EEG electrode localization with 3D iPhone scanning using point-cloud electrode selection (PC-ES)

Objective. Electroencephalography source imaging (ESI) is a valuable tool in clinical evaluation for epilepsy patients but is underutilized in part due to sensitivity to anatomical modeling errors. Accurate localization of scalp electrodes is instrumental to ESI, but existing localization devices are expensive and not portable. As a result, electrode localization challenges further impede access to ESI, particularly in inpatient and intensive care settings. Approach. To address this challenge, we present a portable and affordable electrode digitization method using the 3D scanning feature in modern iPhone models. This technique combines iPhone scanning with semi-automated image processing using point-cloud electrode selection (PC-ES), a custom MATLAB desktop application. We compare iPhone electrode localization to state-of-the-art photogrammetry technology in a human study with over 6000 electrodes labeled using each method. We also characterize the performance of PC-ES with respect to head location and examine the relative impact of different algorithm parameters. Main Results. The median electrode position variation across reviewers was 1.50 mm for PC-ES scanning and 0.53 mm for photogrammetry, and the average median distance between PC-ES and photogrammetry electrodes was 3.4 mm. These metrics demonstrate comparable performance of iPhone/PC-ES scanning to currently available technology and sufficient accuracy for ESI. Significance. Low cost, portable electrode localization using iPhone scanning removes barriers to ESI in inpatient, outpatient, and remote care settings. While PC-ES has current limitations in user bias and processing time, we anticipate these will improve with software automation techniques as well as future developments in iPhone 3D scanning technology.


Introduction
Electroencephalography (EEG) continuously detects electrical activity in the brain using voltage measurements on the scalp and is currently used to diagnose and monitor neurological conditions including epilepsy, traumatic brain injury, psychiatric disorders, and brain tumors [1][2][3].Advantages of EEG include non-invasive, continuous monitoring and high temporal resolution [4][5][6].One important EEG application is the EEG source imaging (ESI) technique, which utilizes sophisticated modeling to spatially estimate the source location within the brain responsible for generating specific electrical potentials found on the scalp.This technique can be used to perform functional mapping or localize paroxysmal brain activity [7].Clinically, ESI adds critical localization and diagnostic value during presurgical planning for refractory epilepsy patients and has been shown to improve the surgical outcomes in these patients [8][9][10].
Despite its clinical and experimental benefit, ESI is underutilized due to cost, anatomical modeling requirements, and processing time [11].ESI is sensitive to modeling errors [5,12], requiring an extensive anatomical modeling process which further limits practical use.Anatomical modeling in ESI typically involves a head model with intracranial features derived from structural imaging and electrode locations derived from in-clinic measurements [13,14].There is general consensus that inaccurately modeled electrode positions result in ESI error.While some studies have found negligible electrode position-based ESI errors in low-density EEG [5,15], other studies in high-density EEG (hdEEG) have found a more significant effect wherein a 1 cm electrode location misplacement can result in a 2 cm ESI estimation error [16].Regardless, ESI accuracy improves with both a high number of electrodes (recommended ⩾ 64) [6] and accurate electrode localization relative to anatomical landmarks [5].Thus, ESI benefits from a dense electrode distribution on the head known as hdEEG [9].Given the importance of electrode localization in ESI and current drawbacks, efficient and accurate electrode localization is an active area of commercial development and research.
Electrode localization approaches include electromagnetic (EM) and ultrasound digitizers [17], photogrammetry systems [18], and structured-light (visible or infrared) systems [16,19].There are trade-offs with each localization technique: EM digitizers can require long recording times (7 min [19]) and yield small to moderate errors (1.0-6.8 mm).Photogrammetry has high accuracy (0.56-1.3 mm error), but requires large, non-portable hardware (figure 1) and long processing times (∼30 min) [18].Structured light approaches are portable but can require subjects to remain still for ∼2 min [19], may be at risk of triggering seizures in certain epilepsy patients (assuming visible structured light systems), and report moderate-to-large errors (median errors of 1.14 [20] to 9.4 mm [19]).None of these techniques can be reliably utilized at the bedside or in the critical care setting without interfering with the routine clinical care.
Here we present, for the first time, a portable and cost-effective electrode localization method using iPhone 3D scanning technology (figure 1).iPhone scanning does not require the patient to remain still for more than a few seconds at a time and can be performed anywhere with approximately 6 inches of space around the head.An opensource custom-Matlab software solution, which we refer to as point-cloud electrode selection (PC-ES), was developed to process the scans and is available on Github [21].The 3D iPhone scans and software were evaluated on 8 subjects with 256-electrode hdEEG caps and compared with results from a Phillips Geodesic Photogrammetry System 3 Research (GPS).Three reviewers localized electrodes using iPhone scan data with PC-ES and photogrammetry data with Geodesic software, resulting in over 6000 electrode locations from each method.The solution proposed here appears competitive to other recent work [19], especially when considering the ubiquity and practicality of iPhones, and relatively low cost.This method could be seamlessly integrated into emergency, ambulatory, or research settings.Comparisons to other scanning solutions in terms of accuracy and scan times are provided in the discussion section.

