2.1. Physical support design

The device was designed around two main modules that enable its use with a smartphone as a digital camera: (i) a camera positioning control module and (ii) a rotation and photo capture module (Fig 1A). Both modules were integrated through an electronic circuit developed to control the system’s actuators, specifically servo and stepper motors. S1 Appendix presents the components used, along with their descriptions, for the construction of the physical support structure.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Design and electronic architecture of the AutoScan3D device.

(A) AutoScan3D device, composed of two modules for smartphone-based image capture. The rotation and photo capture module includes (1) an MG90 servo motor triggering the smartphone’s Bluetooth shutter and a 28BYJ-48 stepper motor with an M4 nut, connected to (2) a telescopic rod repurposed from a selfie stick, enabling controlled object rotation. The assembly is coated in black paint except for the 5/16“ threaded rod, cables, and telescopic rod. The camera positioning control module consists of (3) a circuit housing with openings for the Nema 17 motor shaft (14 mm) and cable routing. Vertical camera movement is managed by (4) a “std h-35 400 mm” telescopic slide rail with linear bearings, secured with MDF joints and fasteners. Elevation control is achieved via a 5/16” threaded rod, coupled to the Nema 17 motor through a modified expansion plug and guided by 5/16” nuts. (5) An MG995 servo motor enhances stability and precise positioning. (B) Electronic circuit diagram illustrating the interconnection of key components, including (1) Nema 17 stepper motor, (2) 100 µF capacitor, (3) MG90 servo motor, (4) MG995 servo motor, (5) Arduino UNO board, (6) 5.5 × 2.1 mm female jack for power supply, (7) Pololu A4988 driver, (8) ULN2003 driver, and (9) 28BYJ-48 stepper motor. The electronic circuit schematic (1B) was designed using KiCAD software (version 8.0). This system facilitates high-precision, multi-angle image capture for anatomical and morphological studies.

https://doi.org/10.1371/journal.pone.0336996.g001

2.2. Program design in Arduino IDE

The program was developed using Arduino IDE 2.3.3. Initially, existing Arduino-based projects and example codes related to motor control were reviewed. Each motor was individually connected to the Arduino UNO to develop and adapt its operational code. Once all necessary information regarding wiring and programming was compiled and tested, the final wiring diagram and an integrated code were designed to encompass all required functions (S2 Appendix). The Arduino IDE interface facilitated structured programming using the C programming language, divided into three main sections: (i) variable declaration for defining operational parameters, (ii) setup section for initializing hardware components, and (iii) loop section containing repetitive instructions for continuous system operation. The 28BYJ-48 stepper motor was connected to the Arduino UNO and ULN2003 driver using a full-size 830-point MB-102 protoboard. During this phase, code tests were conducted to control the motor’s rotation angle, refining parameters for precise operation. The Nema 17 stepper motor was subsequently connected to the Pololu A4988 driveron the same protoboard, where tests established the relationship between the number of motor rotations and the vertical displacement of the rail system. Additionally, the MG90 and MG995 servo motors were integrated into the protoboard, where their rotation direction and angles were adjusted through code testing. These tests identified key programming lines essential for ensuring precise positioning and coordinated motor movement. After finalizing the programming and verifying all configurations, a fully integrated code was developed to synchronize the control of all motors. This final version was uploaded to the Arduino UNO, enabling full system functionality.

2.3. 3D model digitization with AutoScan3D

A workflow (S3 Appendix) was designed to obtain 3D digitizations of an adult skull of the Peruvian booby (Sula variegata, 270 mm height × 60 mm diameter; Tschudi, 1843). The specimen was complete, with an immobile mandible and without a keratinized covering. To assess the performance of the system across diverse anatomical structures, we also included the digitization of a vertebra of the common carp (Cyprinus carpio) and a cervical vertebra of the southern pudu (Pudu puda). All digitizations were performed using the AutoScan3D photogrammetry device and compared with high-resolution micro-CT scans obtained using a Bruker SkyScan 1278 system (version 1.0.5), to validate the accuracy and utility of the proposed device.

