Motion Estimation and Classification in Medical Image Analysis

准确的运动估计和分类在许多医学成像应用中是至关重要的，因为它是诊断和治疗选项的基础。大多数计算机视觉和图像处理方法忽略了时间数据的运动估计方面，它们主要集中在独立分析序列中的图像。这导致了使这些方法适应实际问题的重大问题，因为缺乏时间一致性使得这些方法对于许多现实生活应用程序无效。虽然先验知识的融合在雷达信号处理中能显著提高运动估计精度，但在医学成像中的应用还没有得到很好的研究。此外，在雷达信号处理中推导出的用于估计诸如飞机等物体的运动的数学模型与医学成像应用中产生的组织运动显著偏离。必须使用基本原理推导出新的模型，以准确地描述医学成像应用中发生的运动。最近，人们提出了几种基于深度卷积神经网络的机器学习算法来解决医学图像分割问题。然而，这些方法不提供实现时间上一致的分割的选项。此外，这些方法通常需要大量的训练数据。在接下来的五年里，我计划继续在分析时间医学图像序列这一令人兴奋的领域工作，并打算为运动估计和分类问题开发新的解决方案。提出的研究计划的总体目标是开发新的、自动的和有效的(实时)算法，用于从医学图像扫描中估计和分类组织运动。具体目标包括：i)理论目标：将开发适用于更一般情况的新的跟踪方法。此外，将开发先进的分类方法来区分不同类型的组织或对象运动。这些进展将包括对局部心功能的正常、运动不足、无运动和运动障碍的自动分类。总体目的是开发能够识别运动模式中的细微差别的算法，从而使计算机辅助诊断算法能够尽早发现异常。Ii)实践目标：在实践中，我计划开发数学模型来描述组织运动，以解决基本和具有挑战性的问题，例如，估计和分类心脏左心室的运动模式。拟议的研究计划旨在培养更多的学生，培养出在机器学习和图像处理方面具有独特专业知识的年轻科学家，使他们成为计算机视觉和医学图像分析领域的世界级专家。

Accurate motion estimation and classification is crucial in many medical imaging applications as it serves as the basis for the diagnosis and treatment option. Most computer vision and image processing approaches overlook the motion estimation aspect of temporal data, and they primarily focused on independently analyzing the images in a sequence. This leads to significant problems in adapting these methodologies to practical problems as the lack of temporal consistency render these methods ineffective for many real-life applications. Although integration of prior knowledge has been shown to significantly improve the motion estimation accuracy in radar signal processing, their application has not been studied well for medical imaging. Additionally, the mathematical models derived in radar signal processing for estimating the motions of objects such as aircraft deviate significantly from the tissue motions arising in medical imaging applications. New models have to be derived using fundamental principles to characterize movements that occur in medical imaging applications accurately. Recently, several deep convolutional neural network-based machine learning algorithms have been proposed to address the medical image segmentation problem. However, these methods do not provide an option to achieve temporally consistent segmentation. Additionally, these methods often require a significant amount of training data.  Over the next five years, I plan to continue working on the exciting area of analyzing temporal medical image sequences and intend to develop novel solutions to motion estimation and classification problems. The overall aim of the proposed research program is to develop novel, automated, and efficient (realtime) algorithms for the estimation and classification of tissue motion from medical image scans. Specific objectives include: i) Theoretical objectives: Novel tracking approaches that are applicable under more general circumstances will be developed. Additionally, advanced classification approaches will be developed to discriminate between different types of tissue or object motions. Such developments will include automated classification of normal, hypokinetic, akinetic and dyskinetic motions in regional cardiac function. The overall purpose is to develop algorithms that can identify subtle differences in motion patterns thereby allowing computer-assisted diagnosis algorithms to detect abnormalities as early as possible. ii) Practical objectives: In practice, I plan to develop mathematical models to characterize tissue motion for essential and challenging problems, e.g., estimating and classifying motion patterns for the cardiac left ventricle. The proposed research program intends to train students and produce young scientists with unique expertise in machine learning and image processing to enable them to be world-class experts in the fields of computer vision and medical image analysis.