Methods
In this section we discuss (1) the iPhone data collection process, (2) the PC-ES software and processing steps, (3) subject data collection details, and (4) quantitative and qualitative metrics used to analyze the results.

iPhone data collection
For iPhone data collection, this study uses the Heges 3D scanner application [22], available on iPhone 11 and later models.Heges scanning utilizes the front (FaceID) sensor of an iPhone 12 Pro model-a structured infrared 3D sensor.Heges was preferred due to high precision, protected local data storage, and the capture of detailed scan information including points, colors, and a triangulated surface.However, PC-ES can accommodate iPhone scans taken with other 3D scanning applications exported as point clouds (.ply files).Scans were taken at 1 mm resolution and maximum range.The FaceID scanning field of view was not able to capture a subject's full head in a single scan (see figure 2).We therefore recorded ∼5 overlapping scans generally centered on the front, left, back, right, and top of the head (see figure 2(A)).Subjects only needed to remain still for each individual scan (∼10 s), and the total time required to scan a full head was <2 min.

PC-ES: electrode localization software
PC-ES is a custom Matlab GUI used to label electrodes in each individual scan and merge scans to create a 3D surface of the entire head (see figure 2).The software was designed specifically for the Heges app, but is compatible with any .plyfile containing a partial or full scan of a 256-electrode cap.After uploading each scan, the manual processing steps involve (A) cropping and de-identifying the scan, (B) selecting and labeling fiducial electrodes, and (C) selecting all visible electrodes without labeling (figure 3), referred to as selected electrodes.Cropping is performed to improve processing speed and remove unnecessary background features in the scan.The fiducial electrodes are electrodes with easy to identify labels, determined either by 5-colored stickers placed on electrodes prior to scanning or 17 widely-spaced black electrodes present within the caps used in the study.Obtaining the positions of fiducial electrodes and selected electrodes is optimized through an intuitive user interface.PC-ES presents 4 close-up views of each scan at different angles and projects the current view into a 2D-projected selection plane, improving scan visualization and minimizing processing time.Example views of the PC-ES interface display with fiducial and selected electrodes can be seen in supplementary figures S.1 and S.2.

Automatic labeling algorithm
After the user has made manual selections on all five 3D scans covering a subject's head, automatic electrode labeling is performed (figure 3).The automatic labeling algorithm relies on a nominal 256electrode EEG cap, which we derived from a generic cap model and sensor adjacency information from the GPS software.The algorithm begins with the 4+ manually selected fiducial electrodes, and iteratively attempts to label all the selected electrodes.It first performs a best-fit registration between the fiducial electrodes and the nominal cap, allowing for scalar stretching of the nominal cap in the x,y, and z-directions (see figure 4(A)).It then loops through the current set of labeled electrodes.If there is a corresponding electrode from the nominal-cap within a distance threshold, D A (where A refers to automatic), then the algorithm steps out to adjacent nominal-cap electrodes and attempts to label selected electrodes (figure 4(C)).Electrodes are labeled if there is an unlabeled selected electrode within a distance D A of the adjacent nominal-cap electrode (see  gold sphere in figure 4(C)), otherwise no labeling is performed.When the algorithm finishes looping through electrodes adjacent to currently labeled electrodes, it updates the fit to the nominal cap using the new labeled electrodes.This process is repeated until no new electrodes are labeled.At this point in the automatic processing, we refer to all labeled electrodes (including fiducial electrodes) as auto-labeled electrodes.

