2.1.

System Design

The schematic and a photograph of the assembled system are shown in Figs. 1(a) and 1(b). Five used iPhone XR smartphones (each purchased for approximately £180) were fixed to a hollow aluminum hoop of diameter $\sim 70\mathrm{cm}$ ($\sim 27.6\mathrm{in.}$). Four phones were positioned symmetrically around the hoop, with the fifth iPhone fixed by a flexible phone holder $\sim 35\mathrm{cm}$ above the plane of the hoop to obtain a scan of the top of the subject’s head. All five iPhone front cameras were oriented toward the center of the hoop. The 3D images of the target were acquired using the “Heges” 3D scanning app.¹² The Heges app utilized the front-facing structured illumination (“TrueDepth”) camera of the iPhone to directly capture depth-resolved images in a common point-cloud format.

Fig. 1

(a) Schematic and (b) the photograph of the proposed system. (c) The designed multichannel Bluetooth controller board.

To initiate and control the scanning process of the Heges 3D scanning app on multiple iPhones simultaneously, we developed a custom-designed printed circuit board (PCB) integrated with a multichannel Bluetooth controller and trigger (MDBT50, Raytac, China) [Fig. 1(c)]. The Bluetooth module on our designed PCB board is programmed to send a “volume-down” signal to each of the five iPhones. This signal is recognized by the Heges app as the user pressing a button on the side of the phone to begin image acquisition (functionality that exists in the iPhone and is reminiscent of the shutter release on a traditional camera). After triggering the acquisition, the user partially rotates the hoop 60 deg and then stops the acquisition using the Bluetooth controller. This rotation enhances the coverage and ensures a higher degree of overlap in the acquired point clouds. The increased overlap is essential for achieving accurate point cloud stitching during the subsequent processing. Note that the Heges app was selected from several options available because the developer of the app was willing to add functionality that pushed the resulting images directly from the iPhone to cloud storage without having to actively export the file. This is important to the functionality of our device as it means that the point clouds acquired with all five iPhones immediately and automatically uploaded to the cloud via Wi-Fi to permit instantaneous processing on a local laptop. A Python code running on the laptop downloads the point cloud data from each iPhone using their Internet Protocol addresses, allowing immediate processing to take place.

2.2.

Infant Registration Model and DOT System

Although the methodology proposed here is expected to work for any fNIRS system, it was tested in conjunction with two LUMO high-density DOT (HD-DOT) devices developed by Gowerlabs Ltd.¹³ In the LUMO device, individual hexagonal sensor modules, also referred to as tiles, contain three near-infrared light sources (at 735 and 850 nm) and four optical detectors. This design allows for measurements to be conducted both within each tile and across multiple tiles. By interconnecting multiple tiles through specially designed headgear, users can rapidly construct high-density fNIRS arrays, enabling 3D functional brain imaging. As shown in Figs. 2(a) and 2(b), two devices consisting of 33 and 18 hexagonal modules, respectively, were used to validate the coverage and accuracy of the proposed registration system. The light from the light sources on each LUMO tile is directed to the scalp via short plastic optical fibers called “light-guide” [Fig. 2(c)]. As in the previous applications of the LUMO system, each tile is adorned with a fluorescent green or pink triangular marker on its upper surface, with each corner of the triangle indicating the position directly above one of the tile’s light sources. The triangular marker exists in a flat plane and is precisely positioned, leveraging the fixed and known dimensions of the tile and light guides. Identifying the locations of the three corners of each triangle is sufficient to determine (without approximation) the location of all seven optical contacts to the scalp provided by that tile.

Fig. 2

(a) Photograph of the 33-tile and (b) 18-tile LUMO system. (c) The structure of the tile in the LUMO system. (d) The 3D-printed phantom for system validation and evaluation.

A 3D-printed head model based on the head of a 6-month-old infant was utilized for the validation of the proposed registration system [Fig. 2(d)]. The head model was designed to contain multiple indentations at precisely known positions acquired from the CAD model, into which the light-guides of the LUMO system are to be fitted. As a result, the 3D-printed model allows us to position a LUMO system such that the corners of each triangular sticker (relative to the landmarks of the model) are precisely known. This positioning information can therefore be treated as ground-truth data and used for comparison with the measurements obtained via different positioning methods.

2.3.

Point Cloud Alignment

As each point cloud was obtained in an arbitrary coordinate system, subsequent processing was required to align and “stitch” the point clouds into a complete 3D head model via appropriate spatial transformations. This alignment can either be performed manually or automatically. For manual alignment, all point clouds were imported into the CloudCompare software. The user manually selected at least three pairs of common points from two meshes, often choosing the vertices of the fluorescent green or pink triangular marker. The software then calculated a rigid transformation matrix that minimized the distance between the point pairs to achieve a global alignment between the meshes. To further enhance the alignment accuracy, an iterative closest point (ICP) algorithm was implemented using CloudCompare. This algorithm facilitates the improvement of general alignment between pairs of point clouds, refining the alignment results achieved through the initial rigid transformation process.

