Towards Motion-Robust and Efficient Functional MRI Using Implicit Function Learning

功能性磁共振成像（fMRI）是测量大脑活动和连接的主要方式。在临床上，它开始用于术前计划和评估植物人状态患者的脑功能。它最近也在婴儿认知研究中显示出希望，这是理解人类大脑起源和功能的关键。然而，限制fMRI临床应用的主要挑战之一是其对运动的敏感性，其中头部运动导致fMRI数据中高度有害的伪影，并且可能是功能连接分析中的主要错误来源。这对2至48个月的婴儿来说尤其具有挑战性，在许多情况下，由于头部移动，一半的数据被丢弃，导致重复扫描的显着延迟和成本。对它的需求持续增加将导致医院资源的压力进一步增加，成像工作流程的效率降低。因此，迫切需要技术和工具来消除或减少fMRI扫描中的运动影响。最近，机器学习（ML）和深度学习（DL）技术已经显示出通过从数据中学习来回顾性地纠正运动和伪影，从而减轻运动损坏的希望。然而，这些基于学习的方法中的大多数并不特别关注fMRI，并且由于fMRI中的高存储器要求和应用特定的运动伪影，大多数现有的用于静态和结构MRI的运动校正方法并不直接适用于fMRI。因此，仍然缺乏针对该问题的鲁棒且可靠的技术。随着人们对fMRI数据的需求和可用性的不断增加以及由于运动而导致的重复扫描成本的不断增加，对运动鲁棒性和高效性的fMRI的需求变得越来越迫切，本项目旨在填补fMRI研究中的差距，提出研究基于新型内隐函数学习技术的运动鲁棒性和高效性fMRI。拟议的研究将整合和推进机器学习和医学成像领域的最新研究。在这项研究中，我们将特别考虑婴儿运动轨迹的运动校正，因为婴儿运动会导致功能磁共振成像中大量的数据丢失，代表了最必要和迫切的需要。然而，由于没有太多的低运动婴儿的数据，因为他们也不能很容易地提供无运动控制验证，我们建议使用成人功能磁共振成像的初步可行性研究，在那里可以获得更多的低运动和更好的“地面实况”控制的数据。该项目将创建一种新的隐式函数学习方法，以学习分辨率不可知的无运动fMRI的先验空间，研究数据驱动的先验与特定实例的切片到体积配准的集成，以在模仿婴儿运动的运动场景中进行快速和自适应的运动校正，并评估和验证所创建的方法对成人大脑fMRI扫描数据的影响。这将有助于创建一种新的方法，使高精度，内存效率和强大的功能磁共振成像运动校正与分辨率无关的体积重建。未来的研究将扩展到婴儿的功能磁共振成像鉴于该项目的一个有希望的结果.该项目将有助于机器学习的知识，医学成像和计算机视觉，通过推进在多学科领域的基础和应用研究的最新技术.该项目还将使临床医生和医学图像处理研究人员受益，特别是在fMRI和婴儿方面，为他们提供快速可靠的运动校正工具，以解决当前技术的缺点。患者、医疗保健行业和社会也将受益于医学成像技术的发展，从而改善医疗保健系统和经济。

Functional magnetic resonance imaging (fMRI) is a leading modality to measure brain activity and connectivity. Clinically, it is starting to be used in pre-surgical planning and in assessment of brain function in vegetative state patients. It also recently shows promise in infant cognition research, which holds the key to understanding the origins and functions of the human brain. However, one of the main challenges that constrain the clinical applications of fMRI is its sensitivity to motion, where head movement causes highly deleterious artefacts in fMRI data and can be a major source of error in functional connectivity analysis. This is particularly challenging on infants between the ages of 2 and 48 months, where in many cases half of the data are discarded due to head movement, leading to significant delays and cost for repeated scans. The sustained increase in demand for it would lead to further increased pressures on hospital resources and reduced efficiency of the imaging workflows. Therefore, it is urgent to have techniques and tools to eliminate or reduce the motion effects on fMRI scans.Recently, machine learning (ML) and deep learning (DL) techniques have shown promise to alleviate motion corruption by learning from data to retrospectively correct the motion and artefacts. However, most of these learning-based methods do not specifically focus on fMRI and most existing motion correction approaches for static and structural MRI are not directly applicable to fMRI due to the high memory requirement and application-specific motion artefacts in fMRI. Therefore, there is still a lack of a robust and reliable technique for the problem. With the increasing need and availability of fMRI data and the growing cost for repeated scans due to motion, demand for motion-robust and efficient fMRI are becoming essential.We aim to fill the gap in fMRI research in this project by proposing to investigate motion-robust and efficient fMRI based on novel implicit function learning techniques. The proposed research will integrate and advance state-of-the-art research in machine learning and medical imaging. We will particularly consider motion correction of infant motion trajectories in this study, as infant motion causes substantial data loss in fMRI and represents the most necessary and urgent need. However, as there is not much low-motion infant data available and as they also cannot easily provide motion-free control for validation, we propose to use adult fMRIs for the initial feasibility study, where more data of low motion and better 'ground truth' control can be obtained. The project will create a novel implicit function learning method to learn a prior space for resolution-agnostic motion-free fMRI, investigate integration of the data-driven prior with instance-specific slice-to-volume registration for fast and adaptable motion correction in motion scenarios mimicking infant movement, and evaluate and validate the created approach on data of adult brain fMRI scans with and without infant-like motion. This will conduce to creation of a novel method that enables high-precision, memory-efficient and robust fMRI motion correction with resolution-agnostic volumetric reconstruction. Future research will be extended to infant fMRI given a promising outcome of the project.The project will contribute to knowledge in machine learning, medical imaging and computer vision, by advancing state-of-the-art in both the fundamental and applied research in the multi-disciplinary field. The project will also benefit clinicians and medical image processing researchers especially on fMRI and infant, offering them a fast and reliable motion correction tool that addresses the drawbacks of current techniques. The patients, the healthcare industry and the society will also benefit from the development in medical imaging technologies, with an improved healthcare system and economics resulting from it.