Distortion Correction in Functional MRI with Deep Learning

项目摘要 功能磁共振成像(FMRI)，一种非侵入性的脑活动图谱技术，已经被 广泛应用于认知神经科学和病人护理。组织周围磁场不均匀(B) 在fMRI图像中，界面会在特定的脑区引起严重的几何失真。图像失真 导致fMRI和高分辨率解剖MRI图像之间的配准错误，从而降低 FMRI检测脑活动的空间准确性和灵敏度。在目前的功能磁共振研究中，B诱导 通常使用基于图像的方法来校正重建的幅度图像中的失真 这假定B平滑变化。然而，基于配准的校正(REG-CORR)可以 导致图像伪影和模糊，因为它的假设在B变化的大脑区域崩溃 震级图像中相位信息的快速和遗漏可能会加剧计算误差。这个 该项目的总体目标是开发一种基于深度学习的新方法来准确纠正 用于通过图像重建的几何失真。通过将B效果的物理模型集成到 展开DL网络，可以直接从k-M空间中的复磁共振信号中重建无失真的fMRI图像。 空间，而不假设B的光滑性。提出了基于重建的校正。 (Recon-Corr)算法将使用4050次fMRI扫描的原始k空间数据进行训练和测试，在急性到 慢性疼痛信号(A2CPS)联盟，其中伊利诺伊大学芝加哥分校是主要成员 表演现场。该项目有两个具体的目标：(1)开发一个物理制导的动态链接库算法 同时进行fMRI图像重建和失真校正；(2)系统比较 Recon-Corr方法和传统的Reg-Corr方法的性能比较。通过发展Recon-Corr方法和 利用大型A2CPS功能磁共振k空间数据库，本项目将演示一种准确的功能磁共振成像方法 失真校正，可以提供更好的功能和解剖MRI图像的配准精度。 该项目的成功完成将解决功能磁共振成像(即图像)中一个长期存在的重要问题 失真)，有助于功能磁共振成像在神经科学、病人护理和其他研究领域的应用。

Project Abstract Functional magnetic resonance imaging (fMRI), a non-invasive technique for mapping brain activity, has been widely used in cognitive neuroscience and patient care. Magnetic field inhomogeneities (B) around tissue interfaces can induce severe geometric distortions in specific brain regions in fMRI images. The image distortions lead to errors in the registration between fMRI and high-resolution anatomical MRI images, and thus decrease spatial accuracy and sensitivity of detecting brain activity with fMRI. In present fMRI studies, B-induced distortions are typically corrected in the reconstructed magnitude images using methods based on image registration, which assume a smoothly varying B. However, the registration-based correction (Reg-Corr) can cause image artifacts and blurring because its assumption breaks down in brain regions where B changes rapidly and omission of phase information in the magnitude images can exacerbate calculation errors. The overarching goal of this project is to develop a novel approach based on deep learning (DL) to accurately correct for geometric distortions through image reconstruction. By integrating the physical model of B effects into an unrolling DL network, distortion-free fMRI images will be directly reconstructed from the complex MR signal in k- space, without the assumption about the smoothness of B. The proposed reconstruction-based correction (Recon-Corr) algorithm will be trained and tested with raw k-space data from 4050 fMRI scans, in the Acute to Chronic Pain Signatures (A2CPS) consortium, in which the University of Illinois at Chicago is a primary performing site. The project has two specific aims: (1) To develop a physics-guided DL algorithm for simultaneous fMRI image reconstruction and distortion correction; (2) To systematically compare the performance of Recon-Corr and traditional Reg-Corr methods. By developing the Recon-Corr method and leveraging the large A2CPS fMRI k-space database, this project will demonstrate an accurate method for fMRI distortion correction that can offer better registration accuracy of functional and anatomical MRI images. Successful completion of the project will resolve a long-standing and important problem in fMRI (i.e., image distortion), contributing to fMRI applications in neuroscience, patient care, and other research areas.