Merging algorithm
The merging algorithm first looks through all autolabeled electrodes across all scans and identifies electrodes that have been labeled in multiple scans due to scan overlap.The algorithm counting the number of overlapping electrodes between all pairs of scans and merges the scans with the most common electrodes using a rigid transformation between matching auto-labeled electrodes.Scans are only merged if the root mean square (RMS) error fit is <15 mm, which was heuristically determined.The algorithm then finds the individual scan with the most common auto-labeled electrodes to the current merged scan and repeats the process until all scans have been merged.If the RMS is >15 mm between an individual scan and the merged scan, then the individual scan is removed.The resulting labeled electrodes in the merged scan are referred to as merged electrodes.

Filling gaps algorithm
The filling-gaps algorithm estimates the location of the missing electrodes (fill-gaps electrodes) to generate a full set of 256 electrodes, which we refer to as full-set electrodes.The algorithm first performs a best-fit transformation between all labeled electrodes and the nominal cap and identifies the points on the nominal cap corresponding to missing electrodes.It finds the missing electrode with the most neighboring labeled electrodes within a distance threshold, D F , and labels it and assigns the location to be the nominal cap estimation.It then updates the fit given the newly labeled electrode and repeats the previous step, until all the fill-gaps electrodes are found, creating a full set of 256 electrodes (full-set electrodes).

Projections to the scalp
The final position of every electrode used in the analysis is determined by projecting the point location from the surface of the electrode onto the scalp.We do this by taking the average outward pointing normal direction from all surface triangle elements that fall within 3 mm of the electrode position and projecting the point inwards for a distance equal to the electrode thickness, which is manually entered in the PC-ES program interface.

Subject scans
To validate the clinical utility of the iPhone scanning technique, we conducted a study comparing iPhone and photogrammetry electrode localization.We collected iPhone and photogrammetry data in the same visit for 8 subjects: 3 epilepsy patients, 3 healthy volunteers, and 2 phantoms.The human subjects were enrolled under an IRB-approved study evaluating hdEEG at Dartmouth-Hitchcock Medical Center (DHH #30813).Given that both iPhone and photogrammetry techniques involve significant user input that could introduce bias, all 8 iPhone and photogrammetry datasets were processed independently by 3 separate reviewers, for a total of 6144 electrodes labeled using each method.The hdEEG caps used were Microcel GSN 256-electrode caps with 8-or 12 mm thick electrodes.

Metrics
In the following section, we quantify the number of scans, fiducial electrodes, selected electrodes, and performance metrics.The quantitative performance metrics are the RMS error between the iPhone scan and photogrammetry electrode coordinates, with photogrammetry used as the gold standard measurement for this study.The RMS values are evaluated as a function of the distance thresholds used in the automatic labeling algorithm (both D A and D F ). Qualitative evaluation is performed by comparing full 3D scans, as well as 2D stereographic projections, which allow for a 2D qualitative view of all the electrodes around the subjects' heads.In the projection, The full set of 256 electrodes points are projected up to an x,y-plane at z = 1 from an assumed origin P 0 .In detail, a given point (x, y, z) is projected to where t = (1 − P 0,3 ) / (z − P 0,3 ).Table 1 provides a summary of how electrodes are categorized and referred to in this study.