Our automatic alignment pipeline was performed using python with the Open3D¹⁴ and NumPy¹⁵ packages. The flowchart for our automatic alignment process is shown in Fig. 3. A density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm was first employed to filter out environmental noise. The DBSCAN algorithm groups points in the point cloud together to form clusters, while labeling outliers as noise. By setting appropriate parameters, this allowed us to effectively remove unwanted noise from the data, ensuring a cleaner point cloud cluster for more accurate alignment. A 3D head template obtained from a previous 3D scan of a different infant head model wearing the same DOT device was used for global alignment. As the spatial features (lines, planes, etc.) of the template phantom wearing the DOT device were similar to that of the target model, a fast point feature histograms (FPFH) vector describing the local geometric properties of each point from the point cloud can be calculated separately for both the template and the scans. A random sample consensus (RANSAC) algorithm was utilized to achieve the global alignment of the scans. In each RANSAC iteration, a user-specified number of points were selected from the scans, whereas the corresponding points in the template point cloud were detected by querying the nearest neighbor in the 33-dimensional FPFH feature space. The point pairs were then used to calculate the transformation matrix for the partial point cloud obtained from each scan. The RANSAC algorithm repeatedly estimates and refines the transformation matrix to find the best alignment that maximizes the number of inlier point pairs and minimizes the effect of outliers. Following the global alignment, the individual scans were color-filtered to keep the green and pink colors corresponding to the triangle markers. This step removed the black background and focused on the relevant features, which were also important spatial features utilized in manual alignment procedure. The color-filtering process improved the accuracy of the alignment. Subsequently, an ICP approach was used to refine the alignment further.

Fig. 3

Flowchart of the automatic alignment.

2.4.

Validation Study

2.4.2.

Spatial registration accuracy study on a moving phantom

In the preliminary evaluation of our registration system on dynamic scenarios involving a baby phantom, the 3D printed phantom wearing 18-tile LUMO device was utilized, as detailed in 2.4.1. Prior to scanning, the second experimenter was trained to replicate the movements of the head of a baby based on an open-source video of 6-month-old infants.¹⁶ Before conducting the scans, the experimenter’s capability to accurately reproduce these motions was verified by comparing the simulated movements with video clips simultaneously, whereas the second experimenter was blinded to the video. The phantom was held by the second experimenter while simulating the head movements of a 6-month-old infant as shown in Fig. 4. The proposed system obtained five acquisitions of the dynamic phantom. These acquisitions were subsequently processed and integrated into one whole-head 3D point cloud of the phantom, utilizing both manual and automatic alignment techniques.

Fig. 4

Photograph of the experimenter holding the phantom and mimicking baby head movement.

The 3D positions of cranial landmarks and the vertex points of triangle markers on the LUMO devices were extracted using CloudCompare. These positions were then compared with ground-truth data derived from the 3D printed CAD model. Following the evaluation paradigm outlined in 2.4.1, a rigid transformation was performed based on the cranial landmarks between the manually/automatically merged scan acquired by system and ground-truth data. The Euclidean distance error between the vertex points obtained by the proposed system with two different alignment methods and the true positions was calculated as the localization error.

Furthermore, to isolate the influence of the cranial landmark positioning error, another rigid transformation based on the vertices of the triangle markers on LUMO was implemented. The Euclidean distance error between the vertex points obtained by the proposed system with two different alignment methods and the ground-truth positions was calculated as the localization error after isolating the effect of cranial landmark positioning error.

3.1.

Localization Accuracy Study on the Static Phantom

The localization error of marker vertices and cranial landmarks was determined using the proposed registration system. The marker vertices indicate the position of the optodes on the DOT device, while cranial landmarks are distinctive points on the skull. The average Euclidean distance error between the position obtained using the registration system and the true value was calculated for each marker vertex or cranial landmark in five acquisitions. These marker vertices and cranial landmarks play a crucial role in spatial registration.

The median localization errors of each marker vertices acquired via the three different methods were calculated, as determined via cranial landmark-based registration (the approach most representative of a real-world application) are shown in Fig. 5(a). A box plot illustrating the distribution of localization errors across all marker vertices after the spatial registration process is shown in Fig. 5(b). The median localization errors of all marker vertices for the registration system with manual alignment (Manual) and automatic alignment (Auto), standard photogrammetry (Photogrammetry), and Polhemus were 9.3, 10.0, 8.7, and 9.0 mm, respectively. These errors are within the range typically observed in studies using similar spatial registration methods.

Fig. 5

(a) Localization errors of all marker vertices acquired via cranial landmark-based registration using each employed method including the registration system with manual alignment and with automatic alignment, standard photogrammetry, and Polhemus. Localization errors are represented with a color code on the phantom model. Each value indicates the Euclidean distance between the marker vertex on the original phantom and the node on the model acquired with each method. (b) Boxplots represent the distribution of the errors across all marker vertices after the spatial registration process, separately for each employed method. The box’s bottom and top edges represent the 25th and 75th percentiles, respectively, while the central mark represents the median. Outliers are illustrated as dots.

