A hierarchical face-scalp-cortex mapping for non-invasive neuromodulation targeting without magnetic resonance images guidance

Technical Summary: A Hierarchical Face-Scalp-Cortex Mapping for Non-Invasive Neuromodulation Targeting Without MRI Guidance

Problem Statement
Non-invasive neuromodulation techniques, such as transcranial magnetic stimulation (TMS), transcranial ultrasound stimulation (TUS), and transcranial photobiomodulation (tPBM), offer effective non-pharmacological treatments for psychiatric and neurological disorders. However, precise targeting typically relies on neuronavigation systems dependent on individual magnetic resonance imaging (MRI). This dependency creates significant barriers: MRI scans are costly, time-consuming, inaccessible to many institutions, and uncomfortable for some patients. Furthermore, current MRI-free alternatives often suffer from labor-intensive manual procedures, reliance on operator experience, or a lack of principled theoretical justification for mapping surface landmarks to internal brain targets. There is a critical need for a unified, automated, and explainable framework that enables precise targeting without MRI guidance.

Methodology
The authors propose a unified framework based on a hierarchical face-scalp-cortex mapping, integrated into a markerless robotic neuronavigation system. The workflow consists of three primary stages:

1. Face-to-Scalp Mapping (Cranial Surface Reconstruction):

   • Input: A facial point cloud acquired via a handheld 3D scanner.
   • Model: A 3D generative modeling architecture combining a Variational Autoencoder (VAE) and a Diffusion Model. The VAE extracts topological latent representations of facial and cranial structures, while the diffusion model learns the mapping from facial latent space to cranial latent space.
   • Process: The system takes the facial point cloud as a conditional input to generate a complete 3D cranial surface model, effectively "filling in" the missing scalp and posterior head regions based on learned correlations from a large dataset of 8,195 paired facial and MRI-derived cranial surfaces.
   • Privacy: The generative modeling is trained on a cloud server; only the reconstructed cranial surface is transmitted back to the local system, protecting patient facial privacy.

2. Cortex-to-Scalp Mapping (Target Localization):

   • Coordinate System: The authors introduce a Scalp Angle Coordinate (SAC) system, defined by three anatomical landmarks: the nasion (Nz) and the left/right preauricular points (LPA/RPA). This system represents scalp points using angular coordinates ($S_\theta, S_\phi$) rather than geodesic distances.
   • Mapping Strategy: Using the SLIM dataset (292 healthy participants), the system statistically analyzes the coordinates of cortical targets (in MNI space) to construct a population-level cortex-to-scalp mapping. This mapping transforms a desired MNI target into a specific location on the individual's generated cranial surface.
   • Advantage: This approach decouples the mapping construction from the need for full scalp spatial data, relying instead on the three key landmarks which are easily identifiable on the facial scan.

3. Markerless Robotic Neuronavigation:

   • Hardware: A 6-DOF robotic arm (UR5e) equipped with a TMS applicator and a 6-axis force/torque sensor, guided by a stereo vision camera.
   • Registration: The system performs dynamic registration between the virtual 3D head model and the real patient's head using AI-detected facial landmarks (e.g., eye midpoints, nasal tip).
   • Optimization: To address the inherent uncertainty of AI-based landmark detection, the system employs a dynamic optimization algorithm that refines the virtual model's landmarks to minimize registration errors between the camera space and the rendering space.
   • Execution: The robot positions the applicator at the calculated target and maintains a constant contact force (approx. 4 N) via admittance control to ensure safety and stability during head movement.

Key Contributions

• Unified Framework: The paper bridges the gap between geometric surface reconstruction and surface-to-brain mapping, providing a principled explanation for estimating intracranial targets from facial data alone.
• Generative Modeling: It demonstrates the feasibility of using a diffusion-based generative model to reconstruct complete cranial surfaces from partial facial scans, eliminating the need for manual scalp measurements.
• Scalp Angle Coordinate (SAC) System: A novel coordinate system is proposed that facilitates cortex-to-scalp mapping using only three anatomical landmarks, reducing error accumulation compared to geodesic-based methods.
• Markerless Automation: The integration of this framework into a markerless robotic system removes the need for physical fiducial markers, reducing setup time and operator dependency.

Results

• Surface Reconstruction Accuracy: In a test set of 503 MRI cases, the generative model achieved a mean reconstruction error of less than 1.05 mm for most scalp regions, with higher errors observed in the occipital area and near the ears. Consistency tests across three operators showed high intra-class correlation (ICC > 0.89 for mean distance).
• Targeting Accuracy (In Silico): Using the SAC-based mapping on 292 healthy participants, the mean scalp targeting error ranged from 2.36 mm to 5.70 mm, and cortical targeting error ranged from 2.50 mm to 8.81 mm. Validation on 18 insomnia patients showed that while errors increased slightly with scanner-based inputs compared to MRI-based inputs, the cortical targeting errors for key regions (M1 and DLPFC) remained comparable to other datasets.
• Phantom Validation: On a 3D-printed head phantom, the optimized markerless navigation system achieved a mean positional error of < 5 mm and an angular error of < 3°, outperforming non-optimized landmark methods.
• Clinical Validation (In Vivo): In a study of 18 insomnia patients treated with 1 Hz rTMS targeting the right DLPFC over two weeks, significant improvements were observed in sleep quality (PSQI scores decreased from 13.39 to 8.39; ISI scores decreased from 18.56 to 12.17). Crucially, the study found no significant correlation between the magnitude of targeting estimation errors and the therapeutic outcomes, suggesting the framework's accuracy is sufficient for clinical efficacy.

Significance and Claims
The authors claim that this framework offers a principled, explainable, and automated solution for MRI-free neuromodulation targeting. By integrating fragmented anatomical correlations into a unified workflow, the system addresses the limitations of current manual methods (low accuracy, high variability) and MRI-dependent systems (high cost, low accessibility).

The paper posits that this approach:

• Lowers Clinical Barriers: It makes precise neuromodulation accessible to community hospitals, psychiatric clinics, and rehabilitation centers where MRI is unavailable.
• Reduces Patient Burden: It eliminates the need for pre-treatment MRI scans and manual marker placement, shortening the treatment workflow to approximately 10–15 minutes.
• Enables Automation: The markerless robotic system paves the way for future automated or unmanned neuromodulation services.
• Demonstrates Feasibility: The successful treatment of insomnia patients without MRI guidance validates the clinical utility of the framework, suggesting that the estimation errors fall within the acceptable range of effective stimulation spread.

The authors remain modest regarding limitations, acknowledging that the paired training data is currently limited, that the linear mapping assumption may not perfectly capture complex physical field distributions, and that the in vivo sample size was small. They suggest future work will focus on expanding datasets, improving domain adaptation for scanner variations, and potentially developing wearable robotic devices.