Ultra-high gradient connectomics and microstructure MRI scanner for imaging of human brain circuits across scales

The study addresses the challenge of linking structure and function across scales in the human brain, from macroscopic connections to mesoscopic and microscopic cellular architecture. Diffusion MRI enables non-invasive probing of microstructure by leveraging water diffusion, but robust mapping of the smallest cellular compartments requires faster and stronger gradients than conventional human scanners provide. Building on the Human Connectome Project and the original 3 T Connectome 1.0 scanner (300 mT m⁻¹), the NIH BRAIN Initiative supported the development of Connectome 2.0: a 3 T, head-only, ultra-high gradient scanner designed to bridge micro-, meso-, and macro-scale connectomics. The design goal was to substantially increase gradient strength and slew rate while mitigating peripheral nerve stimulation, improving SNR, and enabling accurate mapping of fine white matter pathways and microstructural features in vivo.

Prior work established diffusion MRI as a powerful non-invasive modality for microstructural imaging, with early methods limited to ex vivo and small-animal systems. The Connectome 1.0 (300 mT m⁻¹) demonstrated macroscopic tractography and biophysical modeling in humans, enabling axon diameter indices down to several microns, though sensitivity to the smallest axons (≈2–3 µm) remained limited. Theory and simulations indicated that higher gradient strengths and optimized waveforms are needed to reach smaller diameters and improve resolution. Head-only high-performance gradient inserts, such as the Siemens 7 T Impulse coil (G=200 mT m⁻¹, SR=900 T m⁻¹ s⁻¹), focused on fMRI at ultra-high spatial resolution, while other 3 T systems (e.g., GE MAGNUS) increased gradient efficiency for microstructure imaging. Advances in parallel RF reception and dynamic field monitoring have been shown to mitigate higher-order eddy-current artifacts and improve accelerated imaging performance. Collectively, these developments provided the foundation and motivation for Connectome 2.0’s architecture.

Scanner and gradient coil: Connectome 2.0 (MAGNETOM Connectom.X, Siemens Healthineers) is a 3 T head-only system with a 3-layer gradient coil designed via target-field optimization and explicit peripheral nerve stimulation (PNS) modeling. The asymmetric, stepped design (inner/middle/outer layer diameters: 44/58/71 cm; patient bore at isocentre: 40 cm) doubles primary layer windings to raise current density and achieves Gmax=500 mT m⁻¹ and SRmax=600 T m⁻¹ s⁻¹ per axis. Coil sensitivity is 0.42 mT m⁻¹ A⁻¹, inductance (x,y,z) 2,250/2,450/1,800 µH, DC resistance 0.28 Ω per axis. Two gradient power amplifiers per axis (1,200 A, 2,250 V each) independently drive two partial coils per axis with minimal mutual coupling. Direct cooling uses stainless-steel tubing within copper filaments, >20 parallel circuits, ~40 l min⁻¹ flow, and >50 temperature sensors; safe steady-state temperatures at 85 °C were achieved at 100% DC with single-axis Grms 175 mT m⁻² and all-axes 150 mT m⁻². Acoustic levels reached 116 dB(A) by NEMA MS4. Linearity errors over a 20-cm sphere: X 6.7%, Y 8.3%, Z 11.7%. PNS modeling and measurements: Design incorporated PNS balancing using an intermediate winding layer to redistribute stimulation thresholds between face and torso, raising worst-case thresholds by ~41% (from 114 to 161 mT m⁻¹). PNS thresholds were experimentally measured in 29 healthy volunteers using 128-cycle trapezoidal pulse trains (flat-top 500 µs; rise times 50–3,000 µs) for single- and combined-axis modes. Connectome 2.0 showed 2.4–4.2× higher thresholds than Connectome 1.0 and 2.5–5.4× higher than state-of-the-art clinical scanners, enabling routine diffusion encoding up to Gmax=500 mT m⁻¹ and SRmax=600 T m⁻¹ s⁻¹ within SAFE/IEC limits. Gradient system characterization and field monitoring: Gradient impulse response functions (GIRF) were measured with 16-channel dynamic field cameras (Skope) using blips and sweeps, capturing axis responses and cross-terms up to second order. Each receive array integrated 16 ¹⁹F probes, positioned to measure higher-order spatiotemporal field perturbations; concurrent field monitoring was used to inform image reconstruction (SENSE/ESPIRiT) and suppress Nyquist ghosting and eddy current distortions beyond dual-polarity GRAPPA. RF coils: Two custom arrays were built. In vivo: 72-channel head receive array on a single-shell helmet (85th percentile sizing) with a local circularly polarized birdcage transmit (16-rung hybrid, tuned at 123.25 MHz) and slotted copper RF shields for both Tx and gradient bore; field probes integrated within the array. Ex vivo: 64-channel whole-brain receive array with dedicated local birdcage Tx and embedded temperature monitoring for controlled long scans; the coil conformed anatomically to brain specimens and included forced air cooling. Circuit designs used silver-plated wire loops, preamplifier decoupling, passive detuning for safety, and g-factor optimization for accelerated acquisitions. Imaging protocols and SNR comparisons: Diffusion MRI used 2D monopolar PGSE-EPI sequences. SNR comparisons were run on Connectome 2.0 with protocols emulating clinical (Gmax=80 mT m⁻¹, SRmax derated to 80 T m⁻¹ s⁻¹), Connectome 1.0 (Gmax=300 mT m⁻¹, SRmax derated to 80 T m⁻¹ s⁻¹) and Connectome 2.0 (Gmax=500 mT m⁻¹, SRmax=600 T m⁻¹ s⁻¹), all with δ=8 ms and matched TR/echo spacing. Temporal SNR was evaluated in cerebral white matter using the last 10 b=0 images, with SynthSeg-based white matter masks and theoretical T2-based predictions. A PGSE-EPI sequence simulator predicted minimal TE for given b-values considering gradient constraints and sequence timing. High-resolution tractography: A healthy adult underwent 1 mm isotropic DWIs at b=1,000 and 2,500 s mm⁻² with 64 directions per shell, reverse PE b=0 for susceptibility correction, and DESIGNER preprocessing (denoising, Gibbs removal, distortion corrections, gradient nonlinearity). Fibre orientation distribution functions were estimated via multi-shell, multi-tissue CSD in MRtrix3; probabilistic tractography seeded 5 seeds per white matter voxel with lmax=8. Datasets were co-registered via inverse-consistent transforms. Microstructure modeling: Ten healthy adults scanned on Connectome 2.0 (matched to 10 Connectome 1.0 participants) underwent multi-b-value acquisitions at Δ=13 and 30 ms (C2.0) versus Δ=19 and 49 ms (C1.0), δ=6 ms (C2.0) vs δ=8 ms (C1.0), with interspersed b=0 and 32–64 directions. Spherical mean signals were fitted with AxCaliber–SMT in white matter (MCMC fitting with fixed longitudinal diffusivity 1.7 µm² ms⁻¹) to derive axon diameter indices and compartment fractions; maps were normalized to MNI152 and analyzed in JHU ROIs with FDR-corrected statistics. SANDI fits in cortical grey matter (short Δ to avoid exchange) used random forest regression to estimate apparent soma radius and compartment fractions; metrics were projected to FreeSurfer surfaces and compared across Brodmann areas with FDR-corrected paired tests.

