Real-Time Tractography-Assisted Neuronavigation for Transcranial Magnetic Stimulation

Abstract

State-of-the-art navigated transcranial magnetic stimulation (nTMS) systems can display the TMS coil position relative to the structural magnetic resonance image (MRI) of the subject's brain and calculate the induced electric field. However, the local effect of TMS propagates via the white-matter network to different areas of the brain, and currently there is no commercial or research neuronavigation system that can highlight in real time the brain's structural connections during TMS. This lack of real-time visualization may overlook critical inter-individual differences in brain connectivity and does not provide the opportunity to target brain networks. In contrast, real-time tractography enables on-the-fly parameter tuning and detailed exploration of connections, which is computationally inefficient and limited with offline methods. To target structural brain connections, particularly in network-based treatments like major depressive disorder, a real-time tractography-based neuronavigation solution is needed to account for each individual's unique brain connectivity. The objective of this work is to develop a real-time tractography-assisted TMS neuronavigation system and investigate its feasibility. We propose a modular framework that seamlessly integrates offline (preparatory) analysis of diffusion MRI data with online (real-time) probabilistic tractography using the parallel transport approach. For tractography and neuronavigation, we combine our open source software Trekker and InVesalius, respectively. We evaluate our system using synthetic data and MRI scans of four healthy volunteers obtained using a multi-shell high-angular resolution diffusion imaging protocol. The feasibility of our online approach is assessed by studying four major TMS targets via comparing streamline count and overlap against offline tractography results based on filtering of one hundred million streamlines. Our development of a real-time tractography-assisted TMS neuronavigation system showcases advanced tractography techniques, with interactive parameter tuning and real-time visualization of thousands of streamlines via an innovative uncertainty visualization method. Our analysis reveals considerable variability among subjects and TMS targets in the streamline count, for example, while 15,000 streamlines were observed for the TMS target on the visual cortex (V1) of subject #4, in the case of subject #3's V1, no streamlines were obtained. Overlap analysis against offline tractograms demonstrated that real-time tractography can quickly cover a substantial part of the target areas' connectivity, often surpassing the coverage of offline approaches within seconds. For instance, significant portions of Broca's area and the primary motor cortex were effectively visualized after generating tens of thousands of streamlines, highlighting the system's efficiency and feasibility in capturing brain connectivity in real-time. Overall, our work shows that real-time tractography-assisted TMS neuronavigation is feasible. With our system, it is possible to target specific brain regions based on their structural connectivity, and to aim for the fiber tracts that make up the brain's networks. Real-time tractography provides new opportunities for TMS targeting through novel visualization techniques without compromising structural connectivity estimates when compared to the offline approach.

Keywords: TMS; brain stimulation; connectivity; diffusion MRI; neuronavigation; tractography.

FIGURE 1

For real‐time tractography‐assisted neuronavigation, we combine information from T1 and dMRI data. In the offline (preparatory) part, necessary inputs for the online (real‐time) part are prepared. The necessary inputs are: (i) fiber orientation distribution (FOD) image, needed for fiber tracking, (ii) anatomical constraints, needed to reduce false positives during tractography by preventing improper termination within white matter and crossing through cerebrospinal fluid, (iii) segmented brain mask, needed to compute the peeled brain surfaces, and (iv) T1 image, to show the grayscale brain image on the peeled surfaces. The online part consists of the neuronavigation and tractography modules that continuously run in a multi‐threaded loop. Neuronavigation and tractography are done using our custom software InVesalius (Souza et al. 2018) and Trekker (https://dmritrekker.github.io/), respectively. The main features used during real‐time operation are inside the yellow boxes.

FIGURE 2

(a) Sensitivity and specificity values based on voxel‐wise overlap between dMRI‐based tractography and ex vivo tracer injection experiments demonstrate the trade‐off in tractography performance for various step size and curvature threshold combinations (81 samples) for each fixed FOD threshold value (Aydogan and Shi 2018). The ROC curve can be traversed by varying the FOD threshold (minFODamp parameter in Trekker). Low values for FOD threshold lead to increased sensitivity at the cost of decreased specificity. By displaying the streamlines generated using low FOD thresholds with more transparency, we provide visual information to the operator regarding increased possibility of false connectivity as a result of the corresponding parameter choice. (b) Fixed parameters used for fiber tracking and the range of values for the varying FOD threshold. The transfer function linearly maps the increasing FOD threshold to opacity values ranging from 20% to 100%.

FIGURE 3

Tractography‐assisted nTMS setup. The display with the InVesalius user interface shows the peeled brain surface and the streamlines obtained from the real‐time TMS coil position. The coil is being held by the operator and tracked by the neuronavigation software that analyzes data from an infrared (IR) camera and retroreflective markers.

FIGURE 4

The variation in tractography performance with respect to the changes in minFODamp parameter for uncertainty visualization was tested using the ISMRM 2015 tractography challenge data. Each red dot represents a score obtained for a tractogram with 10 million streamlines generated by randomly seeding the whole‐brain mask using each of the 10 minFODamp parameters shown in Figure 2. The obtained scores with Trekker are highly competitive against the original submissions to the challenge. The trend shows the expected trade‐off between true and false positives, where increasing the minFODamp parameter decreases bundle overreach, that is, false positives, at the cost of reduced bundle overlap, that is, true positives.

FIGURE 5

Tractograms generated using a seed region placed on the motor cortex show cortico‐spinal‐tract (CST) and connections through the corpus callosum (CC), which are commonly studied in TMS experiments. The three rows show different views of the same tractogram. Tractograms show that increasing minFODamp produces streamlines that may not sample the whole extent of connections. Tractograms with lower minFODamp values reach more regions; however, streamlines lose organization, which can lead to increased false positives. Combination of the tractograms with uncertainty visualization shows all the streamlines. But because streamlines computed with lower minFODamp are shown with more transparency, user is provided with visual information that these connections are more likely to be false positives than other streamlines shown on display.

FIGURE 6

The number of offline‐filtered streamlines. Each value is obtained by selecting streamlines that pass through the seed regions from a whole‐brain tractogram containing 10 million streamlines. Values are reported for each of the 10 repetitions of coil placement, that is, seed point, as well as 10 different minFODamp values.

FIGURE 7

Overlap percentages with respect to the number of streamlines. Here, the streamlines are obtained offline with the same method as the real‐time case. The reference contains 1 million streamlines obtained by combining 10 seed‐based tractograms (see Section 3.2.3). Each of the 10 tractograms are computed with a different minFODamp value and contains 100,000 streamlines. Black dots show the overlap obtained by filtering the whole‐brain tractograms that contain 100 million streamlines for each seed.