1.

Introduction

In neuroimaging modalities, precise spatial alignment of probes (e.g., optodes and electrodes) on the scalp is pivotal. It may enhance spatial resolution, improve source estimation, strengthen between-subject reliability, and thus deepen our understanding of cerebral function.¹ Conventionally, in the absence of spatial registration, the probes are positioned on the cranial surface following the international 10-20, 10-10, or 10-5 electroencephalogram (EEG) systems.² These systems use relative positions that are individually determined based on the head circumference (i.e., the total distance from the inion to the nasion). To facilitate the probe mounting process, manufacturers offer an assortment of caps designed to accommodate a range of head circumferences, with typical increments of 2 cm. However, this method does not guarantee consistent coverage across different brain areas across participants and is subject to expected deviations in relative positioning.³ Moreover, even minor displacements can markedly impact the accuracy of source estimation.¹ The challenges intensify when considering optical neuroimaging techniques, such as functional near-infrared spectroscopy (fNIRS) and high-density diffuse optical tomography. In a conventional fNIRS configuration, optical sources and detectors, referred to as optodes, are usually affixed to the scalp at an optimal distance of 30 mm, determined by the absorption characteristics of near-infrared light. Yet, the fixed distance between optodes renders it unfeasible to align the caps according to EEG systems, resulting in reduced anatomical consistency across participants with varying head sizes.⁴ To address these challenges, brain imaging studies necessitate a more accurate spatial registration methodology that is both anatomically precise and practical in terms of time and implementation. Traditional electromagnetic three-dimensional (3D) digitizers have been considered the gold standard for highly accurate spatial registrations.⁵ With the device, the separate probes can be marked on the head using an electromagnetic tracking method. However, their drawbacks include high costs, time-consuming consecutive probe registration, interference with medical implants (e.g., cochlear implants), and impracticality in certain populations.^6,7 To address these limitations, less cost-intensive photogrammetric approaches gained popularity, utilizing a single camera for video recording around the head, enabling parallel acquisition of multiple probes and faster data collection.^7,8 Yet, the computationally expensive and time-consuming reconstruction of 3D shapes from two-dimensional video frames poses challenges, limiting immediate assessment after recording.⁶ Adding a depth-sensing instrument (e.g., an infrared camera) to the setup offers multiple advantages. It facilitates quick 3D reconstruction, allowing immediate evaluation of the 3D mesh and avoiding additional experimental time due to lengthy reconstruction processes. Furthermore, the infrared camera is less sensitive to background information, allowing more effective recording under a variety of conditions. These advantages collectively make it the preferred choice for user-friendliness.

Despite these advantages, the efficacy of the system had not been validated under actual experimental (or clinical) conditions, nor were there clear guidelines for its use. Our goal was to fill these gaps.

In this technical note, we aim to evaluate the efficacy of a commercially available infrared 3D scanner as a potential solution for accurately localizing probes on the scalp of human participants. For evaluation, we used magnetic resonance imaging (MRI) scans as in Refs. 9 and 10. We also provide guidelines for data anonymization to promote data sharing among researchers and to ensure compliance with privacy and ethical standards. Finally, we provide the complete source code for anatomical mapping and integration into the MNE-Python library, a widely used neuroimaging data processing toolbox.^11,12

2.

Materials and Methods

The study followed the Declaration of Helsinki guidelines and was approved by the KEK-Bern (study number 2020-02978), and all participants provided written informed consent.

To estimate the spatial resolution, we registered probe positions using an infrared 3D registration method and performed MRI scans on five voluntary participants. We positioned an fNIRS optode holder cap on the participants’ heads and fixated nitroglycerin capsules inside the optode holders and on the nasion, left pre-auricular (LPA), and right pre-auricular (RPA) anatomical landmarks. The fNIRS cap used in this investigation is identical to that detailed in our published article.¹³ The cap is a custom assembly, constructed from the components provided in the manufacturer’s holder kit (Shimadzu, Kyoto, Japan). A total of 25 optodes were included in the study, spaced at an inter-optode distance of 30 mm. The configuration covers temporal, visual, and prefrontal brain regions. After mounting the cap, we performed an MRI scan on the participants. Subsequent to the MRI scan, the participants sat down on a chair, and we performed three successive infrared 3D scans under identical conditions [Fig. 1(a)].

