The emergence of non-invasive transcranial brain mapping techniques, such as functional Near-Infrared Spectroscopy (fNIRS) and Transcranial Magnetic Stimulation (TMS), extends our ability to functionally map the human brain. FNIRS can monitor neural-activity-related hemodynamics taking place in specific areas with optode dyads set on the scalp surface (Scholkmann et al., 2014). Due to its advantages in portability, affordability, and insensitivity to head motion (Boas et al., 2014), fNIRS has demonstrated potential in studies on children (Vanderwert and Nelson, 2014) and patients (Ehlis et al., 2014) and in studies that require high ecological validity (Cui et al., 2012). TMS can regulate brain activity by inducing an intense magnetic field from a coil set on the scalp. With different parameters, TMS can either excite, or inhibit the focal area beneath the coil (Hallett, 2007), thus providing a tool to investigate the causal relationship between brain and behavior (Pascual-Leone et al., 2000). It is also a potential therapeutic route for psychiatric disorders (Lopez-Ibor et al., 2008).

When using these transcranial techniques, an important prerequisite is to accurately place the fNIRS optodes or the TMS coil at correct positions on the scalp to ensure the coverage of cortical regions of interest (ROI). Currently, this procedure is facilitated by two classes of methods. One class is MRI-dependent methods, which use the subject's own structural magnetic resonance (MR) images to guide placement (Herwig et al., 2001). Such methods provide high accuracy: error of a few of millimeters (Sparing et al., 2008). However, MRI scanning is expensive and uncomfortable, and not always available in many institutes. Even when available, MRI scanning would exclude subjects with incompatible cardiac pacemakers, and the noisy and narrow scanning cavity may impose extra burden on subjects. Therefore, in practice, most fNIRS studies, and many TMS studies are conducted without MRI data. In such a situation, optode/coil placement based on the international 10–20 system (10/20), as a MRI-free probe placement method, is commonly adopted (Herwig et al., 2003; Sparing et al., 2008). The 10/20 system (Jasper, 1958) is a proportional landmark system for the scalp, consisting of four initial reference points on the head and landmark points defined at specific relative distances from the reference points. Studies using cadaver (Jasper, 1958), CT (Homan et al., 1987), and MRI (Lagerlund et al., 1993; Okamoto et al., 2004) have found that each 10/20 landmark on the scalp corresponds to a specific cortical area, and this cranio-cerebral correspondence can be generalized across different subjects (Okamoto et al., 2004). Accordingly, once the 10/20 landmarks are precisely identified on the scalp, desired cortical areas can then be accessed with optodes/coils properly set according to the 10/20 landmarks (Herwig et al., 2003).

However, this widely-used method faces problems in reliability and time cost, mainly due to the manual measurement procedure prescribed by Jasper. The 10/20 system consists of 21 points covering the whole head surface (including the Fpz and Oz). According to Jasper, it takes four steps to complete the identification of all 21 landmark points in a one by one manner. Moreover, identification of 10–20 landmarks in later steps depends on the positions of 10/20 landmarks determined in prior steps (Jasper, 1958; Milnik, 2009). Such a tedious and error-prone procedure, involving both manual measurement and marking of 10/20 landmarks on the head of the subject, makes it difficult for the experimenter to maintain high reliability. Such manual operation is also time-consuming. In our experience, it takes a moderately-proficient experimenter about 16 min to manually complete the whole process. Such a time consuming process imposes an extra and possibly heavy burden on subjects, and may reduce the ecological validity of an experiment. Therefore, manual 10/20 landmark identification has become a bottleneck of the 10/20-based MRI-free probe placement approach for transcranial mapping techniques, especially for methods with relatively high spatial specificity.

To address these issues, a semi-automatic approach is proposed in the present study to identify the 10/20 landmarks on the head of a subject in a fast and reliable way. The proposed method is validated by a set of simulation experiments. Then a real experiment was conducted to evaluate the proposed method in both reliability and time cost.

The International 10/20 system and the previously established cranio-cerebral correspondence (Okamoto et al., 2004) are important for transcranial brain mapping techniques such as fNIRS and TMS. Lacking cerebral anatomic information in MRI-free contexts, coils of TMS and optodes of fNIRS are usually set based on the 10/20 system to permit both proper ROI coverage and spatial consistency among subjects and experiments. Aiming to identify 10/20 landmarks on the head of a subject in a fast and reliable way, a semi-automatic 10/20 method based on a novel S3R_VM algorithm and a user-friendly visually-guided navigation system has been proposed in this study.

