Learning diffusion MR from commercially available protocols: bringing advanced tractography into routine neurosurgical practice

扩散磁共振成像(DMRI)纤维束成像的准确定位可以通过识别应该保留以避免功能缺陷的白质(WM)束来辅助外科计划。然而，仍有一些缺点限制了临床上最佳的气管造影术。首先，信号建模依赖于采集参数，包括信噪比(SNR)、信号幅度(b值)和扩散加权梯度的最小数目(b-VEC)，以实现高质量的局部纤维取向建模。然而，采集时间有限，商业扫描仪可能无法提供强大的最先进的dMRI采集。其次，纤维跟踪方法在如何区分体素内不同的复杂纤维配置方面存在内在的未解决的挑战，在体素中，轴突可以交叉、接吻、弯曲或扇出。第三，对于通常不考虑的束位置，纤维跟踪算法中存在可变性。大多数最先进的光路成像方法不估计WM路径的不确定性，并且通常执行后处理用户引导的滤波来消除虚假流线。在这个博士项目中，我的目标是开发一种端到端的方法，以确保对临床获得的dMRI进行最佳的气管造影术，从而在神经外科手术中提供更准确的指导。我将通过提高从商业dMRI采集中恢复复杂局部纤维方向的能力来做到这一点，并使用深度学习为术前指导提供束不确定性测量。为了实现这一目标，我将：1)开发卷积神经网络(CNN)方法来改进商业MRI模型拟合；2)计算纤维束成像的不确定性量化；3)整合纤维束成像的特定特征以改进模型拟合并验证用于胶质瘤患者的管道。

Accurate localization of diffusion magnetic resonance imaging (dMRI) tractography can aid surgical planning by identifying white matter (WM) tracts that should be preserved to avoid functional deficits. However, there are still shortcomings that limit optimal tractography in a clinical setting. Firstly, signal modeling depends on acquisition parameters including signal-to-noise ratio (SNR), signal magnitude (b-values), and the minimum number of diffusion-weighting gradients (b-vecs) for high-quality local fiber orientation modeling. However, acquisition time is limited, and commercial scanners may not provide robust state-of-the-art dMRI acquisitions. Secondly, fiber tracking methods have intrinsic unresolved challenges in how to distinguish between different complex fiber configurations within a voxel, where axons can cross, kiss, bend, or fan out. Thirdly, there is variability in the fiber tracking algorithms with regards to tracts locations that usually are not taken into account. Most of the state-of-the-art tractography approaches do not estimate the uncertainty of WM pathways, and post-processing user-guided filtering is often performed to eliminate spurious streamlines. In this Ph.D. project, I aim to develop an end-to-end approach to ensure optimal tractography for clinically acquired dMRI and thereby provide more accurate guidance during neurosurgery. I will do that by improving the ability to recover complex local fiber orientations from commercial dMRI acquisitions and providing a tract uncertainty measure for preoperative guidance using deep learning. To achieve this I will: 1) Develop a convolutional neural network (CNN) approach to improve commercialdMRI model fitting; 2) Compute tractography uncertainty quantification, 3) Incorporate tractography specific features to improve model fitting and validate pipeline on patients with gliomas.