Quantifying faces three-dimensionally in orthodontic practice

Revealing the representative facial traits of different sagittal skeletal types: decipher what artificial intelligence can see by Grad-CAM

OBJECTIVES

Aesthetic improvement is a significant concern in dental therapy. While orthodontic treatment primarily targets hard tissue, the impact on soft tissue and the extent of these changes remains empirical. This study aims to unveil the intricate relationship between facial soft tissue and skeletal types using artificial intelligence (AI) analysis.

METHODS

First, we collected a dataset of 1044 3-side-photographs and categorized them based on cephalometric measurements. After pre-processing and data augmentation, samples were fed to two independent models (Sfa, Res model) for training and testing. After validating that the Sfa model could accurately recognize the skeletal types based merely on photographs, Grad-CAM algorithm was utilized for model decipherment. Verification of the vital traits were carried out by facial adjustment simulation.

RESULTS

The Sfa model demonstrated superior accuracy (0.9293) in identifying skeletal types based solely on soft tissue, compared to the Res model (0.8395) and even trained orthodontists (0.764), testifying our hypothesis that AI could be more capable of processing imperceptible cues compared to mankind. Intriguingly, Grad-CAM revealed that cheek volume, forehead, chin and nasolabial traits could be representative features of each type, exceeding the traditional knowledge which merely concerns mandible and chin.

CONCLUSION

By constructing a deep learning model as a classifier and then decipher it with Grad-CAM, we revealed the subtle and unnoticed cues associating skeletal and soft tissue, as well as provided a novel approach that could aid practitioners in devising tailored treatment plans for enhanced esthetic outcomes.

CLINICAL SIGNIFICANCE

The proposed AI methods offer valuable assistance to practitioners in identifying uncoordinated facial traits that may detract from a patient's attractiveness. By incorporating these insights into customized treatment plans, dental therapy can maximize esthetic benefits for individual patient.

Correlation Analysis of Nasal Septum Deviation and Results of AI-Driven Automated 3D Cephalometric Analysis

The nasal septum is believed to play a crucial role in the development of the craniofacial skeleton. Nasal septum deviation (NSD) is a common condition, affecting 18-65% of individuals. This study aimed to assess the prevalence of NSD and its potential association with abnormalities detected through cephalometric analysis using artificial intelligence (AI) algorithms. The study included CT scans of 120 consecutive, post-traumatic patients aged 18-30. Cephalometric analysis was performed using an AI web-based software, CephX. The automatic analysis comprised all the available cephalometric analyses. NSD was assessed using two methods: maximum deviation from an ideal non-deviated septum and septal deviation angle (SDA). The concordance of repeated manual measurements and automatic analyses was assessed. Of the 120 cases, 90 met the inclusion criteria. The AI-based cephalometric analysis provided comprehensive reports with over 100 measurements. Only the hinge axis angle (HAA) and SDA showed significant (p = 0.039) negative correlations. The rest of the cephalometric analyses showed no correlation with the NSD indicators. The analysis of the agreement between repeated manual measurements and automatic analyses showed good-to-excellent concordance, except in the case of two angular measurements: LI-N-B and Pr-N-A. The CephX AI platform showed high repeatability in automatic cephalometric analyses, demonstrating the reliability of the AI model for most cephalometric analyses.

An automated parametric ear model to improve frugal 3D scanning methods for the advanced manufacturing of high-quality prosthetic ears

Ear prostheses are commonly used for restoring aesthetics to those suffering missing or malformed external ears. Traditional fabrication of these prostheses is labour intensive and requires expert skill from a prosthetist. Advanced manufacturing including 3D scanning, modelling and 3D printing has the potential to improve this process, although more work is required before it is ready for routine clinical use. In this paper, we introduce a parametric modelling technique capable of producing high quality 3D models of the human ear from low-fidelity, frugal, patient scans; significantly reducing time, complexity and cost. Our ear model can be tuned to fit the frugal low-fidelity 3D scan through; (a) manual tuning, or (b) our automated particle filter approach. This potentially enables low-cost smartphone photogrammetry-based 3D scanning for high quality personalised 3D printed ear prosthesis. In comparison to standard photogrammetry, our parametric model improves completeness, from (81 ± 5)% to (87 ± 4)%, with only a modest reduction in accuracy, with root mean square error (RMSE) increasing from (1.0 ± 0.2) mm to (1.5 ± 0.2) mm (relative to metrology rated reference 3D scans, n = 14). Despite this reduction in the RMS accuracy, our parametric model improves the overall quality, realism, and smoothness. Our automated particle filter method differs only modestly compared to manual adjustments. Overall, our parametric ear model can significantly improve quality, smoothness and completeness of 3D models produced from 30-photograph photogrammetry. This enables frugal high-quality 3D ear models to be produced for use in the advanced manufacturing of ear prostheses.