Results
There were 5.1 ± 0.8 scans (average ± standard deviation) ranging from 4 to 7 across the 8 subjects.Electrode selection using the PC-ES interface resulted in 4.5 ± 1.5 (3-7) fiducial electrodes and 88.1 ± 15.6 (68-110) selected electrodes per scan across all reviewers and subjects.The average electrode locations across reviewers for photogrammetry and PC-ES from the full-set electrodes were projected into 2D along with a heatmap of the corresponding standard deviation from the average electrode location (see figure 5).These projections demonstrate high precision with a standard deviation <1 mm (dark blue) over large regions for all subjects using photogrammetry and all subjects except Subject 6 for PC-ES.The overall median and mean ± standard deviation of the electrode differences from photogrammetry alone were 0.53 mm and 0.95 ± 1.48 mm, respectively, whereas for PC-ES they were

Threshold determination
The automatic labeling (D A ) and filling-gaps distance thresholds (D F ) were explored over a wide range of values.We report RMS errors at different stages of automatic processing (auto-labeled, merged, and full-set electrodes) as a function of (D A ).We further report the corresponding number of merged electrodes found (figure 6(A)) and the RMS of the fullset electrodes in figure 6(B) versus the filling-gaps distance threshold.Informed by these metrics, the optimal value for D A was heuristically chosen to be 10 mm, aiming to balance obtaining the smallest RMS error while achieving a maximum number of electrodes found.The best value for D F was chosen to be 150 mm to balance the slightly rising RMS error with the decreasing standard deviation as D F increases (figure 6).At D A = 10 mm and D F = 150 mm, the RMS errors were 4.5 ± 1.8, 4.5 ± 1.3, and 5.4 ± 1.5 mm for the auto-labeled, merged, and full-set electrodes respectively, with 216.2 ± 34.7 electrodes found across all subjects and reviewers.

Quantitative and qualitative assessment
Tables 2 and 3 show the RMS values and number of electrodes found for each subject and each reviewer.The RMS values vary across both subjects and reviewers, suggesting electrode localization accuracy depends on the quality of both the iPhone scan and processing.Figures 7 and 8 show the full head scan and 2D projections with RMS values (for full-set electrodes) respectively, for every subject and every reviewer.Overall, the 3D head scans appear realistic and exhibit only minimal visual artifacts from scan merging, with a few exceptions such as Subject 4 processed by Reviewer 1.Additionally, there is only one case in which a partial head was reconstructed (Reviewer 1, Subject 6), for which scan reprocessing could improve the results.In this case, one of the scans was removed due to the 15 mm merging threshold.Two other scans were omitted because of the merging threshold (Reviewer 1 on Subject 4 and Reviewer 2 on Subject 8), but due to overlaps in scans their merged scans still qualitatively appear complete.The projections in figure 8 reveal that errors are generally largest in the face and at the posterior base of the skull.Additional factors were explored on individual scans to better understand the sources of errors, as shown in figure 9. Specifically, figures 9(A)-(E) shows multiple combinations of RMS of auto-labeled electrodes from separate scans, number of fiducial electrodes and number of auto-labeled electrodes, and individual scan percent areas.we obtain more auto-labeled electrodes if there were more fiducial electrodes.The scan area comparisons (figures 9(D) and (E)) appear to imply that a larger scan will increase the RMS errors, which could indicate that the scans themselves contribute to the localization errors observed.This was not significant when considered across subject head volumes (supplementary figure S.6).Figures 9(A) and (F) do not yield a significant correlation, which indicates that the number of fiducial electrodes and average number of overlapping (auto-labeled) electrodes do not clearly impact the resulting RMS.

Timing
The time required for manual processing was somewhat large, averaging 15-20 min per scan.Given that there were generally 5 scans, this yielded a total manual processing time per subject of 75-100 min.Comparatively, photogrammetry processing time is reported as 30-40 min [18].This is a current drawback that could be improved by software automation strategies such as machine learning.After the manual processing is complete, the automatic labeling and merging take on average 7.8 and 7.6 s, respectively-a negligible duration compared to manual processing.

