Concurrent multi-task learning-based deep convolutional neural networks for high resolution assessment of in-vivo cardiac microstructure

博士项目的目标：开发一种基于并行多任务学习的深度卷积神经网络（CNN），用于高度欠采样的螺旋MRI数据的数据采集和图像重建，具有最小的伪影。部署这些方法用于心脏微结构的高效，高分辨率的体内扩散张量心血管磁共振。在对照组和心肌梗死（MI）患者中验证这些方法。项目描述：心肌细胞（心肌细胞）和心肌细胞群（称为小薄片）的复杂排列和动力学对正常心脏功能至关重要。扩散张量心血管磁共振（DT-CMR）是一种独特的MRI方法，基于测量水的自扩散，提供微观组织结构的信息。DT-CMR可以推断心肌细胞和小薄片的方向，其在收缩期间重新定向，并提供对细胞外空间、膜完整性和心肌细胞方向一致性的变化敏感的测量。这种新方法越来越多地用于研究疾病的微观变化。作为我们正在进行的皇家布朗普顿医院心血管磁共振单元心脏微观结构研究的一部分，我们已经证明，使用STEAM技术可以在心动周期的多个时间点采集DT-CMR，这是一种低信噪比（SNR）技术，因此需要在低分辨率下进行许多信号平均。因此，病理影响右心室和心房，心肌变薄的慢性梗死组织和小的局部变化不能充分investigated.We最近开发了一种技术，采样数据沿着两个交错的螺旋路径分裂的采集更高分辨率的DT-CMR数据在两个心跳。然而，交错DT-CMR技术对由磁场中的局部变化和两个螺旋之间的运动引起的伪影敏感。尽管存在这些伪影，但我们能够将体内DT-CMR采集的平面内分辨率从2.8x2.8mm2提高到1.8x1.8mm2[5]。不幸的是，采集两个交错需要将已经很长的扫描时间加倍，每个切片和心脏相位通常需要20次18秒的屏气。相比之下，从单个交错重建数据节省了时间，提高了患者的舒适度，并避免了一些相关的伪影。并行成像经常用于减少MRI中所需的数据，但螺旋STEAM DT-CMR中使用的SNR损失，长计算时间和小视场是不相容的。在最近的工作中，我们已经建立了深度卷积神经网络（CNN）在MRI重建[6]和去噪DT-CMR数据集[7]中的有效性。在螺旋DT-CMR[8]的模拟中，我们沿着螺旋轨迹回顾性地对DT-CMR图像进行了欠采样，并使用完全采样的数据作为地面实况来训练CNN。我们证明了有效的去除混叠伪影的欠采样因子高达4，在这个有前途的，但早期的试点数据。在这里，我们的目标是建立在这些初始计算模拟的基础上，并在具有挑战性的患者队列中开发一种适用于CNN的有效和强大的高分辨率螺旋DT-CMR的临床体内工具。学生将开发一种新型的基于并发多任务学习的深度CNN，用于优化数据采集（即寻求最佳螺旋轨迹）并同时实现高保真图像重建（即去除欠采样和其他伪影）。将在皇家布朗普顿医院使用我们最先进的3 T西门子维达扫描仪在模拟和体内测试这些方法。

Aim of the PhD Project:Develop a concurrent multi-task learning-based deep convolutional neural network (CNN) for data acquisition and image reconstruction of highly-undersampled spiral MRI data with minimal artefact.Deploy these methods for efficient, high-resolution in-vivo diffusion tensor cardiovascular magnetic resonance of cardiac microstructure.Validate these methods in controls and patients with myocardial infarction (MI).Project Description:The complex arrangement and dynamics of heart muscle cells (cardiomyocytes) and groups of cardiomyocytes known as sheetlets is vital to normal cardiac function. Diffusion tensor cardiovascular magnetic resonance (DT-CMR) is a unique MRI method providing information on microscopic tissue structures, based on measuring the self-diffusion of water. DT-CMR can infer the orientation of cardiomyocytes and sheetlets, which reorientate during contraction and provide measures sensitive to changes in extracellular space, membrane integrity and coherence of cardiomyocyte orientation. This novel method is increasingly used to investigate the microscopic changes underlying disease.As part of our ongoing investigations into cardiac microstructure at the Cardiovascular Magnetic Resonance Unit, The Royal Brompton Hospital, we have shown that acquisition of DT-CMR at multiple timepoints in the cardiac cycle is possible using the STEAM technique, which is a low signal to noise ratio (SNR) technique and therefore requires many signal averages at low resolution. As a result, pathologies affecting the right ventricle and atria, the thinned myocardium of chronically infarcted tissue and small focal changes cannot be adequately investigated.We recently developed a technique which samples data along two interleaved spiral paths to split the acquisition of higher resolution DT-CMR data across two heartbeats. However, interleaved DT-CMR techniques are sensitive to artefacts caused by localised changes in the magnetic field and motion between the two spirals. Despite these artefacts, we were able to increase the in-plane resolution of in-vivo DT-CMR acquisitions from 2.8x2.8mm2 to 1.8x1.8mm2[5].Unfortunately, acquiring two interleaves requires doubling the already long scan times, with typically twenty 18s breath holds required per slice and cardiac phase. In contrast, reconstructing data from a single interleave saves time, improves patient comfort and avoids some of the associated artefacts. Parallel imaging is frequently used to reduce the data required in MRI, but the associated loss of SNR, long computation times and the small field of view used in spiral STEAM DT-CMR are incompatible.In recent work we have established the effectiveness of deep convolutional neural networks (CNNs) in MRI reconstruction[6] and denoising DT-CMR datasets[7]. In simulations of spiral DT-CMR[8], we retrospectively undersampled DT-CMR images along spiral trajectories and trained a CNN using the fully sampled data as the ground truth. We demonstrated effective removal of aliasing artefacts with undersampling factors of up to 4 in this promising, but early stage pilot data. Here, we aim to build on these initial computational simulations and develop a clinically applicable in-vivo tool for CNN enabled efficient and robust high-resolution spiral DT-CMR in challenging patient cohorts. The student will develop a novel concurrent multi-task learning-based deep CNN for optimising data acquisition (i.e. seeking optimal spiral trajectories) and achieving high-fidelity image reconstruction (i.e. removal of undersampling and other artefacts) simultaneously. These methods will be tested in simulations and in-vivo using our state-of-the-art 3T Siemens Vida scanner at The Royal Brompton Hospital.