Fig. 1

Study procedure. (a) Flowchart showing the different steps of our proposed 3D infrared scanning method (yellow box). We validated the method using MRI scanning (grey box). (b) We employed a custom-built fNIRS cap designed to evaluate accuracy across the temporal, visual, and prefrontal regions. To enhance detectability on infrared 3D scans, the holders were marked using colored 3D-printed markers or painted. Nitroglycerin markers were attached to the optode holders and anatomical landmarks, as depicted in the left image (captured after the MRI scan).

2.2.

Face Anonymization

The faces in the images captured by the 3D infrared camera were anonymized after scanning. This avoids any workaround to preserve participant confidentiality during the scanning procedure, such as covering the face or wearing a mask. For this process, we utilized 3D Builder (Microsoft Corporation, Redmond, Washington, United States, ver. 20.0.4.0). We imported the mesh files (.obj, .mtl, and .png) and kept the default unit of the files, and if the software offered it, we let it repair the mesh. In the first step, we oriented the mesh into a transverse view and rotated along the $Z$-axis until the nasion was on the $X$-axis [Fig. 2(a)]. Next, in a sagittal view, we used the split tool to remove the lower half of the mesh. We set the cutting plane to retain visibility of essential anatomical landmarks (nasion, LPA, RPA) and their proximities [Fig. 2(b)]. A cylinder was then added, positioned, and reshaped to cover the remaining face while preserving anatomical landmarks. Once the cylinder was correctly positioned, we used the subtract tool to remove the cylinder from the main mesh [Fig. 2(c)]. We saved the final mesh in .glb format to retain color information.

Fig. 2

Processing of 3D Mesh. (a) In 3D Builder, we rotated the mesh along the $Z$-axis to place the nasion on the $X$-axis. The head coordinate system is shown in the top right corner, and the applied orientation is shown in the bottom right corner of each step. (b) Using the split tool, we removed the lower half of the mesh. (c) We placed a cylinder to cover the face while preserving the necessary anatomical landmarks. (d) We labeled the anonymized mesh in MeshLab. We placed markers on the head using the PickPoints tool, with the resulting coordinates and labels listed on the right.

It is advisable to decide on the anonymization step beforehand as the anonymized mesh has a different coordinate system than the original mesh, necessitating subsequent adjustments in the labeling.

3.

Results

The inter-optode distance on the cap was fixed at 30 mm. The 25 optodes were connected to form a total of 24 channels for analysis. For five participants, it meant 125 data points for MRI-based scans and 375 data points for the infrared 3D scans, as it was repeated three times. The average head circumference of the participants was 56.6 cm with a standard deviation (SD) of 2.9 mm and a range of 54 to 61 cm. First, we analyzed the positioning of the nitroglycerin markers across participants (i.e., accuracy of standard cap fixation). On average, the Euclidean distance of these markers among different participants was 15.25 mm (SD = 8.0 mm), with the largest variance observed at optode “D1” which had a maximum error of 43 mm among participants. The accuracy of the infrared 3D registration was evaluated by comparing the infrared 3D registered positions with the reference positions obtained from the MRI scans. The average error with the landmark-based alignment was 5.69 mm (SD = 1.73). This error is significantly smaller than with standard cap fixation ($p<0.001$). The individual participants are visualized in Fig. S1 in the Supplementary Material. Using the landmark-free alignment, the error was 2.55 mm (SD = 1.01). The reproducibility of the infrared 3D registration process was evaluated by comparing the positions from consecutively performed infrared 3D scans. The average error with the landmark-based alignment was found to be 3.43 mm (SD = 1.62), and that with the landmark-free alignment was 1.72 mm (SD = 0.77). The average inter-optode distance for the selected channels was 29.28 mm (SD = 1.12) based on the MRI scans and 29.43 mm (SD = 1.96) based on the infrared 3D scans.