Influence of the nutritional status on facial morphology in young Japanese women

Evidence regarding the possible influence of nutritional status on the facial morphology has thus far been insufficient. We examined whether or not the physical body compositions and dietary behaviors were correlated with any morphological characteristics of the face. One hundred and fifteen young Japanese women participated. Variables representing the dietary behaviors were extracted from self-reported survey data, and corresponding three-dimensional (3D) facial images and body compositions were examined. Multivariate analyses identified significant relationships between the nutritional status and facial topography (p < 0.05). The clustering method revealed the existence of three dietary condition patterns ("balanced diet", "high-calorie-diet" with obesity tendency, and "imbalanced low-calorie-diet" with sarcopenic obesity tendency). Among these three patterns, a round face (increased facial width; analysis of variance [ANOVA], p < 0.05) was observed in the high-calorie-diet pattern, while the imbalanced low-calorie-diet pattern showed a more masculine face (increased face height, decreased eye height, increased non-allometric sexual shape differences; ANOVA, p < 0.05), thus suggesting the possibility of sex-hormonal influences. In summary, the body composition and dietary behaviors were found to influence the facial morphology, and potential biological influences were discussed.

Development of novel artificial intelligence systems to predict facial morphology after orthognathic surgery and orthodontic treatment in Japanese patients

From a socio-psychological standpoint, improving the morphology of the facial soft-tissues is regarded as an important therapeutic goal in modern orthodontic treatment. Currently, many of the algorithms used in commercially available software programs that are said to provide the function of performing profile prediction are based on the false assumption that the amount of movement of hard-tissue and soft-tissue has a proportional relationship. The specification of the proportionality constant value depends on the operator, and there is little evidence to support the validity of the prediction result. Thus, the present study attempted to develop artificial intelligence (AI) systems that predict the three-dimensional (3-D) facial morphology after orthognathic surgery and orthodontic treatment based on the results of previous treatment. This was a retrospective study in a secondary adult care setting. A total of 137 patients who underwent orthognathic surgery (n = 72) and orthodontic treatment with four premolar extraction (n = 65) were enrolled. Lateral cephalograms and 3-D facial images were obtained before and after treatment. We have developed two AI systems to predict facial morphology after orthognathic surgery (System S) and orthodontic treatment (System E) using landmark-based geometric morphometric methods together with deep learning methods; where cephalometric changes during treatment and the coordinate values of the faces before treatment were employed as predictive variables. Eleven-fold cross-validation showed that the average system errors were 0.94 mm and 0.69 mm for systems S and E, respectively. The total success rates, when success was defined by a system error of < 1 mm, were 54% and 98% for systems S and E, respectively. The total success rates when success was defined by a system error of < 2 mm were both 100%. AI systems to predict facial morphology after treatment were therefore confirmed to be clinically acceptable.

Statistical shape modelling for the analysis of head shape variations

The aim of this study is, firstly, to create a population-based 3D head shape model for the 0 to 2-year-old subjects to describe head shape variability within a normal population and, secondly, to test a combined normal and sagittal craniosynostosis (SAG) population model, able to provide surgical outcome assessment. 3D head shapes of patients affected by non-cranial related pathologies and of SAG patients (pre- and post-op) were extracted either from head CTs or 3D stereophotography scans, and processed. Statistical shape modelling (SSM) was used to describe shape variability using two models - a normal population model (MODEL1) and a combined normal and SAG population model (MODEL2). Head shape variability was described via principal components analysis (PCA) which calculates shape modes describing specific shape features. MODEL1 (n = 65) mode 1 showed statistical correlation (p < 0.001) with width (125.8 ± 13.6 mm), length (151.3 ± 17.4 mm) and height (112.5 ± 11.1 mm) whilst mode 2 showed correlation with cranial index (83.5 mm ± 6.3 mm, p < 0.001). The remaining 9 modes showed more subtle head shape variability. MODEL2 (n = 159) revealed that post-operative head shape still did not achieve full shape normalization with either spring cranioplasty or total calvarial remodelling. This study proves that SSM has the potential to describe detailed anatomical variations in a paediatric population.

Surface-based 3-dimensional cephalometry: An objective analysis of cranio-mandibular morphology

INTRODUCTION

The purpose of this study was to apply a previously reported homologous model for quantifying and visualizing the three-dimensional (3D) cranio-mandibular morphology.

METHODS

Twenty-eight 3D cone-beam computed tomography (CBCT) images of Japanese patients (7 to 13 years of age) showing skeletal Class 1 malocclusions were used in this study. Wire mesh fitting was conducted for each patient based on the assignment of landmarks to each 3D surface (homologous modeling). Cranial and mandibular images were generated using surface rendering of the CBCT images. The mean and standard deviation of each point on the wire mesh were calculated as the normative mean. Two operators conducted the process twice for seven patients, with a one-week interval between fittings. The intra- and inter-examiner reliability were determined using the 95% confidence interval minimal detectable change (MDC₉₅) for random error and intra-class correlation coefficients (ICCs).

RESULTS

The intra- and inter-examiner reliability of the mesh fitting method were almost perfect. Normative ranges of the cranial and mandibular surfaces of Japanese were determined.

CONCLUSIONS

A clinical method that enables practitioners to quantify and visualize the hard tissues of a patient's face in three dimensions with almost perfect reliability is presented. This method allows practitioners to evaluate how patients' skeletal characteristics differ from normative means in three dimensions.