Evaluation of two stereophotogrametry software for 3D reconstruction of virtual facial models

ABSTRACT Objective: The present study aimed to evaluate the accuracy of 3D facial soft tissue virtual models produced by two photogrammetry softwares (AgiSoft Photoscan and 3DF Zephyr Free), when compared to those created by cone beam computed tomography (CBCT). Methods: Ten patients were submitted to two sequences of photographs performed with a DSLR camera (with and without the aid of a ring flash) and CBCT scans. Each photo series for each patient was processed with the softwares, and at the end, five models of each patient were generated: 1) CBCT, 2) AAL (Agisoft Ambient Light), 3) AFL (Agisoft Flash Light), 4) ZAL (Zephyr Ambient Light), and 5) ZFL (Zephyr Flash Light). Color coded maps and root-mean-square (RMS) distances were used to compare the photogrammetry models to the CBCT ones. Results: One sample t-test showed significant differences between all methods versus CBCT. The worst results were seen in the ZAL group (discrepancies up to 5.17mm), while the best results were produced by AAL group (discrepancies up to 2.11mm). Conclusions: It can be concluded that this type of virtual facial models are reasonably accurate, although not perfect, and considering its lower biological and financial cost, they may play an important role in specific situations.


INTRODUCTION
Adequate planning is key for orthodontic success, prognosis definition, as well as for treatment duration prediction.
Comprehensive planning demands diagnostic techniques to identify aesthetical, functional and anatomical aspects. However, a divergence in the diagnosis of orthodontic problems among professionals is still very frequent. 1 Facial analysis is a valuable tool for quantitative and qualitative assessment, and its application is increasingly demanded by professionals and patients. 2,3 At the beginning of the 20th century, Angle stated that the orthodontist would be able to classify malocclusion only by facial evaluation. 4 But it was only in 1999 that the importance of soft tissue assessment in diagnosis and treatment planning was highlighted, and its use was advocated, recognizing the face as a determining factor in orthodontic planning. Currently, facial aesthetics is being increasingly considered in the decision making process of teeth extractions, sagittal and asymmetry camouflage, as well as orthognathic surgery. [6][7][8][9][10] Most popular facial analysis methods include photographs and cephalometric measurements based on two-dimensional (2D) images. These have inherent limitations, such as a significant amount of radiographic projection error, distortion, inaccurate duplication of measurements, significant variation in the position of reference points, among others. 11,12 After the introduction of three-dimensional (3D) technologies in dentistry, 3D imaging systems are increasingly being used instead of 2D ones, especially in cases of craniofacial deformities. In fact, three-dimensional image provides more detailed and realistic information on craniofacial soft and hard tissues, and allows for easier, faster and more reliable analysis, even though some methods still have limitations. [13][14][15] Although Cone-Beam Computed Tomography (CBCT) could be considered the most accurate 3D imaging method for diagnosis and follow-up of orthodontic treatment results, its use is limited due to its high costs and, mainly, due to the exposure of patients to ionizing radiation. 16,17 Laser scanning, stereophotogrammetry (SPG), video-image, structured light scanners among other methods for obtaining 3D images without the use of radiation have already been proposed. [18][19][20][21] SPG scanners like the 3dMDface system and the Di3D system are able to generate very realistic and accurate 3D facial models. 22-24 However, these scanners need a dedicated and costly hardware, which may reduce their clinical application. Therefore, the search for alternative 3D facial models acquisition methods, with lower financial and biological costs is relevant in the current orthodontic scenario. In this context, Chaves LB, Barbosa TL, Casagrande CPM, Alencar DS, Capelli Jr J, Carvalho FAR -Evaluation of two stereophotogrametry software for 3D reconstruction of virtual facial models 6 some methods are promising for the clinical popularization of facial scanning, such as the Microsoft Kinect scanner, which generated average results, with a low cost hardware. 25 A viable alternative might be software that allows the 3D reconstruction from a series of 2D pictures, without the need of any specialized hardware rather than a photo camera, such as

MATERIAL AND METHODS
Sample size calculation was performed using the GPower3.1 software (University of Kiel, Germany) with a one sample t-test.
Based on parameters of a study with similar methodology 25 and considering a power of hypothesis test of 80%, a level significance of 0.05, to detect a difference in measurements of 2mm with a standard deviation of 2.03, at least eight volunteers would be needed for this research.
Ten patients were selected for this study, five males and five females, with a mean age of 24.4 years, who needed a fullhead CBCT scanning at the beginning of orthodontic treatment.
Exclusion criteria involved patients who were already using orthodontics appliances, with syndromes and /or craniofacial deformities, with a beard and/or mustache, with tattoos on the face, or with severe facial asymmetries. The selected patients agreed to participate with the use of their images in the study and signed an informed consent form. This study was approved     Table 1. One sample t-test was used to verify if the recorded differences were statistically different from 0. The level of significance was set at 0.05.