2.3.1. Photographic environment setup.

A setup was installed, consisting of a 2 × 1.6 m black fabric backdrop, a flexible-neck desk lamp equipped with a 15W LED bulb (E27), and a foldable 110 cm translucent diffuser. The light source was positioned at a 45° angle relative to both the camera and the object, approximately 1.5 m away from the skull, with the diffuser placed 50 cm from the bulb to enhance light distribution. The black backdrop was arranged parallel to a window in the museum room, allowing for the entry of indirect natural light (Fig 2A). The skull was mounted by inserting a piece of cork into the foramen magnum, serving as a base to secure a bamboo rod. This rod was affixed with transparent adhesive tape to the tip of the telescopic cylindrical bar of the device, enabling the vertical positioning of the skull with the beak oriented upwards. For image capture, a Samsung Galaxy A15 5G (width: 76.8 mm, main camera: 50 MP, f/1.8 aperture) was used, configured to capture images at 12 MP resolution (to optimize file size and processing efficiency in Agisoft Metashape) with auto-exposure/auto-focus lock to maintain consistent intrinsic camera properties (e.g., focal length, exposure) during photogrammetry, mounted on the vertical rail (Fig 2C). The iPhone 8 (width: 67.3 mm, camera: 12 MP, f/1.8 aperture) was similarly configured with auto-exposure/auto-focus lock, ensuring compatibility with the phone holder (maximum opening: 80 mm). The auto-exposure/auto-focus lock feature, available in the native camera apps of both smartphones, ensured stable image capture parameters, critical for accurate alignment in Agisoft Metashape [7]. A detailed protocol for constructing and operating AutoScan3D, including camera settings, is provided in Appendix 3 to ensure replicability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Operational configuration and imaging workflow of the AutoScan3D system.

(A) A smartphone is mounted on the vertical rail, capturing images of a Sula variegata skull (Tschudi, 1843). (B) Completed physical support design, displaying the final device and its two modules: the camera positioning control module and the rotation and photo capture module. Frontal view of the camera positioning control module. Frontal view of the rotation and photo capture module. (C) The smartphone holder features a maximum opening of 80 mm, accommodating devices with a width of 80 mm or less. (D) Base of the rotation and photo capture module, housing an MG90 servo motor with a plastic arm that functions as a trigger, activating the Bluetooth shutter button of the smartphone camera.

https://doi.org/10.1371/journal.pone.0336996.g002

2.4. 3D model digitization with micro-CT

The digitization process using micro-CT was conducted with a Bruker SkyScan 1278 system, version 1.0.5. Three-dimensional (3D) images were generated from datasets obtained from the skull bones, employing semi-automatic segmentation to create 3D models. The acquisition parameters included a source voltage of 59 kV, a source current of 692 µA, and a voxel size of 51.489 µm. The total exposure time during the scanning process was approximately 1 hour, 34 minutes, and 50 seconds. The processing and analysis of the images were carried out using the Slicer software, version 5.6.2 [8], with the Slicer Morph extension [9]. The TIFF files generated during digitization were converted into NRRD format, facilitating the 3D model reconstruction. Finally, the reconstructed model was imported into the Blender 4.2 3D modeling software, where the vertex count of the model was verified.

Prior to deviation analysis, the 3D models obtained with AutoScan3D were spatially aligned to the corresponding micro-CT reference models to ensure geometric comparability. A manual pre-alignment was first performed using three easily identifiable anatomical landmarks (e.g., rostral tip, occipital condyle, and foramen magnum margin) to achieve an approximate initial positioning. This was followed by a rigid-body registration using the Iterative Closest Point (ICP) algorithm implemented in GOM Inspect 2023 (GOM GmbH, Germany). The algorithm iteratively minimized the mean square distance between homologous vertices of both meshes until convergence was reached. The quality of alignmentwas confirmed by verifying that the root mean square (RMS) residual error between the reference (v_ref) and test (v_test) vertices was consistently below 0.2 mm. After alignment, deviation maps and statistical reports were generated to quantify mean, standard, and maximum Euclidean distances (‖v_ref – v_test‖) between models.