The median Euclidean distance error between the true positions and the cranial landmarks obtained by the proposed registration system with manual and automatic alignment, standard photogrammetry, and Polhemus were, respectively, 10.0, 11.5, 10.9, and 8.2 mm. Moreover, a rigid transformation matrix was calculated using the vertices of all the triangular markers. This approach eliminated the potential localization error caused by user’s inaccurate cranial landmark poisoning. After isolating the effect of cranial landmark positioning error, the median localization errors of all marker vertices for the registration system with manual alignment (Manual) and automatic alignment (Auto), standard photogrammetry (Photogrammetry), and Polhemus were 1.8, 3.0, 2.2, and 3.7, mm respectively. The localization errors across all marker vertices and a box plot illustrating the distribution of the errors after isolating the effect of cranial landmark positioning error are shown in Figs. 6(a) and 6(b).

Fig. 6

(a) Localization errors of all marker vertices using each employed method after isolating the effect of cranial landmark positioning error. Localization errors are represented with a color code on the phantom model. Each value indicates the Euclidean distance between the marker vertex on the original phantom and the node on the model acquired with each method. (b) Boxplots represent the distribution of the errors across all marker vertices after isolating the effect of cranial landmark positioning error, separately for each employed method. The box’s bottom and top edges represent the 25th and 75th percentiles, respectively, while the central mark represents the median. Outliers are illustrated as dots.

3.2.

Localization Accuracy Study on the Moving Phantom

The median localization errors of all marker vertices acquired via the system with two different alignment methods were calculated, as determined via cranial landmark-based registration. The localizations errors of the dynamic phantom were compared with the static phantom as shown in Fig. 7(a). A box plot illustrating the distribution of localization errors across all marker vertices after the spatial registration process is shown in Fig. 7(b). The median localization errors of all marker vertices for the registration system with manual alignment and automatic alignment under dynamic phantom situation were 9.5 and 9.3 mm, respectively. In contrast to the static phantom scenario, the movement of the phantom had a relatively minor impact on the localization errors. However, it did lead to an increased occurrence of outliers at the junctions between the scanned images from different smartphones. This is due to the more severe blurring at the edges of the scanned image caused by the phantom movement, as shown in Fig. 8.

Fig. 7

(a) Localization errors of all marker vertices acquired via cranial landmark-based registration using the proposed system, with two alignment methods under static and dynamic phantom. Localization errors are represented with a color code on the phantom model. Each value indicates the Euclidean distance between the marker vertex on the original phantom and the node on the model acquired with each method. (b) Boxplots represent the distribution of the errors across all marker vertices after the spatial registration process, separately for each situation. The box’s bottom and top edges represent the 25th and 75th percentiles, respectively, while the central mark represents the median. Outliers are illustrated as dots.

Fig. 8

Screenshot of the acquired 3D head model. The subfigure illustrates the severely blurred triangle vertices due to the phantom movement.

After isolating the influence of the cranial landmark positioning error, the median localization errors of all marker vertices acquired via the system with two different alignment methods were calculated. The localization errors of the dynamic phantom were compared with the static phantom as shown in Fig. 9(a). A box plot illustrating the distribution of localization errors across all marker vertices after the spatial registration process is shown in Fig. 9(b). The median localization errors of all marker vertices for the registration system with manual alignment and automatic alignment under dynamic phantom situation were 2.6 and 3.5 mm, respectively, after the cranial landmark positioning error isolation. The movement of the phantom indeed resulted in an increase of localization errors across all vertices of the triangle marker, particularly in the context of the manual alignment method. In addition, the phantom’s movement also contributed to an elevation in the occurrence of outliers at the junctions between scanned images obtained from different smartphones, which is also due to the significant blurring at the edges of the smartphone scans.

Fig. 9

(a) Localization errors of all marker vertices acquired using the proposed system after isolating cranial landmark positioning error, with two alignment methods under static and dynamic phantom. Localization errors are represented with a color code on the phantom model. Each value indicates the Euclidean distance between the marker vertex on the original phantom and the node on the model acquired with each method. (b) Boxplots represent the distribution of the errors across all marker vertices after cranial landmark positioning error isolation, separately for each situation (static/moving phantom). The box’s bottom and top edges represent the 25th and 75th percentiles, respectively, while the central mark represents the median. Outliers are illustrated as dots.

3.3.

Scan Coverage Study

The aligned whole-head 3D model of the phantom acquired from one of the scans using the proposed registration system was shown in Fig. 10. As the vertices are indicators of the position of the source and detector on the LUMO device, the number of vertices covered by the registration system relative to the total number of vertices on the LUMO device was calculated as the coverage rate, indicating its ability to cover the positions of the sources and detector as a registration system. No vertices were missing in any acquisitions, and some vertices at the margins of the scanned point cloud were blurred. An average of 0.6 vertices was blurred (3, 1, 1, 0, and 0, respectively), indicating that the proposed registration system was able to capture nearly all (99.4%) of the target’s relevant anatomical features.

Fig. 10

Screenshot of the acquired 3D head model. The subfigure illustrates the blurred triangle vertices. These blurred vertices usually appear at the margin of a single iPhone scan.