RESULTS
The one sample t-test results showed that there was a statistically significant difference between the ARIs (all with p < 0.05) in all groups compared. Models generated in 3DF Zephyr showed poor accuracy. When comparing CBCT vs ZAL, the mean differences were <4mm for all ARI, except for DN and NT. Better results were found when comparing CBCT vs ZFL, where the mean differences were < 3.2mm. Models generated in Agisoft showed acceptable accuracy for both groups (AAL and AFL). When comparing CBCT vs AAL, the mean differences were < 2.3mm. Comparing CBCT vs AFL, the mean difference was < 2.2mm (Fig 5 and Table 1).   The nose region was the most critical for all the evaluated situations; being challenging to accurately represent its morphology with all the evaluated methods. This limitation is remarkable but doesn't preclude its application in orthodontics since, the nose do not undergo major modifications following orthodontic treatment. This way, these softwares have shown promise for facial soft tissues evaluation when a tomographic scan is not justifiable due to its biological and/or financial cost. 16 The worst accuracy was observed for all methods in the NT region. This fact may be explained due to its simpler anatomy, being a small and usually smooth region without any marked anatomical feature, which implies a more challenging scenario for the software's algorithm to distinguish it from the background.
This NT region is also hard to be reconstructed from CBCT scans, due to its delicate soft-tissue-only composition, which barely attenuates the X-ray and therefore compromises its acquisi- In the present study, the photographs were taken indoors, with ambient light and, for a certain series of photos, with the aid of a ring flash. This fact may be limiting and determinant for the occurrence of shadows in certain ARIs, but it represents the ambient lightening of a typical dental office. New studies should be carried out in ambients with better lighting, to evaluate its influence on the quality of the generated models.
Another variable that could influence the models generated by the photo series is the quality of the camera sensor and the settings used. Since the methods evaluated by the present study are suitable for clinical follow up evaluation, the authors choice of a basic DSLR with ringflash was based on the hardware setup normally used by clinical dentists. Professional cameras, with their state-of-the-art sensors, could be able to compensate for the limited availability of light in the ambient, which would reduce artifacts generation in the two-dimensional images, making it easier, for the software, to identify landmarks to reconstruct the models. However, their higher prices would limit the diffusion of the proposed technique.
Chaves LB, Barbosa TL, Casagrande CPM, Alencar DS, Capelli Jr J, Carvalho FAR -Evaluation of two stereophotogrametry software for 3D reconstruction of virtual facial models 20 Likewise, especially in the 3DF Zephyr instruction manual, it is recommended to overlap 70-80% of surface in each photo, limiting the angles formed between them. It is advisable to make as many photographs as possible so that better final results may be obtained. However, this photo additions would increase the duration of acquisition, which could make the technique less viable.
Since the purpose of this study wass to verify the accuracy of the methodology, the number of photos taken was chosen to use the maximum potential of the software. Due to the lack of methodologies reported in the literature to date, and the constant evolution of software programming, less photos could be tested in future papers. However, the method is applicable for clinical use, with 5-minute average execution time for each photo sequence.
Other scientifically proven methods that use SPG have the disadvantage of their high acquisition cost²³, and new facial reconstruction methods are being developed, turning SPG more practical, 25 mainly by using mobile applications for images acquisition. However, these methods also need to be tested for accuracy, as there are no studies to this effect.
The models obtained through software were generated in three-dimensional meshes that were compared to those produced by CBCT. However, the same models with texturing of Chaves LB, Barbosa TL, Casagrande CPM, Alencar DS, Capelli Jr J, Carvalho FAR -Evaluation of two stereophotogrametry software for 3D reconstruction of virtual facial models facial tissues (Fig 6 ) are more pleasing to laypeople, and would allow subjective analyzes of patients submitted to the method, 3 which could be a good tool for qualitative comparison of results and treatment follow-up.