2.5. Validation

The validation of AutoScan3D revealed morphological disparities between photogrammetry-generated and micro-CT models, particularly in the ventral and caudal regions of the Sula variegata skull. To address concerns about low resolution and insufficient anatomical detail (e.g., fused skeletal elements, obliterated sutures), we optimized the Agisoft Metashape workflow and conducted tests with three objects: a Sula variegate skull (270 mm height × 60 mm diameter), a pudu (Pudu puda) vertebra (~30 mm height × 20 mm diameter), and a common carp (Cyprinus carpio) vertebra (~15 mm height × 10 mm diameter). Each object was scanned with 120 images (six vertical positions, 20 photographs at 18° intervals) using a Samsung Galaxy A15 5G (f/1.8 aperture) and iPhone 8 (12 MP, f/1.8, 67.3 mm). The optimized workflow, using “Model from Depth Maps” and refined camera alignment, yielded models with enhanced resolution: the skull model had 198,376 vertices, 38.2 MB, and a mean deviation of ±1.9 mm (180% dispersion coefficient); the pudu vertebra had 182,150 vertices, 35.8 MB, and ±1.7 mm; the carp vertebra had 165,820 vertices, 32.4 MB, and ±2.0 mm. These models showed ~35% improved clarity in fine structures (e.g., nasofrontal suture, vertebral laminae) compared to the initial workflow (132,952 vertices, ± 2.7 mm), reducing fusion artifacts. The Samsung Galaxy A15 5G provided slightly sharper textures (12 MP vs. 48 MP), but software optimization, not the smartphone, was the primary factor improving quality. Model scaling was performed in GOM Inspect by aligning AutoScan3D models to micro-CT reference models using a best-fit algorithm, ensuring consistent dimensions (e.g., skull length: 270 mm). Alignment used iterative closest point (ICP) registration, achieving sub-millimeter precision. While micro-CT (5,938,596 vertices, 926 MB) remains superior for internal structures, AutoScan3D’s optimized models are suitable for museographic, educational, and forensic applications. To quantify the geometric accuracy of the 3D models generated with AutoScan3D, the reconstructed meshes were compared against the corresponding micro-CT reference models using GOM Inspect 2023 (GOM GmbH, Germany). A Best-Fit Alignment was performed using the Iterative Closest Point (ICP) algorithm after a preliminary manual registration based on three anatomical landmarks to ensure consistent orientation. Once aligned, the Deviation Analysismodule was applied to compute the Euclidean distance between homologous vertices of both meshes, defined as the scalar magnitude of the displacement vector (‖v_ref – v_test‖). The mean deviation (mm) was calculated as the arithmetic mean of these distances, while the standard deviation (SD) represented their variability. The maximum deviation (mm) corresponded to the largest scalar distance observed between the two surfaces. Each deviation value was expressed as a scalar magnitude, allowing the analysis of geometric error independently of directionality. All results were exported as deviation maps and numerical reports for subsequent statistical comparison.

2.6. Surface analysis in GOM inspect

Surface analysis was performed using GOM Inspect [3,6]. This software enabled the superimposition of the 3D models generated with AutoScan3D and micro-CT, using the micro-CT model as the reference due to its superior precision compared to photogrammetry [3]. Through this approach, quantitative differences in the position of vertices or points between both models were evaluated. The results were graphically represented in a color-coded model, highlighting areas with variations and providing a clear visualization of surface discrepancies. This method allowed for the identification and quantification of differences in terms of position and resolution. Additionally, the vertex count observed in Blender 4.2 was recorded as an additional measure of model resolution, complementing the data obtained through GOM Inspect. This analysis provided an objective and detailed validation of AutoScan3D’s capabilities in comparison to micro-CT.