Performance of the proposed navigation method was evaluated with an experiment involving identification and marking of 10/20 landmarks on a medical head model. With respect to reliability, the navigation method provided an over-all variance at 2.93 ± 0.73 mm across 20 runs conducted by four different operators, which is about half of the variance from the traditional manual measurement approach (6.41 ± 1.14 mm). In terms of time cost, the navigation method also showed an advantage over the manual approach. Specifically, the time cost of the navigation method consists of three parts. The first part is the manual sampling of the head surface. Using an EEG cap to assist in uniformly sampling points, it takes about 1.5 min to complete the sampling process. The second part is the time to execute the S3R_VM algorithm, which is about 0.5 s for a modern desktop PC. The third part is the navigation process. On average, it takes 28 s to navigate to each of the 10/20 landmarks. Moreover, it is worth noting that operators are allowed to locate 10/20 landmarks in arbitrary sequence, since all of the 10/20 landmarks have already been identified in virtual space. For manual measurement, however, 10/20 landmarks must be measured in a sequence determined by the 10/20 rules (Jasper, 1958; Milnik, 2009). According to our experience, to manually measure Cz, i.e., the first landmark in the measurement sequence, the time cost is about 2 min. While, for P4, the last landmark in the measurement sequence, the time cost is about 16 min. Thus, to determine a single 10/20 landmark, time cost of the proposed method is similar to the lower limit of the traditional manual measurement time cost.

The navigation method also shows ergonomic advantages compared to traditional manual measurement. First, the proposed method allows the experimenter to avoid tedious measurements and calculations as well as memorizing the 10/20 measurement rules, thus reducing both the burden on the experimenter and the possibility for mistakes. Second, as indicated in the real experiment, a significant difference between inter- and intra- operator variance existed under manual measurement, but was not found under the navigation method (Figure 8). This indicates that the navigation method shows better robustness to operator experience and familiarity with the 10/20 system and thus is more suitable for beginners or less-proficient experimenters.

The crust algorithm used in the navigation method (Figures 2A,B) is a validated and prevailing interpolating surface reconstruction algorithm in the field of computational geometry (Amenta et al., 1998; Lim and Haron, 2012). However, when directly used on the sparsely sampled points (i.e., a small sample size) from the head surface, the crust algorithm does not provide good accuracy (Figures S2, S3). Considering that the human head is smooth and ellipsoid-like, we added a spherization step (Figure 2C). The combined procedure provides higher reconstruction accuracy, allowing smaller sample sizes (Figure S2). For example, when the sample size is set at 21, the average reconstruction accuracy of our combined procedure is 2.09 mm, while the crust algorithm alone provides an equivalent accuracy only when the sample size is increased to larger than 76. Moreover, as shown in Figure 5, When sample size is 21, the error of the resultant 10/20 landmarks virtually measured on the reconstructed head surface, i.e., output of the S3R_VM, will be under 3 mm. Another important character of S3R is that it is insensitive to variability in the sample points (i.e., allows sampling flexibility). As indicated in Figure 5, for sample sizes larger than 21, differences in reconstruction error between different sample point sets is <2 mm, and such differences decrease with increasing sample size. Such sampling flexibility makes the navigation method easy to use.

The theoretical definition of the 10/20 system was proposed long ago. However, how to actually determine the 10/20 landmarks remained ambiguous, as reviewed in Jurcak et al. (2007). To eliminate the uncertainty, efforts have been made in several studies (Jurcak et al., 2005, 2007). These studies eliminated the ambiguity in the location of the four reference points on the scalp and proposed a standardized measurement procedure based on the original definition of 10/20. This deterministic measurement rule was implemented with a virtual measurement algorithm on MR images. As a result, for a given set of MR images, positions of 10/20 landmarks can be uniquely determined. This deterministic way of finding 10/20 landmarks was named UI10/20, short for unambiguously illustrated 10/20. This refined and deterministic UI10/20 rule was used in this study as the theoretical basis for the virtual 10/20 measurement process.

