Connections of white matter are typically inferred from diffusion MRI using deterministic or probabilistic tractography algorithms^1-4. These algorithms operate by taking as input orientation distributions functions, or a set of peaks, and by connecting adjacent voxels based on peaks that are maximally aligned¹. However, not all fascicles are organized so that peaks along their course in the brain are maximally aligned. A perfect example of such caveat is the Meyer’s loop (ML), a highly curved fascicle known to exhibits a narrow turn, kissing/crossing regions, and changes in fascicle dispersion⁵. This fascicle is well understood from conventional Klinger dissection and histological analysis, and yet, the virtual reconstruction of the ML with tractography typically remains incomplete. Common methods encode our knowledge of the anatomy by solely using Boolean ROIs⁶ to include or exclude streamlines. We hypothesise that a new magnetic-ROI can help incorporate a priori information about the course of the pathway and lead to an improved delineation of the ML. The main goals of MAGNEtic Tractography (MAGNET) are thus to 1) increase the accuracy of tractography by selecting a specific direction based on a prior anatomical knowledge, and 2) reduce the total streamline calculation burden by avoiding an exponential search of all possibilities (i.e. computationally infeasible and inaccurate as it increases the amount of false positive^{4, 7}).

The idea is to provide a preferential direction given by V_magnet when a streamline enters a predefined magnetic-ROI. The most popular tracking equation¹ is of the form: V_next = argmin_k α(V_in ,V_k) (eq. 1), where V_next is the next direction to propagate and α is the angle between the incoming direction V_in and the orientation of the k^th peak V_k in the voxel. Instead, when tracking inside the new magnetic-ROI, we propose to follow the V_k that is most aligned with V_magnet. The propagation equation becomes then: V_next = { argmin_k α(V_magnet,V_k) if inside the magnetic-ROI; eq. 1 otherwise }. If a voxel contains a single direction, the propagation naturally resumes with eq. 1. To measure the effect of our new evolution equation, MAGNET was applied on simulated data and on in vivo data obtained from 15 control subjects (10.2±3.1 years).

Synthetic dataset: A noiseless synthetic dataset consisting of 2 bundles crossing at 45° was generated using Phantomas⁸. A seed-ROI (Fig. 1a, purple box) and a magnetic-ROI (Fig. 1a, red box) were positioned in the peaks field. Streamline propagation was then qualitatively observed by activating MAGNET in real-time⁹.

Human datasets: Diffusion scans were performed using a multi-direction (90) and multi-b-value (range: 400-3000) scheme (TR/TE: 5700/89 ms, 1.7 × 1.7 × 2 mm³, CUSP90¹⁰) on a Siemens 3T Trio MRI. Multi-fiber model estimation was done using Diffusion Compartment Imaging (DCI)¹¹, resulting in up to 3 main peaks per voxel. Tractography of the left optic radiation was performed using the following parameters: step size: 1 mm, θ_max: 45°, maximum consecutive steps in the cortex: 5, min/max length: 60/200 mm. White matter/grey matter masks extracted from subject-specific anatomical T1-weighted images (1 mm isotropic) were used as tracking masks⁴. 3375 seeds were interactively⁹ placed anterolaterally to the left lateral geniculate nucleus (LGN)⁵, with initial seed direction oriented in the left direction (Fig 2a, "S"). An inclusion (AND) planar-ROI positioned at the midbody of the optic radiation and an exclusion (NOT) sagittal plane acted as filtering regions. To maximize the extent and coverage of the ML, magnetic-ROIs were then placed around the medial, anterior and lateral tip of the ML (Fig. 2a, b). These magnetic-ROIs favored the selection of DCI peaks that were oriented toward the visual cortex.

Fig. 1 shows the effect of MAGNET on synthetic data. Activating the magnetic-ROI enabled specific selection of DCI peaks oriented toward the 45° pathway (Fig. 1c). Fig. 2 illustrates the magnetic-ROIs positioning for a single subject, as well as a conventional view of ML tractography. Fig.3 qualitatively shows that MAGNET successfully recovered a larger extent of ML for 15 subjects compared to traditional Boolean ROIs. Corresponding quantitative results are summarized in Tab.1. Finally, Fig. 4 shows an accurate comparison between MAGNET-based delineation of ML and histological drawing. We showed that MAGNET can accurately reconstruct ML in all subjects. It effectively improved streamline coverage, and significantly reduced the ML-temporal pole and ML-inferior horn distances, crucial information for preoperative planning of temporal lobe surgery¹². In future work, we plan to apply the magnetic operator inside automatically segmented anatomical structures (e.g. parcels). Our MAGNET technique is expected to enable unprecedented delineation of neural circuits not visible with conventional tracking algorithms.