3.1. AutoScan3D design

3.1.3. System operation.

The image capture speed was one photograph every three seconds, and the total process duration—including setup and automated acquisition—was approximately 30 minutes per specimen. The photographic setup included a matte black background and a fixed LED lamp to ensure uniform illumination and contrast (Fig 2). Background removal was performed using the CANVA platform, which facilitated the alignment of photographs in the Photo Alignment step in Agisoft Metashape. Each 3D model was generated from 108 photographs (three sets of 36 images captured at 10° intervals and different vertical camera angles), using the AutoScan3D automated platform. Image processing in Agisoft Metashape, using the “Model from Depth Maps” workflow with refined camera alignment, required approximately 25 minutes. Thus, the complete digitization process for each object—including scanning and model reconstruction—took approximately 55 minutes. The final Sula variegata skull model consisted of 198,376 vertices, had a file size of 38.2 MB, and presented realistic photographic texture but no internal anatomical structures. In comparison, the micro-CT model was monochromatic, lacked texture, but included highly detailed internal anatomy, with 5,938,596 vertices and a file size of 926 MB, as measured in Blender 4.2 (Fig 3). In addition to the skull, two vertebral elements were digitized to evaluate system performance on smaller and more complex anatomical units: a cervical vertebra of the southern pudu (Pudu puda) and a vertebra of the common carp (Cyprinus carpio). The Pudu vertebra model consisted of 1,730,180 vertices and a file size of 165 MB (Samsung scan), while the Carp vertebra model had 152,331 vertices and a size of 14.5 MB. Models generated with the iPhone 8 exhibited higher vertex counts (1265451 for Pudu, 1249810 for Carp) and slightly improved detail, although software optimization remained the principal determinant of quality. All models were scaled and aligned to micro-CT references using GOM Inspect and the iterative closest point (ICP) algorithm, achieving sub-millimeter accuracy. These results demonstrate the system’s capacity to produce detailed surface models suitable for museographic, educational, and comparative anatomical applications.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. 3D model of a (A) skull of Sula variegata skull, (B) a cervical vertebra of the southern Pudu (Pudu puda) and (C) a vertebra of the Common carp (Cyprinus carpio) generated with AutoScan3D.

The photographic texture provides a realistic appearance to the model, along with details such as visible markings made with a graphite pencil.

https://doi.org/10.1371/journal.pone.0336996.g003

3.2. Validation of AutoScan3D

The validation of AutoScan3D revealed morphological disparities between photogrammetry-generated and micro-CT models, particularly in the ventral and caudal regions of the Sula variegata skull. To address concerns about low resolution and insufficient anatomical detail (e.g., fused skeletal elements, obliterated sutures), we optimized the Agisoft Metashape workflow and conducted tests with three specimens: a Sula variegata skull (270 mm height × 60 mm diameter), a Pudu (Pudu puda) vertebra (~30 mm height × 20 mm diameter), and a Common carp (Cyprinus carpio) vertebra (~15 mm height × 10 mm diameter). Each object was scanned using 120 images (six vertical positions, 20 photographs at 18° intervals) with both a Samsung Galaxy A15 5G (50 MP camera, f/1.8 aperture) and an iPhone 8 (12 MP, f/1.8, 67.3 mm focal length). The optimized workflow, which included the use of “Model from Depth Maps” and refined camera alignment, resulted in enhanced 3D reconstructions. For example, the Samsung scans yielded models with the following characteristics: Sula skull: 198,376 vertices (38.2 MB, mean deviation ±1.9 mm, 180% dispersion coefficient, Pudu vertebra: 173,0180 vertices (165 MB, mean deviation ±0.327 mm), Carp vertebra: 152,331 vertices (14.5 MB, mean deviation ±0.089 mm). In contrast, the iPhone 8 reconstructions showed: Pudu vertebra: 1265451 vertices (120 MB, mean deviation ±0.038 mm) and Carp vertebra: 1249810 vertices (119 MB, mean deviation ±0.213 mm). These results indicate a significant improvement over initial models (132952 vertices, ± 2.7 mm), with up to ~35% increased clarity in fine anatomical features (e.g., nasofrontal suture, vertebral laminae) and reduced fusion artifacts. While the Samsung device offered slightly sharper textures, the critical factor for enhanced resolution was the optimization of the photogrammetric pipeline, not the device itself. Model scaling and alignment were performed in GOM Inspect using a best-fit algorithm with iterative closest point (ICP) registration to micro-CT reference models (5,938,596 vertices, 926 MB), achieving sub-millimeter accuracy. Although micro-CT remains superior for internal and highly detailed structures, the optimized AutoScan3D models are suitable for museographic display, educational applications, and some forensic contexts. Complete validation metrics are summarized in Table 1 and Fig 4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Quantitative comparison of 3D models generated with AutoScan3D and reference datasets.

https://doi.org/10.1371/journal.pone.0336996.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Surface deviation analysis between AutoScan3D and micro-CT 3D models for three osteological specimens.

