Next generation of ultrasound imaging using ultrafast acquisition and machine learning

PhD项目的目标：该项目旨在开发先进的生物医学超声图像重建技术，利用超快超声采集产生的大量数据，以及用于超快超声和3D采集所采集的多GB/s数据的快速/实时数据处理的机器学习算法，2D和3D中的新型图像重建技术，使用机器学习和成像物理学知识来实现前所未有的图像质量扩展机器学习算法以具有时间分量，利用超快速采集获得的极高时间分辨率数据，探索该技术在心血管疾病中的应用。疾病和癌症项目描述/背景：在过去的十年中，由于电子、计算和换能器技术的进步，超快和3D超声技术正在迅速扩展生物医学超声领域。虽然超快采集技术为更好的图像质量和信息内容提供了令人兴奋的机会，但仍然存在重大挑战。首先，数据量是显著的（每秒多个GB），并且现有方法的计算成本阻止了它们的真实的实现。其次，即使具有超快能力，采集和处理策略的原理仍然在很大程度上依赖于经典方法，与标准超声相比，这些方法只能产生图像质量的边际改善。该项目的目的是通过在整个图像形成链中开发基于深度学习模型的方法来设计和评估新型图像重建和数据处理技术，以便显著地加速成像并提高图像质量。超声领域现有的机器学习研究大多集中在图像后处理，而深度学习模型和算法在图像重建中的应用在很大程度上是一个未开发的领域。作为一个路线图，在这个项目中，学生将：探索现有的先进图像重建算法的前景，这些算法比经典方法生成的图像质量更好，但目前太慢。这些先进的算法包括最小方差方法，相干因子，稀疏正则化。这些方法都遭受非常缓慢的重建，目前不适合临床使用。在对各种方法进行评估后，学生将设计神经网络架构，通过实时提供最佳的采集和重建参数来加速这些方法。探索使用深度学习和我们的物理知识（例如使用声波模拟）来直接重建具有上级图像质量的图像。代替依赖于简单的几何声学近似，如目前所做的，它是可能的，使用声学传播的物理和设计重建策略的基础上更复杂和现实的模型。深度神经网络将用于通过结合物理模拟的端到端训练将测量数据的反演规则化为理想图像。压缩的问题，以一个合适的空间也将进行探讨，以确保计算负担兼容的真实的时间应用。

Aim of the PhD Project:The project aims to develop advanced biomedical ultrasound image reconstruction technology, taking advantage of the large amount of data generated by ultrafast ultrasound acquisition, and machine learning algorithms forfast/real-time data processing of the multiple GB/s data acquired by ultrafast ultrasound and 3D acquisitions,novel image reconstruction technologies in both 2D and 3D, using machine learning and knowledge of the imaging physics to achieve unprecedented image qualityexpanding the machine learning algorithms to have temporal components, taking advantage of the very high temporal resolution data obtained by ultrafast acquisitionexplore the applications of the techniques in cardiovascular disease and cancerProject Description / Background:In the last decade, ultrafast and 3D ultrasound techniques are rapidly expanding fields in biomedical ultrasound thanks to the advances in electronics, computing and transducer technologies. While ultrafast acquisition technologies offer exciting opportunities for better image quality and information content, significant challenges still exist. Firstly, the amount of data is significant (multiple GBs per second) and the computational cost of the existing approaches prevents their real time implementation. Secondly even with the ultrafast capability the principles of acquisition and processing strategies still largely rely on classical approaches, which only produce a marginal improvement of image quality compared to standard ultrasound.The aim of the project is to design and evaluate novel image reconstruction and data processing technologies by developing deep learning model based approaches throughout the image formation chain, in order to significantly speed up the imaging and improve image quality. Most existing machine learning studies in the field of ultrasound have been focused on post image-processing, and the application of deep learning models and algorithms for image reconstruction is largely an unexplored area.As a roadmap, in this project the student will:Explore the landscape of existing advanced image reconstruction algorithms for US which generate better image quality than the classic approach, but currently are too slow. These advanced algorithms include the Minimum Variance methods, Coherence factor, sparse regularization. Such approaches all suffer from very slow reconstruction and currently not suitable for clinical use. After evaluation of the various methods, the student will design neural network architectures that can speed up such methods by providing in real-time optimal acquisition and reconstruction parameters.Explore the use of deep learning and our knowledge of physics (e.g. using acoustic wave simulation) to directly reconstruct images with superior image quality. Instead of relying on the simple geometrical acoustic approximation as currently done, it is possible to use the physics of the acoustic propagation and design reconstruction strategies based on more complex and realistic models. Deep neural networks will be used to regularize the inversion of measured data to ideal images via end to end training incorporating the simulation of physics. Compression of the problem to a suitable space will also be explored, to ensure a computational burden compatible with real time application.