kidney Calculi; Knowledge; Learning; Location; Machine Learning; Measurement; Metals; Modeling; Morphologic artifacts; Needles; Network-based; Noise; Nonionizing Radiation; North America; Operative Surgical Procedures; Output; Overweight; Pain; Patients; Prevalence; Procedures; Process; Radiology Specialty; Readability; Resolution; Scientist; Signal Transduction; Source; Structure; Techniques; Testing; Time; Tissues; Translations; Ultrasonography; United States National Institutes of Health; Visualization; Work; algorithm training; base; convolutional neural network; cost; image guided; image guided intervention; imaging scientist; improved; in vivo; innovation; instrument; interest; metallicity; obese patients; obese person; radiologist; signal processing; tool

Project Summary Despite the widespread prevalence of ultrasound imaging in hospitals today, the clinical utility of ultrasound guidance is severely hampered by clutter and reverberation artifacts that obscure structures of interest and com- plicate anatomical measurements. Clutter is particularly problematic in overweight and obese individuals, who account for 78.6 million adults and 12.8 million children in North America. Similarly, interventional procedures of- ten require insertion of one or more metal tools, which generate reverberation artifacts that obfuscate instrument location, orientation, and geometry, while obscuring nearby tissues, thus additionally hampering ultrasound im- age quality. Although artifacts are problematic, ultrasound continues to persist primarily because of its greatest strengths (i.e., mobility, cost, non-ionizing radiation, real-time visualization, and multiplanar views) in comparison to existing image-guidance options, but it would be significantly more useful without problematic artifacts. Our long-term project goal is to use state-of-the-art machine learning techniques to provide interventional radiologists with artifact-free ultrasound-based images. We will initially develop a new framework alternative to the ultrasound beamforming process that removes needle tip reverberations and acoustic clutter caused by multipath scattering in near-field tissues when guiding needles to the kidney to enable removal of painful kidney stones. Our first aim will test convolutional neural networks (CNNs) that input raw channel data and output human readable images with no artifacts caused by multipath scattering and reverberations. A secondary goal of the CNNs is to learn the minimum number of parameters required to create these new CNN-based images. Our second aim will validate the trained algorithms with ultrasound data from experimental phantom and ex vivo tissue. Our third aim will extend our evaluation to ultrasound images of in vivo porcine kidneys. This work is the first to propose bypassing the entire beamforming process and replacing it with machine learning and computer vision techniques to remove traditionally problematic noise artifacts and create a fundamentally new type of artifact-free, high-contrast, high-resolution, ultrasound-based image for guiding interventional procedures. This work combines the expertise of an imaging scientist, a computer scientist, and an interventional ra- diologist to explore an untapped, understudied area that is only recently made feasible through improvements in computing power, advances in computer vision capabilities, and new knowledge about dominant sources of image degradation. Translation to in vivo cases is enabled by our clinical collaboration with the Department of Radiology at the Johns Hopkins Hospital. With support from the NIH Trailblazer Award, our team will be the first to develop these tools and capabilities to eliminate noise artifacts in interventional ultrasound, opening the door to a new paradigm in ultrasound image formation, which will directly benefit millions of patients with clearer, easier-to-interpret ultrasound images. Subsequent R01 funding will customize our innovation to addi- tional application-specific ultrasound procedures (e.g., breast biopsies, cancer detection, autonomous surgery).