Surface comparison maps generated in GOM Inspect show the spatial deviation between AutoScan3D-generated 3D models and reference micro-CT models for three specimens: a skull of Sula variegata, a cervical vertebra of the southern pudu (Pudu puda), and a vertebra of the common carp (Cyprinus carpio). Models were created using two different smartphones (Samsung Galaxy A15 5G and iPhone 8), and deviations are visually represented using a color scale ranging from red (positive deviation, protrusion in the AutoScan3D model) to blue (negative deviation). Green areas indicate optimal alignment, with differences within ±1 mm. For each model, 88 homologous surface points were analyzed to quantify geometric discrepancies using best-fit alignment and Iterative Closest Point (ICP) registration. These points were used to compute key accuracy metrics such as mean distance, standard deviation, and extreme deviations (maximum positive and negative values), as summarized in Table 1. The color bar on the right of each model indicates the magnitude of the deviation in millimeters. Labels on the models highlight representative localized differences at specific anatomical landmarks.

https://doi.org/10.1371/journal.pone.0336996.g004

3.3. Comparison of 3D models

The comparative analysis of three-dimensional models generated via AutoScan3D and micro-CT revealed substantial morphological variations across multiple anatomical perspectives of the cranium and vertebrae. The systematic examination demonstrated technology-specific capabilities and constraints in terms of resolution, morphological detail capture, and modeling accuracy. Analysis of the cranial and caudal aspects of the Sula variegata skull revealed four primary distinctions between the imaging methodologies (Fig 5). The micro-CT-generated model exhibited enhanced resolution of quadrate bone articular surfaces within the mandible from the caudal perspective, while the cranial view demonstrated superior visualization of internal osseous structures, particularly the premaxillary trabecular architecture. In contrast, the AutoScan3D model displayed palatine bone hypertrophy and fusion of the parasphenoid medial processes with the quadrate and mandibular elements. Lateral view comparison identified five key morphological disparities. The AutoScan3D model exhibited complete fusion of the nasofrontal suture, absence of primary nasal aperture fusion sulci, and imperceptible intraramal suture definition. Furthermore, the interorbital septum demonstrated incomplete formation, with quadratojugal-mandibular fusion evident. Examination of the inferior aspect revealed three significant variations in the AutoScan3D model: fusion of mandibular dentary elements, reduced definition and hypertrophy of the palatine bone and its maxillary process, and notable thickening of the pterygoid and basisphenoid structures. The dorsal perspective of the AutoScan3D model demonstrated absence of nasofrontal sutural detail and exhibited characteristic granular surface morphology. Comparable disparities were also observed in the vertebral models of Pudu puda and Cyprinus carpio. In the pudú cervical vertebra, the micro-CT model allowed precise delineation of the vertebral laminae, articular facets, and internal trabecular structure, whereas the AutoScan3D model exhibited smoothed contours, reduced definition of articular surfaces, and partial loss of foraminal boundaries. Similarly, in the carp vertebra, micro-CT rendered detailed internal architecture and subtle concavities along the centrum, which were either simplified or absent in the photogrammetric reconstructions. Both AutoScan3D models presented evidence of surface fusion artifacts and exaggerated thickness in bony prominences, particularly along the transverse processes. These observations were systematically documented in S1 video.mp4, providing visual documentation of the comparative morphological analyses between the two imaging modalities.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Comparison of 3D skull models of Sula variegata generated by AutoScan3D and micro-CT.

The images highlight anatomical and morphological differences between the models produced by both technologies in five views: (A, A’) Anterior view: AutoScan3D exhibits fewer internal details compared to micro-CT, which reveals bony trabeculae in the premaxilla. (B, B’) Posterior view: micro-CT provides greater detail in the quadrate bone articulations, whereas AutoScan3D shows thickening of the palatine bone and absence of the medial parasphenoid processes. (C, C’) Lateral view: AutoScan3D reveals bone fusions, such as the naso-frontal suture and the quadratojugal bone fused with the mandible, in addition to an incomplete interorbital septum. (D, D’) Superior view: micro-CT provides greater clarity in the naso-frontal suture, whereas AutoScan3D displays a grainy surface texture. (E, E’) Inferior view: AutoScan3D shows fusion of the dentary bones, thickening of the palatine bone and the basisphenoid, with less overall detail compared to micro-CT. These differences reflect the limitations of AutoScan3D in resolution and internal modeling, as well as the strengths of micro-CT in capturing internal details and anatomical precision.