Discussion
The results presented here suggest that PC-ES may provide a feasible and accessible tool for aiding accurate ESI and its widespread utility.ESI is increasingly used for both research and clinical applications as it allows noninvasive localization of electrical activity and understanding of cognitive processes within the brain [24].Clinically, ESI can be used to diagnose, monitor, and help guide therapies for pathologic, physiologic, and functional neurologic disorders.Example use cases include, but are not limited to, surgical planning for epilepsy treatment, attention deficit hyperactivity disorder diagnosis and neurofeedback monitoring [11,25].However, ESI is limited by its precision, accessibility, and ill-posed modeling constraints.To enhance the accuracy and resolution of ESI, MRI/CT patient-specific meshes are often used along with precise electrode localizations collected from an EEG lab photogrammetry (or equivalent) system.Such systems are large, static, expensive and require specialized trainingultimately limiting their accessibility.
We have shown here the first results supporting PC-ES as a potential mobile, low-cost, opensource solution for ESI in a study covering over 6000 registered electrodes.To reduce bias, three reviewers processed each of the 8 subjects, and photogrammetry was used as the clinical gold standard for comparison with the PC-ES program.The overall median error distances for all subjects and all reviewers for the auto-labeled, merged, and full-set electrodes were 3.1 ± 1.3 mm, 3.0 ± 0.7 mm, and 3.4 ± 1.0 mm, respectively.An average of 216.2 ± 34.7 out of 256 electrodes were found across all subjects and reviewers.Subject scan times were completed in less than 2 min total with less than 10 s required per scan.Full processing time from iPhone scanning to completion was approximately 75-100 min and all scans were easily deidentified prior to processing.While variability was present across reviewers (table 3, figures 8 and S.4, S.5) the differences in median RMS were not significant (p > 0.05), demonstrating stability of PC-ES across users and subjects.
While PC-ES is comparable to alternative approaches such as optical, EM (Polhemus), and IR scanners (1.1 mm-9.4 mm) [18][19][20][26][27][28], our errors are higher than the reported photogrammetry errors (0.56-1.3 mm) [18].Other comparable algorithms and digitization approaches under development include structured sensor 3D scanning [19,20], video processing [28] and SLR photogrammetry [27].To our knowledge no other study has presented iPhone-based (or mobile phone-based) results, a choice made to improve accessibility and affordability of the PC-ES approach.While these algorithms achieved comparable or lower error values, it is important to emphasize the results presented here represent a 256 electrode hdEEG cap.The accurate capture and localization of 256 electrodes adds significant complexity over a cap of 68 or 128 electrodes, for example.Additionally, by incorporating multiple reviewers and subjects we aimed to investigate the robustness and repeatability of the PC-ES method; this approach incorporated real-world variability and error to the PC-ES results as compared to single subject studies (supplementary table S.1).Lastly, the PC-ES median error of 3.0 mm (merged) or 3.4 mm (full-set) includes the inherent photogrammetry and scanning resolution error (i.e. to be conservative we permitted for compounding errors in our final reported values).This is particularly noteworthy as the best accuracy structured light study (1.1 mm median error), to the best of our understanding, did not utilize a clinically-validated gold standard for registration distance error comparison [20].Thus, these results may present, for the first time, the most accurate structured light mobile application for patient-specific electrode localization when compared to a clinical gold standard.
While these results support the feasibility of PC-ES as a research or clinical tool, several limitations were present in the study which may be contributing to the observed errors.First, the scanner quality and third-party Heges software inherently limit the precision and accuracy of PC-ES algorithms.The iPhone 12 FaceID scanner was reported to have a 0.5 mm accuracy [29], and the Heges software was found to be most effective at 1 mm resolution.Both limitations are well-positioned to improve with advancements in mobile 3d scanning sensors, GPU processing speed (e.g.70% increase in iPhone 15), and available 3rd party processing applications (e.g.SureScan 3D and many alternatives).In the future, alternate applications and scanning approaches will be explored.PC-ES accepts any PLY source file which include points, colors, and a triangulated surface, allowing it to be application agnostic-i.e.ready to evaluate with other sensors/applications.