- Hardware performance: Head-only 3-layer gradient coil achieved Gmax=500 mT m⁻¹ and SRmax=600 T m⁻¹ s⁻¹ per axis, yielding 5× higher gradient performance than Connectome 1.0 and >20× typical clinical systems. Coil sensitivity was 0.42 mT m⁻¹ A⁻¹; inductance (x,y,z)=2,250/2,450/1,800 µH; DC resistance=0.28 Ω. - PNS thresholds: Design raised worst-case thresholds by ~41% via intermediate layer balancing. In 29 volunteers, Connectome 2.0 exhibited 2.4–4.2× higher thresholds than Connectome 1.0 and 2.5–5.4× higher than state-of-the-art clinical scanners, enabling routine use of the expanded hardware performance space. - SNR gains and TE reductions: Shorter Δ and δ allowed TE reductions of 13–50% vs Connectome 1.0 across b-values up to 40,000 s mm⁻². SNR gains reached ~2× at b=40,000 s mm⁻² vs Connectome 1.0 and ~4× at b=5,000 s mm⁻² vs clinical protocols; theoretical predictions and in vivo measurements aligned. - RF coil performance: The 72-channel in vivo coil increased SNR by ~1.5× in peripheral (cortical) regions and ~5% centrally vs a standard 32-channel head coil, and provided improved acceleration (lower g-factors), enabling at least one additional unit of acceleration for matched noise amplification. The 64-channel ex vivo coil delivered ~1.73× higher average SNR than the 72-channel in vivo coil in phantom comparisons, particularly near the sample periphery; both arrays integrated 16-channel dynamic field monitoring. - High-resolution tractography: With 1 mm isotropic DWIs acquired in ~30 minutes, Connectome 2.0 resolved fine diencephalic and brainstem pathways (e.g., mammillo-tegmental tract) and reconstructed internal/external/extreme capsules more robustly than a Connectome 1.0 protocol. - Microstructure sensitivity: AxCaliber–SMT showed improved sensitivity to smaller axons, consistent with a resolution limit scaling ∝ Gmax⁻¹/². In the posterior corona radiata, axon diameter indices were 2.45 ± 0.15 µm on Connectome 2.0 vs 4.04 ± 0.48 µm on Connectome 1.0, indicating capture of smaller-diameter axons and improved precision (lower SD, better hemispheric consistency). SANDI revealed stronger cortical contrasts in intrasoma fractions on Connectome 2.0 (e.g., BA3a 0.36 ± 0.02 vs BA4 0.32 ± 0.02, BA3b 0.27 ± 0.03, BA1 0.21 ± 0.02; FDR-corrected P<0.001), whereas Connectome 1.0 detected limited contrast. Estimated soma radii were smaller on Connectome 2.0 (8.52 µm) vs Connectome 1.0 (9.26 µm; P<0.001), suggesting sensitivity to a greater proportion of smaller cells. - Image fidelity: Dynamic field monitoring and higher-order reconstruction effectively mitigated eddy-current-induced trajectory deviations and Nyquist ghosting beyond standard linear-phase corrections and dual-polarity GRAPPA. - Operational metrics: Maximum technical sound level was 116 dB(A); safe steady-state temperatures at 85 °C under high-duty cycles were achieved with direct cooling; GIRF characterization captured higher-order and cross-axis responses.