https://doi.org/10.1371/journal.pone.0336996.g005

4.1. MorphoScan design and technical implementation

A key advantage of AutoScan3D is its cost-effectiveness, with a total construction cost of approximately USD 90 for hardware components, including an Arduino UNO, stepper motors, and a repurposed selfie stick. However, the photogrammetry workflow relies on Agisoft Metashape Professional, with a one-time license cost of USD 3,499 for the full commercial version, which increases the total investment when included. For academic and educational users, a restricted educational license is available for USD 549, applicable to accredited institutions, their employees, and students for non-commercial use (Agisoft LLC, 2024; https://www.agisoft.com/buy/licensing-options/). This option enhances accessibility for resource-constrained researchers. While micro-CT systems involve a substantially higher institutional investment—typically between USD 260,000 and 750,000 for acquisition—the relevant metric for users is the operational scanning cost. In practice, the operational rate at facilities such as the Bruker SkyScan 1278 averages around USD 100 per hour (USD 50–150 per scan, depending on duration and voxel size). Thus, micro-CT scanning can be cost-effective per use when the infrastructure already exists, but access remains centralized in specialized facilities. AutoScan3D, by contrast, eliminates the need for such infrastructure, offering negligible operating costs once assembled and enabling independent use in laboratories, classrooms, and museums. This operational autonomy, rather than equipment cost alone, represents the main advantage of the system. To ensure a comprehensive cost comparison, AutoScan3D’s total initial cost, including the full Agisoft Metashape license, is approximately USD 3,589 (or USD 639 with the educational license)—still substantially lower than the acquisition cost of micro-CT but higher than its per-scan cost.

Notably, AutoScan3D’s hardware cost (USD 90) is over 30 times lower than Medina et al.’s (2020) setup, and open-source alternatives like Meshroom could further reduce software costs, as suggested by [7]. Recent comparative evaluations between Agisoft Metashape and Meshroom (AliceVision framework) have shown that the open-source software achieves equivalent geometric accuracy and surface reconstruction quality for small to medium-sized objects, with mean deviations typically below 2–3 mm and similar mesh completeness. Although Meshroom requires longer processing times and offers fewer optimization controls, it remains a reliable option for educational and low-budget photogrammetry applications, reinforcing the feasibility of AutoScan3D in resource-limited settings. The use of a smartphone instead of a professional camera, while limiting resolution, contributes to this cost advantage, aligning with the system’s goal of accessibility for institutions with limited resources. Comparative analysis with [1] demonstrates that AutoScan3D’s automated design achieves comparable geometric accuracy (mean deviation of ±1.9 mm vs. ± 1.5 mm, Table 1) at a lower hardware cost and with reduced operator dependency due to automation. This economic advantage, together with the system’s reproducibility and ease of operation, supports AutoScan3D as a complementary, low-cost photogrammetric tool suitable for educational and museographic digitization, particularly in low-budget settings.

4.2. Comparative analysis and validation

Fundamental operational differences between micro-CT and photogrammetric methodologies determine their respective imaging capabilities. Micro-CT, based on X-ray imaging, provides volumetric data with sub-millimetric resolution and a much higher point density than AutoScan3D (Table 1), enabling the visualization of internal and external structures with great detail [13,14]. In contrast, photogrammetry is restricted to the object’s surface but provides true-color textures that can be valuable for visual and educational purposes. The larger file size of micro-CT datasets reflects their greater volumetric richness and surface fidelity rather than inefficiency. When needed, these volumetric datasets can be converted into lighter surface meshes or decimated models without significant loss of geometric accuracy, depending on the analytical objective. These distinctions agree with prior reports [6], which emphasize micro-CT’s superior spatial resolution (voxel sizes ~10–50 µm) and photogrammetry’s advantages in cost and portability. The geometric deviation metrics used in this study were calculated as scalar Euclidean distances (‖v_ref – v_test‖) between homologous vertices after rigid-body alignment of the AutoScan3D meshes to the micro-CT reference models. Alignment was performed in GOM Inspect 2023 using the Iterative Closest Point (ICP) algorithm following a manual pre-alignment based on anatomical landmarks. The mean deviation represents the arithmetic mean of these vertex-to-vertex distances, the standard deviation (SD) reflects their variability, and the maximum deviation corresponds to the largest scalar distance observed. Because all deviations were expressed as magnitudes, these metrics quantify geometric error independently of vector direction, providing a reproducible and objective measure of model fidelity across acquisition systems. AutoScan3D’s automated system—combining a smartphone camera (≤ 80 mm width) with Arduino-controlled stepper motors—introduces specific trade-offs that help explain the observed performance differences. The smartphone sensor’s limited dynamic range and depth-of-field, together with the single-light illumination setup, contribute to minor deviations in complex regions such as the ventral skull surface (mean deviation ± 2.7 mm, Table 1). These design choices distinguish AutoScan3D from professional photogrammetric configurations, which typically use DSLR cameras and multi-angle lighting to improve feature detection [15].