Second, scanning environment has the potential to affect electrode reconstructions, as insufficient light may partially explain the increased position error for electrodes at the base of the skull.Further, the decrease in performance significantly correlating to the increase in scan area (figures 9(D) and (E)) suggests that the scanning technique itself may be contributing significant error.This may be compensated for with collection of additional scans and use of more controlled ambient lighting.To test the effect of increasing number of scans we collected data on an additional well-lit phantom subject and increased the total scan number incrementally from 5 to 15. Increasing scan number decreased merged electrode position error from 3.4 mm to 2.8 mm, and full-set error from 4.3 mm to 3 mm (supplemental figure S.7).Thus, additional improvement may be realized by increasing the number of user-collected scans, as well as ensuring a well-lit ambient environment.
Third, the manual processing required is time intensive (∼10-15 min per scan) and introduces reviewer-based variability.In the future it is feasible that a scan of the whole head may be collected removing the need for the merged scan step.Alternatively, standardized training may help reduce reviewer variability and future work will focus on further automation to reduce the manual processing burden.Fourth, as mentioned above, the use of an hdEEG 256 electrode cap may have introduced additional variability.As seen in figure 8 the majority of errors are present around the face and lowest electrodes on the back of the head, areas only sampled under an hdEEG layout.While hdEEG is a promising area of research, other standard clinical EEG electrode caps range from 32-128 electrodes.If we downselect from our 256 electrode cap to prioritize the regions of the scalp closely matching electrodes of a traditional clinical EEG cap, we see a decrease in our median RMS error across all subjects and all reviewers (figure 10).Specifically, average median errors of 2.85 and 2.93 mm are observed using the electrodes which correlate to the international 10-20 or 10-10 systems, respectively.
The last limitation potentially contributing to errors is our inclusion of all data.When considering the performance across subjects and reviewers, it is apparent that Subject 6 had consistently higher errors present when compared to all subjects.To be conservative in this study we did not remove hypothesized mislabeled electrodes or have a reviewer reprocess a scan that had high errors.In this way we best captured the likely outcomes of this preliminary software under basic training should it be utilized by others.However, if we recognize Subject 6 as a potential outlier, likely due to an imprecise iPhone scan or mislabeled electrode and complete the same analysis without Subject 6 we see an overall improvement in total error (figure 10 red dashed line) with a minimum median error of 2.41 mm.Further, if excluding Subject 6 the resulting median errors for the international 10-20 and 10-10 systems are 2.77 mm and 2.82 mm, respectively.These sub-analyses further support the potential performance and utility of PC-ES.
Beyond clinical applications, PC-ES has the potential to be useful in settings where photogrammetry and MRI imaging are less accessible, such as mobile, acute or research locations.It is important to note that mobile low-cost patient-specific meshing and precise electrode localization have potential implications far beyond localizing interictal epileptiform activity in epilepsy patients.Potential additional clinical applications for PC-ES may include transcranial magnetic stimulation, traumatic brain injury rehabilitation, neurofeedback-based ADHD regulation, brain-computer interface, and behavioral research studies.Additionally, PC-ES as a solution is well positioned for future performance improvement as technologic continue to develop.Specifically, the Heges application <1.0 mm precision settings, which would enable significant improvements in electrode identification, scanning errors, and automation abilities; however, would regularly crash under those settings making data collection impractical.As iPhone (and mobile) scanning technologies become more precise, accessible, and robust, the PC-ES software will be further enhanced.
Future work will primarily focus on automating user-driven manual processing steps in the PC-ES software and improving computational efficiency.With further development and image processing integration a fully automated algorithm should be feasible.Should full automation not be possible, fiducial-based automated electrode selection between scans, similar to the current photogrammetry software, will enable significant reduction in manual processing time and is under development.Additionally, a patient study is currently underway to compare the ESI accuracy between PC-ES and photogrammetry within an epileptic patient cohort, although as noted above the optimal utility may be beyond the surgical application.Future resources under development will include training materials and continued support from our lab for adoption and integration of the PC-ES software for research use.