Connectome 2.0’s ultra-strong head gradients and integrated high-sensitivity RF arrays improve diffusion MRI sensitivity and precision across scales, enabling shorter TE, higher SNR, and robust mapping of microstructural features and fine fibre tracts in individuals. The raised PNS thresholds expand the usable gradient performance space while maintaining safety, and dynamic field monitoring ensures high-fidelity reconstruction under strong diffusion encoding. Compared to Connectome 1.0, the system reduces systematic biases in axon diameter estimation and increases precision at the single-participant level. Relative to other high-performance systems (e.g., 7 T Impulse with lower Gmax but higher SRmax), Connectome 2.0 targets mesoscale microstructure imaging at 3 T with a design optimized for diffusion sensitivity rather than primarily fMRI. These advances support individualized, anatomically precise tractography of clinically relevant pathways and quantitative microstructure mapping that could guide personalized neuromodulation and other image-guided interventions. By improving sensitivity to smaller axons and cells, the platform enhances statistical power for studying variability in tissue organization and connectional anatomy within and across individuals.

The Connectome 2.0 scanner integrates a 3-layer, head-only gradient coil (Gmax=500 mT m⁻¹, SRmax=600 T m⁻¹ s⁻¹), advanced RF receive arrays with embedded field monitoring, and robust cooling to deliver a step-change in human diffusion MRI performance. It enables high-resolution tractography in practical scan times and increases sensitivity and precision for microstructural metrics such as axon diameter indices and soma size across cortical and subcortical regions. These engineering advances, aligned with BRAIN Initiative goals, pave the way for precision human neuroscience and clinical applications that non-invasively probe microscopic tissue changes over time. Future work includes broader deployment of ultra-high-performance gradient systems, leveraging faster slew rates for diverse diffusion encoding paradigms, expanding cohort studies to clinical populations, and continued development of artifact-robust reconstruction and biophysical modeling. Stronger and faster gradients adopted by vendors will further permeate scientific and clinical imaging, enabling earlier detection and monitoring of disease processes at cellular scales.

- Gradient nonlinearity: Increased worst-case nonlinearity within the 20-cm sphere (up to 11.7% on Z) is a trade-off for higher Gmax; while geometric distortions can be corrected, uniformity-related pixel-size errors at FOV edges cause unrecoverable spatial resolution loss. - PNS constraints: Despite raised thresholds, biological limits still cap the maximal usable performance in some sequences and participants; operational slew rates on certain legacy protocols may require derating to avoid cardiac stimulation. - Physical constraints: Head-only gradient bore (inner diameter 44 cm) and in vivo helmet sized to the 85th percentile may limit accommodation of larger heads; tight radial spacing imposes integration constraints for Tx/Rx arrays and field probes. - Acoustic and thermal loads: Technical noise levels (116 dB(A)) require standard hearing protection; high-duty cycles demand robust cooling, and sustained high-gradient operation is bounded by thermal limits. - Generalizability and study design: Validation focused on healthy adults with matched cohorts across scanners; power analyses were not conducted and certain replications were limited. Performance comparisons with clinical protocols at very high b-values were theoretical due to excessive SNR loss in clinical systems. - Artifact susceptibility: Strong diffusion encoding induces higher-order eddy currents necessitating concurrent field monitoring and advanced reconstruction; without these, standard linear ghost corrections may be insufficient.