Despite these constraints, AutoScan3D’s automated rotation and capture modules improve image consistency and reduce operator-dependent variability relative to manual photogrammetry [15]. The smaller file sizes (27.9 MB) and realistic photographic textures facilitate easy manipulation in software such as Blender 4.2, making the models suitable for educational visualization and specimen documentation where surface detail is the primary requirement. Its geometric performance falls within the range reported for other photogrammetric studies (average error 1–5% [1,7]), though it remains below the spatial resolution achievable with micro-CT. The micro-CT data were used solely as a high-resolution reference to validate AutoScan3D’s geometric consistency, not as a competing imaging modality. While we acknowledge that low-cost hand-held laser scanners may represent a more direct alternative for surface-based digitization, their inclusion was beyond the scope of this study. The main objective was to validate AutoScan3D’s automation performance and geometric reliability, rather than to benchmark all available scanning technologies. The micro-CT dataset was selected because it provides a consistent and high-precision volumetric reference, allowing for objective quantification of surface deviation under controlled conditions. Moreover, entry-level laser scanners (typically USD 400–600) often require proprietary calibration targets, specific lighting environments, and licensed software, which would have contradicted the low-cost and open-source design principles that guided AutoScan3D’s development. In addition, low-texture bone surfaces are known to pose reconstruction challenges for both photogrammetry and consumer-grade laser scanners, as previously reported in comparative imaging studies. Future work may include testing AutoScan3D against structured-light or laser-based systems under standardized acquisition protocols, once the current prototype’s mechanical and optical parameters have been fully optimized. The most relevant methodological comparison is with other photogrammetric approaches, particularly that of Medina et al. (2020) [1], which shares similar acquisition principles. In this context, AutoScan3D achieved comparable surface accuracy (± 1.9–2.7 mm) while reducing hardware costs by more than an order of magnitude. It is important to note that the specimens analyzed by Medina et al. (2020) included avian material with plumage and soft tissues, which differ substantially in surface texture and reflectance from the skeletal specimens used in this study. Consequently, this comparison was not intended to establish direct geometric equivalence but rather to contextualize AutoScan3D’s relative efficiency and order-of-magnitude accuracy among photogrammetric approaches operating under different material and optical conditions. Nevertheless, these models remain dependent on lighting, camera quality, and limited viewing angles, and they cannot capture internal morphology; therefore, their application should focus on external-surface analyses and teaching contexts. Future improvements—such as integrating higher-resolution cameras, multi-angle capture, or improved lighting—could further enhance accuracy [16]. While micro-CT’s internal imaging capabilities are technologically unmatched [13], AutoScan3D’s design emphasizes affordability, reproducibility, and portability. This makes it a complementary tool for educational and museographic digitization, particularly in settings with limited access to advanced imaging infrastructure. To further contextualize its performance, AutoScan3D was compared with other photogrammetric and scanning approaches. The system completed full digitization in approximately 90 minutes (45 minutes for image capture and 45 minutes for processing in Metashape), substantially reducing operator time compared with manual photogrammetry pipelines [1]. This efficiency results from its Arduino-controlled automation, which ensures consistent image acquisition without manual repositioning.

