Real-Time Freehand Ultrasound Molecular Imaging with Deep Learning

项目摘要 近年来，早期检测降低了乳腺癌的死亡率;然而，目前的成像工具对乳腺癌的诊断效果不佳。 阳性预测值，并可能导致过度诊断。超声分子成像（UMI）是一种很有前途的 一种工具，可以提供无创、非电离、实时、徒手乳腺癌肿瘤评估， 照顾。UMI使用靶向超声造影剂（UCA）来区分良性和恶性 病变，并有可能减少假阳性率和过度诊断。然而，较差的UMI图像质量 导致研究人员将实时和徒手成像的贝内用于更好的图像质量， 更长的检查时间，更高的UCA剂量，以及可能错过的目标。该项目将开发一种新的实时 基于深度学习的徒手UMI方法，实现乳腺癌的出色检测。 提出了一种时空卷积神经网络（CNN）方法来专门检测粘附的UCA， 其实时指示疾病，同时抑制来自组织和游离UCA的背景噪声。物理学- 将构建乳房UMI驱动模拟器，并使用体模和水箱测量值进行校准。 模拟UMI时间序列的数据库将被组装并用于训练时空CNN以识别 遵守UCA。最后，CNN将部署在一个原型实时徒手UMI系统上， 超过每秒30帧该系统将在UMI乳腺癌检测研究中进行测试， 转基因小鼠模型。本研究将使用靶向高度特异性的新型生物标志物（B7-H3）的UCA。 并被认为与进展为浸润性乳腺癌的可能性密切相关。 实时徒手方法非常重要，因为它使操作员能够自由地询问（和重新访问） 通过单次UCA注射实现多个目标，减少检查时间和UCA剂量。所提出的方法是 创新，因为它使用CNN来更好地利用UMI期间已经获得的数据。所提出的系统 将能够使用B7-H3的实时徒手UMI以高图像质量检测乳腺癌， 减少假阳性和过度诊断，从而减少不必要的测试、活组织检查、成本和患者痛苦。 K99阶段将提供分子成像，UCA合成， 姐妹和靶向、动物研究设计、小动物UMI、临床UMI和其他所需专用制剂 过渡到一个教师的位置和一个独立的研究生涯研究成像方法的UMI。这些 培训活动将提供完成R 00阶段拟定动物研究所需的技能。

Project Summary Earlier detection has reduced breast cancer mortality in recent years; however, current imaging tools have poor positive predictive value and likely contribute to overdiagnosis. Ultrasound molecular imaging (UMI) is a promising tool that can provide noninvasive, non-ionizing, real-time, freehand breast cancer tumor assessment at the point of care. UMI uses targeted ultrasound contrast agents (UCAs) to differentiate between benign and malignant lesions and has the potential to reduce false positive rates and overdiagnosis. However, poor UMI image quality has led researchers to trade the benefits of real-time and freehand imaging for better image quality, resulting in longer exam times, higher UCA dosage, and potentially missed targets. This project will develop a new real-time freehand UMI approach based on deep learning to achieve excellent detection of breast cancer. A spatiotemporal convolutional neural network (CNN) approach is proposed to specifically detect adherent UCAs, which indicate disease, in real-time while suppressing background noise from tissue and free UCAs. A physics- driven simulator of breast UMI will be constructed and calibrated with phantom and water tank measurements. A database of simulated UMI time series will be assembled and used to train the spatiotemporal CNN to identify adherent UCAs. Finally, the CNN will be deployed on a prototype real-time freehand UMI system, operating in excess of 30 frames per second. This system will be tested in a UMI study of breast cancer detection in a transgenic mouse model. This study will use UCAs targeted to a novel biomarker (B7-H3) that is highly specific for cancer and is believed to correlate strongly with likelihood of progression to invasive breast cancer. The real-time freehand approach is significant because it enables the operator to freely interrogate (and revisit) multiple targets with a single UCA injection, reducing exam times and UCA dosage. The proposed approach is innovative because it uses CNNs to better utilize data that is already acquired during UMI. The proposed system will be able to detect breast cancer with high image quality using real-time freehand UMI of B7-H3, potentially reducing false positives and overdiagnosis and thus unnecessary tests, biopsies, costs, and patient distress. The K99 phase will provide dedicated training and career growth opportunities in molecular imaging, UCA synthe- sis and targeting, animal study design, small animal UMI, clinical UMI, and other dedicated preparation needed to transition to a faculty position and an independent research career studying imaging methods for UMI. These training activities will provide the skills necessary to complete the proposed animal study in the R00 phase.