Making Sense of the Data Trove Hidden in Medical Ultrasound Signals

超声是一种安全、快速、经济的成像手段。虽然它是最受欢迎的医学成像方式之一，但其潜力尚未被充分开发。具体地说，目前利用时间超声数据的处理技术是有限的，尽管其具有非常高的帧速率能力。此外，原始超声数据不适合可视化，因此被转换为通常被称为B型超声的灰度级图像。然而，这种转换是有损的，会破坏原始数据中的大部分信息。这两个因素构成了一个未得到充分利用的信息宝库，我的两项研究努力旨在探索这些信息。 推力1.超声弹性成像使用以高帧速率采集的原始超声信号来揭示组织的临床相关机械特性，这些特性在B型超声图像中通常是看不见的。这是一个临床上和商业上成功的新兴领域，因此，它是一个很好的例子，可以通过更好地利用超声数据来实现什么。然而，从超声信号估计这些机械性能存在几个技术挑战，我的研究计划的第一个重点是使用新型机器学习(ML)技术来解决这些挑战，以进一步提高弹性成像的性能。 后向散射定量超声使用原始超声数据来估计组织特性，例如衰减、后向散射系数和有效散射体大小。这些特性与细胞属性有关，如大小和形状，是非常重要的疾病生物标志物。然而，目前的超声成像技术没有提供这些特性。定量超声中一个尚未解决的重要挑战是其较高的估计方差，这阻碍了该方法的临床应用。我的研究计划的第二个重点是使用新的ML方法来解决这些挑战。 提高超声波能力的重要性有四个方面。首先，在广泛使用超声波的外科干预中，它可以导致更好的诊断和指导。其次，由于超声波是一种廉价的成像手段，它降低了加拿大的医疗成本。第三，由于超声波广泛可用，它可以改善获得医疗保健的机会，特别是在加拿大的偏远地区，那里的患者目前必须飞往较大的城市进行诊断。第四，提取组织属性，如弹性和散射属性，使评估超声图像的主观性减少，并减少咨询专家临床医生的需要，潜在地进一步降低了医疗成本并提高了可及性。 随着最近医疗保健和ML领域的技术进步，对该领域专业人员的需求已达到顶峰。拟议的研究计划将有助于培训高素质的人才，帮助加拿大保持在医学成像和ML领域的领先地位。

Ultrasound is a safe, fast and cost-effective imaging modality. Although it is one of the most popular medical imaging modalities, its potential has not been fully explored. Specifically, current processing techniques that utilize temporal ultrasound data are limited, despite its very high frame-rate capability. In addition, raw ultrasound data is not suitable for visualization, and as such, is converted to gray-scale images that are commonly referred to as B-mode ultrasound. This conversion, however, is lossy and destroys most of the information in raw data. These two factors constitute a trove of underutilized information that my two research thrusts aim to explore. Thrust 1. Ultrasound elastography uses raw ultrasound signals acquired at high frame-rates to reveal clinically relevant mechanical properties of the tissue, which are often invisible in the B-mode ultrasound images. It is a clinically and commercially successful emerging field, and as such, it is a good example of what can be achieved by better exploiting ultrasound data. However, there are several technical challenges in estimating these mechanical properties from ultrasound signals, and the first thrust of my research program aims to tackle them using novel Machine Learning (ML) techniques to further enhance the performance of elastography. Thrust 2. Backscatter quantitative ultrasound uses raw ultrasound data to estimate tissue properties such as attenuation, backscattering coefficients and effective scatterer size. These properties are related to cell attributes, such as size and shape, and are very important biomarkers of disease. However, current ultrasound imaging technology does not provide these properties. An important unresolved challenge in quantitative ultrasound is its high estimation variance, which has hindered clinical utility of this approach. The second thrust of my research program focuses on solving these challenges using novel ML methods. The importance of improving the capabilities of ultrasound is fourfold. First, it can lead to better diagnosis and guidance of surgical interventions, where ultrasound is extensively used. Second, as ultrasound is an inexpensive imaging modality, it reduces the cost of healthcare in Canada. Third, as ultrasound is widely available, it can improve access to healthcare especially in remote regions of Canada where patients currently have to fly to larger cities for diagnosis. And fourth, extracting tissue properties such as elasticity and scattering properties makes assessing ultrasound images less subjective, and reduces the need to consult expert clinicians, potentially further reducing the healthcare cost and improving accessibility. With recent technological advancement in healthcare and ML, demand for professionals in the field has peaked. The proposed research program will contribute to training of highly qualified personnel to help Canada maintain a leading role in medical imaging and ML.