The acquisition of a single 3D mesh with the infrared 3D scanner takes 2 min, the reconstruction is almost instant, and the labeling process with a certain level of experience takes less than 5 min.^7,17 The results with the landmark-based and landmark-free alignment are summarized in Table 2.

Table 2

Registered optode positions.

4.

Discussion

In this technical note, we presented a method for performing spatial registration of probe positions for neuroimaging applications. Such improved mapping is pivotal, especially in optical measurements, where the fixed inter-optode distance poses challenges with the alignment of standard EEG landmarks. The utilized infrared 3D scanner is commercially available, has been applied in various healthcare applications, is cost-effective, and requires only an iPad to operate. It is also simple and easy to use, allowing 3D models of objects to be captured by simply walking around them, with an acquisition time of 2 min and almost instant reconstruction of a single 3D mesh.^6,17

We aimed to evaluate the system under realistic experimental or clinical conditions. Therefore, we tested on human participants, and we used MRI scans to determine the precise positions of optode holders, which served as our reference for evaluation.

First, we evaluated the accuracy of different cap fitments by comparing the registered positions among participants based on MRI. This error is particularly important for studies that do not include spatial registration. In these cases, the measurement cap is typically placed according to the EEG system, and it is assumed that the measurement channels cover the corresponding EEG positions. However, this can never be achieved due to the fixed inter-optode distance. The average Euclidean distance between participants was 15.25 mm (SD = 8.0), with one optode showing a maximum error of 43 mm (i.e., the accuracy of standard cap fixation, Table 2). This large optode displacement, if uncorrected, could lead to misinterpretation in the analysis, as these channels might cover regions that have different signal characteristics and different functions.^4,18,19

Then, we determined the spatial accuracy of the infrared 3D scanning method by comparing its results with the MRI-derived reference points. By performing this comparison using landmark-based alignment, we aimed to provide a robust estimate of accuracy, as optode positions are determined in relation to anatomical landmarks. This comparison yielded an average error of 5.69 mm (SD = 1.73) (i.e., the accuracy of 3D registration using landmark-based alignment, Table 2). The spatial registration provides a significant improvement, compared with the standard cap fixation. Comparing our method with other, even closely related, methods is challenging due to differences in the determination of ground truth, the number of participants, and the number of anatomical landmarks used, all of which affect the final accuracy. Nevertheless, methods using comparable or even more accurate ground truth information reported slightly higher errors using landmark-based alignment both for photogrammetric and electromagnetic techniques (Table 3). An overview of previously reported methods can be found in Table S1 in the Supplementary Material. Taberna et al.²¹ reported the Euclidean distance between the detected electrode positions and their closest points on the MRI-based head shape, rather than MRI-derived reference points, making direct comparison difficult. Jaffe-Dax et al.⁷ used reference points and reported a smaller error of 3.4 mm. The reason for this is that they used additional landmarks mounted on the head and cap, therefore allowing a non-rigid registration (i.e., deformation of the head shape). However, this approach requires additional cap preparation time for marker placement, and due to the photogrammetric approach, the reconstruction time constrains the immediate feedback of the recording.⁶ The advantage of our method is the immediate reconstruction and the use of visible landmarks on the head, eliminating the need for any additional markers. Notably, the SD of our method’s accuracy (1.73 mm) was lower than expected with the standard EEG cap (5.5 mm, as per Ref. 3). This suggests that the infrared 3D registration method can support not just fNIRS but also EEG studies by offering improved spatial accuracy.

Table 3

Accuracy of different registration methods.

To provide further insights into the accuracy of the system and to ease the comparison with other methods, we conducted additional analyses:

In conclusion, our study presents a comprehensive solution for integrating spatial registration into neuroimaging research protocols. We demonstrate the use of a commercially available device for measurements and 3D reconstructions, along with guidelines for data anonymization and labeling. The acquired positions can be loaded and processed in the toolbox of choice. We provided a solution for MNE-Python integration. The scripts to reproduce these results are freely available at Ref. 23. To apply to individual data, researchers need to save their optode positions into a .pp file for each participant (described in Sec. 2.3) and specify the channels used in their unique montage in the script.