Besides guiding optode placement, the navigation method can also be applied to MRI-free registration for fNIRS (Singh et al., 2005; Tsuzuki and Dan, 2014). In experiments with multi-channel fNIRS, registration is an important post hoc localization process to determine the observed cortical area of the given probe set. This process can be briefly described as follows. First, positions of 10/20 landmarks are identified on the head of the subject and digitized together with the channel locations, which are often simplified as the mid-point of optode dyads. Second, an affine transformation is estimated based on the digitized 10/20 landmarks, and used to transform the channels locations into the reference database established in MNI coordinates (Okamoto et al., 2004). Finally, in MNI coordinates, each transformed channel is projected to the cortex (Okamoto and Dan, 2005) of brains in the database, and projected points are integrated as a probabilistic estimation of the registration result. With this process, the observed anatomical region can be localized within a predictable error without requiring MR images of the subject (Singh et al., 2005; Tsuzuki and Dan, 2014). In this framework, 10/20 is essential for estimating the affine transformation from the head surface of the subject into MNI coordinates and thus the accuracy of 10/20 landmarks would affect the final accuracy of the MRI-free registration. In practice, 10/20 landmarks are commonly marked with manual measurement on the head of the subject and digitized with a 3D digitizer, which introduces problems in reliability. Thus, reliable 10/20 landmarks identified in virtual space with our proposed system, i.e., the output of S3R_VM, is expected to improve the performance of MRI-free registration of fNIRS. It is noteworthy that for registration, the registration algorithm can be directly executed on the output of S3R_VM in virtual space. Thus, the time cost of the navigation step is eliminated when the navigation method is applied to the registration process, as compared to probe placement.

In the real experiment, the 4 initial reference points are controlled to be the same for both manual measurement and the proposed navigation method since we wanted to focus on the variance induced in the measurement procedure. In practice, however, variability will also be added due to the manual digitization of the four reference points, especially for Iz, which usually cannot be clearly distinguished by visual inspection. Although this variance would not change our comparison conclusions since it affects both methods equally, this variance should be reduced as much as possible when applying the navigation method in practice. One way is to digitize each reference point multiple times and take the average of the positions. To solve the problem of finding Iz, a potential approach introduced in recent work, is to determine the location of Iz using the positions of Nz, AL, and AR, which are more likely to be precisely located on the scalp (Tsuzuki et al., 2016).

The ultimate goal of the proposed method is to facilitate the optode/coil placement on the scalp. In current vision, the target of navigation is limited within the 21 10/20 landmarks. This raises a limitation that at locations outside those of the 21 10/20 landmarks, experimenters are unable to precisely describe the location of the probe on the scalp or to directly infer the cortical area underneath. There are several possible directions to enrich the knowledge of such cranio-cerebral correspondence. One straightforward direction is to utilize higher-density versions of the UI10/20, such as UI10/10 and UI10/5 (Jurcak et al., 2007) which are extensions of the 21 landmark UI10/20, to 64 or 241 landmarks uniformly distributed on the head surface. However, it has been suggested that 241 landmarks might be the upper limit of resolution for a 10/20 derived system, considering the possible overlap between neighboring landmarks induced by inter-subject variance (Jurcak et al., 2007). For fNIRS, another direction is to combine the proposed system with virtual registration technology (Tsuzuki et al., 2007). Given the geometry of a probe holder, the spatial distribution of each channel can be simulated on a reconstructed head surface and then registered into MNI coordinates via 10/20 landmarks estimated by the proposed system. This combination could generalize the cranio-cerebral correspondence from landmarks of 10/20 and derived systems toward arbitrary positions on a head surface. Though facing several technical challenges, we believe this would be a promising solution to optode set placement for multichannel fNIRS.

In conclusion, the current study can be regard as a methodological advance in utilizing cranial-cortical correspondence established under the 10/20 framework for transcranial brain mapping techniques. As spatial resolution and specificity of transcranial technologies improve, methodologies for probe placement require more attention, particularly in terms of precision, but also in terms of time cost. In addition, to reduce the burden on the experimenter, user-friendly interfaces are also important. We believe our method successfully addresses these issues.

CZ, ZL, LD, and XX designed the research; XX, HZ, and WL developed the algorithm and system; XX, XY, and HZ performed the experiments, collected the data and analyzed the data; XX, XY, LD, and CZ drafted the work; CZ, ZL, LD, HZ, and WL revised the manuscript and contributed important intellectual content.

This work was supported by the National 973 Program (Grant 2014CB846100), the National Natural Science Foundation of China (Grants No. 61431002, 61273287 and 31221003), NCET (Grant 11-0046), and the Major Project of the National Social Science Foundation (Grant 12&ZD228).