项目概要 尽管超声成像在当今医院中广泛普及，但超声成像的临床应用 混乱和混响伪影严重阻碍了指导，这些伪影模糊了兴趣和COM的结构。 褶皱解剖测量。对于超重和肥胖的人来说，杂乱问题尤其严重，他们 北美有 7860 万成年人和 1280 万儿童。同样，介入程序- 十种需要插入一个或多个金属工具，这会产生混淆仪器的混响伪影 位置、方向和几何形状，同时遮挡附近的组织，从而额外妨碍超声成像 年龄素质。尽管伪影存在问题，但超声波继续存在主要是因为其最大的优点 比较优势（即移动性、成本、非电离辐射、实时可视化和多平面视图） 现有的图像引导选项，但如果没有有问题的伪影，它会更加有用。 我们的长期项目目标是使用最先进的机器学习技术来提供干预 放射科医生可以使用基于超声的无伪影图像。我们将首先开发一个新的框架替代方案 超声波束形成过程消除了针尖混响和声学杂乱 将针引导至肾脏以切除疼痛的肾脏时，近场组织中的多路径散射 石头。我们的第一个目标是测试输入原始通道数据并输出的卷积神经网络（CNN） 人类可读的图像，没有由多径散射和混响引起的伪影。次要目标 CNN 的目标是学习创建这些新的基于 CNN 的图像所需的最少参数数量。 我们的第二个目标是利用来自实验体模和离体的超声数据来验证经过训练的算法 组织。我们的第三个目标是将我们的评估扩展到体内猪肾脏的超声图像。这部作品是 首先建议绕过整个波束成形过程并用机器学习和计算机代替 视觉技术可以消除传统上有问题的噪声伪影，并创建一种全新的 无伪影、高对比度、高分辨率、基于超声的图像，用于指导介入手术。 这项工作结合了成像科学家、计算机科学家和介入研究人员的专业知识。 放射学家探索一个尚未开发、研究不足的领域，该领域直到最近才通过改进变得可行 计算能力、计算机视觉能力的进步以及有关主要来源的新知识 图像质量下降。通过我们与该部门的临床合作，可以转化为体内病例 约翰·霍普金斯医院放射科。在 NIH 开拓者奖的支持下，我们的团队将 第一个开发这些工具和功能来消除介入超声中的噪声伪影，开放 开启超声图像形成新范式的大门，这将直接造福数百万患有以下疾病的患者 更清晰、更容易解读的超声图像。随后的 R01 资金将定制我们的创新 特定应用的超声程序（例如，乳腺活检、癌症检测、自主手术）。

项目成果

Deep Learning to Obtain Simultaneous Image and Segmentation Outputs From a Single Input of Raw Ultrasound Channel Data.

• DOI: 10.1109/tuffc.2020.2993779
• 发表时间: 2020-12
• 期刊: IEEE transactions on ultrasonics, ferroelectrics, and frequency control
• 作者: Nair AA;Washington KN;Tran TD;Reiter A;Lediju Bell MA
• 通讯作者: Lediju Bell MA

The Generalized Contrast-to-Noise Ratio: A Formal Definition for Lesion Detectability.

• DOI: 10.1109/tuffc.2019.2956855
• 发表时间: 2020-04
• 期刊: IEEE transactions on ultrasonics, ferroelectrics, and frequency control
• 作者: Rodriguez-Molares A;Rindal OMH;D'hooge J;Masoy SE;Austeng A;Lediju Bell MA;Torp H
• 通讯作者: Torp H

CohereNet: A Deep Learning Architecture for Ultrasound Spatial Correlation Estimation and Coherence-Based Beamforming.

• DOI: 10.1109/tuffc.2020.2982848
• 发表时间: 2020-12
• 期刊: IEEE transactions on ultrasonics, ferroelectrics, and frequency control
• 作者: Wiacek A;Gonzalez E;Bell MAL
• 通讯作者: Bell MAL

Detection of COVID-19 features in lung ultrasound images using deep neural networks.

使用深度神经网络检测肺部超声图像中的 COVID-19 特征。

• DOI: 10.1038/s43856-024-00463-5
• 发表时间: 2024
• 期刊: Communications medicine
• 作者: Zhao,Lingyi;Fong,TiffanyClair;Bell,MuyinatuALediju
• 通讯作者: Bell,MuyinatuALediju

Deep Learning for Ultrasound Image Formation: CUBDL Evaluation Framework and Open Datasets.

• DOI: 10.1109/tuffc.2021.3094849
• 发表时间: 2021-12
• 期刊: IEEE transactions on ultrasonics, ferroelectrics, and frequency control
• 作者: Hyun D;Wiacek A;Goudarzi S;Rothlubbers S;Asif A;Eickel K;Eldar YC;Huang J;Mischi M;Rivaz H;Sinden D;van Sloun RJG;Strohm H;Bell MAL
• 通讯作者: Bell MAL