Conclusion
This study presents, for the first time, PC-ES: an open source [21], low cost, portable electrode localization software application.PC-ES utilizes iPhone scanning to remove barriers to ESI in clinical, ambulatory, and research settings.While PC-ES has current limitations in processing time and RMS error, we anticipate these will improve with software automation, continued work and future developments in iPhone 3D scanning technology.

Figure 2 .
Figure 2. (A).Five individual overlapping 3D scans from a gelatin head phantom, and (B).Two views of the corresponding merged full-head 3D scan where individual scans are colored corresponding to their scan location.

Figure 3 .
Figure 3. Block diagram of the manual and automatic processing steps.The manual processing involves cropping and de-identifying the scan, followed by finding fiducial electrodes, and selecting (unlabeled) electrodes.The automatic process involves automatically labeling the electrodes per scan, merging scans, and filling gaps for a given subject.

Figure 4 .
Figure 4. (A).An example partial scan with 5 (manually-labeled) fiducial electrodes (green dots within spheres representing distance thresholds at these locations), and a best-fitting nominal EEG cap (blue dots and lattice).(B).Shows a zoomed-in view of one electrode from (A). (C).Shows the same zoomed-in view illustrating how adjacent electrodes are labeled.The 5 adjacent nominal-cap electrodes (cyan circles) determined via the lattice, and the yellow distance-threshold sphere (DA = 10 mm) are shown.If there is an (unlabeled) selected electrode (red dot) that lies within the distance threshold, then the nominal EEG electrode label is given to the corresponding electrode.
Figure 9(F) shows RMS from the merged scans versus the average number of overlapping (auto-labeled) electrodes between those scans.Comparisons in figures 9(B) and (C) yield intuitive results: 9B implies the RMS decreases if there are more fiducial electrodes, and 9 • C implies

Figure 6 .
Figure 6.(A).RMS errors from auto-labeled, merged, and full-set electrodes, and the corresponding number of merged electrodes found versus the automatic labeling distance threshold.B. RMS errors of the full-set electrodes versus the filling-gaps distance threshold.Note DF = 150 mm in A and DA = 10 mm in B.

Figure 7 .
Figure 7.View of the merged full-head iPhone scans from the 8 subjects performed by Reviewer 3.These images give a qualitative view of the resulting merged scans.The QR-code allows with Schol-AR[23] an augmented reality view of 8:P2, which allows for rotations of the scans.

Figure 8 .
Figure 8. Projection of the 8 subjects performed by Reviewer 3 (see supplemental figures S.4.and S.5.for Reviewers 1 & 2).Black dots represent merged electrodes that were labeled via the auto-labeling algorithm whereas red dots depict filling-gaps electrodes.The colormap depicts the RMS of the full-set electrodes giving a visual depiction of where errors occur.The bottom right image of Subject 8 shows the colormap on a view of the 3D scan.One can additionally view Subject 8 using augmented reality via Schol-AR [23].

Figure 9 .
Figure 9. (A)-(E) Shows multiple combinations of RMS of separate scans, number of electrodes, NE, before and after auto-labeling (i.e.number of fiducial versus number of auto-labeled electrodes), and individual scan percent areas (relative to the area of the full merged, head scan).(F).Shows RMS from the merged scans versus the average number of overlapping electrodes between those scans.Red lines show linear regression results with p-values noted.

Figure 10 .
Figure 10.Electrode selection implications on median error results.Top: sub-selected scalp electrodes (magenta ellipse) extend from above the left ear to right ear and from the forehead electrodes (front) to Oz. Bottom: average median error of the sub-selected electrodes across all reviewers and subjects (blue line) and excluding Subject 6 (red dashed line) by number of electrodes selected (centered on top of cranium).Electrodes and corresponding electrode numbers are determined by expanding or contracting the magenta ellipse for each subject and reviewer.The average median errors are 2.86 and 2.93 mm for the international 10-20 and 10-10 systems, respectively.

Table 1 .
Definitions of all the types of electrodes considered.

Table 3 .
RMS errors (average ± standard deviation) and number of merged electrodes found per reviewer.