In terms of model detail, AutoScan3D produced meshes with 132,952 vertices (Table 1)—adequate for educational and display purposes but less precise than commercial photogrammetric or structured-light scanners, such as Artec Spider (0.05–0.1 mm accuracy [7]). However, those systems cost approximately USD 20,000 and require controlled environments and trained operators. When compared to other low-cost automated prototypes [15], AutoScan3D offers similar accuracy (errors ~1–3 mm) and greater operational simplicity due to smartphone-based acquisition. Its realistic textures also reduce post-processing demands relative to manual protocols [1]. Ease of use remains a key advantage. Although operation requires basic Arduino programming knowledge, automation minimizes the need for precise manual camera handling, thereby improving reproducibility. Compared with subscription-based platforms such as RealityCapture [1], AutoScan3D combined with Agisoft Metashape—or open-source alternatives like Meshroom—provides a flexible and affordable processing workflow. Nonetheless, the current single-axis configuration limits image coverage for complex shapes; future iterations could incorporate multi-axis motion to enhance geometric completeness [16]. Overall, AutoScan3D demonstrates a balanced compromise among cost, speed, and reproducibility. It improves operational efficiency compared with manual photogrammetry and offers simplicity comparable to other low-cost automated systems. Rather than replacing high-end scanners, AutoScan3D provides an accessible framework for teaching, training, and specimen documentation where high precision and volumetric data are not essential.

4.3. Limitations

AutoScan3D exhibited several inherent limitations due to its current design. The camera’s vertical-only movement may limit accuracy for objects with complex, curved geometries, such as highly domed skulls, due to restricted angular coverage. Future designs could incorporate a multi-axis positioning module to enhance capture flexibility [16]. Its reliance on optimal lighting conditions without an internal light regulation system makes its use challenging in unfavorable environments, potentially affecting the quality of the captured images. Additionally, its inability to record internal structures restricts its application to surface-based three-dimensional models, distinguishing it from technologies such as micro-CT. The system operates exclusively with mobile phones, limiting its versatility compared to cameras with superior technical capabilities. Motor control depends directly on modifications to Arduino code, requiring users to have basic programming knowledge. This lack of a graphical user interface (GUI) for dynamic adjustments presents a learning curve for users with no prior coding experience. Moreover, the vertical movement range is constrained by the dimensions of the telescopic rail, while the distance between the object and the camera must be manually adjusted, introducing potential alignment, and focusing errors. Another key limitation is that during operation, the phone screen remains covered, preventing real-time monitoring of the field of view, which may compromise precision in photographic capture. To mitigate these challenges, strict protocols for lighting and alignment were implemented, along with detailed user guides (S4 Appendix) to assist those with less technical expertise. Future improvements will focus on automated adjustments and a more intuitive interface, aiming to enhance user experience and minimize errors in image capture and processing.

4.4. Projections

The development of AutoScan3D should prioritize the integration of electronic enhancements to improve its functionality and usability. Implementing a user interface with screens and control buttons would allow direct motor adjustments based on object characteristics, eliminating the need for manual modifications in the Arduino code. A structural redesign to support professional cameras would enhance its load capacity and stability, broadening its applicability to contexts requiring higher precision. Additionally, the creation of a secondary interface, either as dedicated software or a mobile application, would facilitate real-time capture control and image exportation, significantly improving operational efficiency. To optimize three-dimensional model quality, an alternative approach could involve capturing photogrammetric data from individual bones, allowing for more detailed reconstructions. These individual models could then be aligned and merged using editing software, resulting in a more accurate and comprehensive final representation. Such improvements would address AutoScan3D’s current limitations and bring its performance closer to that of advanced imaging technologies like micro-CT in applications where internal modeling is not required. Beyond anatomical and museum applications, AutoScan3D holds significant potential in forensic investigations [17], providing an accessible and efficient method for three-dimensional skull reconstruction. Its ability to generate high-resolution digital models with realistic texture makes it particularly valuable for facial approximation, forensic anthropology, and skeletal analysis in human identification cases. Unlike micro-CT, which is costly and requires specialized facilities, AutoScan3D offers a portable, low-cost solution, facilitating its use in fieldwork, crime scene documentation, and mass disaster victim identification. The automated capture system ensures standardized imaging, reducing human error and enhancing reproducibility in forensic reconstructions. Additionally, its compatibility with widely used photogrammetry software enables forensic experts to integrate 3D models into legal investigations, courtroom presentations, and comparative analyses with existing forensic databases. Future improvements, including expanded camera compatibility and real-time monitoring, could further strengthen its forensic applications, establishing AutoScan3D as an invaluable tool in criminal investigations and forensic research.