Facial morphology analysis in osteogenesis imperfecta types I, III and IV using computer vision

1 INTRODUCTION

Osteogenesis Imperfecta (OI) is a rare, heritable bone disease. It is caused, in most cases, by dominant mutations in COL1A1 or COL1A2, the genes encoding collagen type I; however, mutations in about 20 other genes can also lead to OI.1 The main clinical feature of OI is increased bone fragility, and growth is often impaired.2 OI’s range of phenotypic severity is wide, and the disease's severity is assessed using the Sillence clinical classification that distinguishes four major OI types: type I (‘mild’), type IV (‘moderate’), type III (‘severe’) and type II (‘lethal’).3 No cure for OI is available, and treatment usually involves intravenous bisphosphonates to strengthen bones and, thus, decrease fracture rates.4

The disease's genetic alterations can result in significant craniofacial implications.5 Qualitative assessments of affected patients’ facial morphology include a triangular facial shape, low-set ears and frontal bossing.6-8 To our knowledge, no previous quantitative study examining the facial morphology of individuals with OI has been published.

Quantitative facial analysis can be performed using anthropomorphic measurements or facial proportions; however, phenotypical manifestations are difficult to describe using these methods, so genetic diseases’ facial manifestations are often described qualitatively.9 This qualitative approach is highly prone to bias and is time-consuming.

The current study's goals were to establish a reproducible method of assessing facial morphology, to compare the facial morphology of individuals with OI with a normocephalic control group, to define the facial characteristics of OI types I, III and IV quantitatively, and to train a supervised machine learning model to detect whether a subject is affected by OI, based on their facial morphology. Our null hypothesis was that no statistically significant differences occurred in the facial morphology of subjects affected by different OI types compared to a normocephalic control group. This study was supported by the Brittle Bone Disorder Consortium (BBDC). The BBDC is part of the Rare Diseases Clinical Research Network (RDCRN), an initiative of the Office of Rare Diseases Research (ORDR) at the National Institute of Arthritis and Musculoskeletal and Skin (NCATS).

2 MATERIALS AND METHODS 2.1 Study population

This case-control study was conducted with a total of 306 participants (145 male and 161 female) aged 3-55 years (Table 1) The BBDC 7701 study took place at the Shriners Hospital for Children—Canada, and 173 subjects who had been diagnosed with OI participated. These participants were grouped according to their OI type. The study's groups shared similar age and sex distributions (Table 1).

TABLE 1. Measurement and statistical testing results Control OI-I OI-III OI-IV§ P value N (M/F) 133 (68/65) 88 (42/46) 28 (9/19) 57 (26/31) P (Chi Square) = .32 Mean Age (SD) 17.7 (7.7) 19.7 (14) 15.4 (8.4) 17.1 (8.4) P (ANOVA) = .19 Mean Ratio 1 (SD) 0.612a (0.026) 0.595 (0.026) 0.604 (0.036) 0.602 (0.031) P (ANOVA) = <.05 Mean Ratio 2 (SD) 0.799ab (0.027) 0.784 (0.038) 0.774 (0.030) 0.791 (0.031) P (ANOVA) = <.05 Mean Ratio 3 (SD) 1.26abc (0.075) 1.33b (0.106) 1.40c (0.124) 1.33 (0.097) P (ANOVA) = <.05 Mean LFH (SD) 0.569(0.022) 0.571 (0.027) 0.562 (0.035) 0.567 (0.026) P (ANOVA) = .47 Note Bonferroni's post hoc analysis a: P < .05 in comparison to OI-I b: P < .05 in comparison to OI-III c: P < .05 in comparison to OI-IV.

The control group comprised 133 patients who were recruited from the McGill Undergraduate Dental Clinic for orthodontic treatment and who agreed to participate in this study. Control subjects were randomly selected to approximate the OI groups’ age and gender, representing 75% of the total number of OI-affected subjects. We aimed at a control-group size equal or greater to 70% of our OI sample subgroups combined, given the limitations to the number of control participants available and the time required to extract the study's data. An orthodontist reviewed patients’ facial photographs to ensure that the control-group subjects had a normocephalic facial type and no major skeletal discrepancies or malocclusions. All participants from the clinic signed an informed consent form. Additionally, the patients whose photographs were used in this study's figures signed a release form. This study was approved by the McGill University Faculty of Medicine Institutional Review Board.

2.2 Facial photographs

Our analysis used antero-posterior facial photographs. These pictures were all taken by the same operator using a Canon Rebel T5™ camera with a Tamson™ SpDi macro 90 mm lens and a Canon Macro Ring Lite MR-14 EXII Flash™. The camera settings comprised: AV mode, an aperture of f/11, an ISO of −125, a ‘faithful’ picture style, a flash-mode white balance, image quality of ‘JPEG L’, an ‘evaluative’ metering mode, a ‘single shooting’ drive mode, autofocus (with subjects 1.5 m away from the camera lens) and portrait orientation. Subjects faced the camera head-on and were unable to rotate or tilt their heads. The entirety of subjects’ heads and necks was included in their respective pictures. All of the study's photographs were taken with a white background. Subjects were required to clear their faces of any hair. Any pictures that did not meet these requirements of a proper extra-oral orthodontic photograph were excluded from the study.

2.3 Analysis of facial photographs

Landmark coordinates were extracted from the subject photographs using the PFLA programme.10 This programme uses deep learning models11, 12 to detect and annotate faces with 68 landmarks (Figure 1A). These landmarks’ coordinates were used for our statistical analyses with the R programming language. The predictive machine learning models were implemented in Python 3 and MATLAB.

image

Face Detection and landmark placement. A, Automatic face detection (the green rectangle) and 68 landmarks’ annotation. B, Four facial ratios computed using the landmark coordinates from the PFLA programme

The proposed morphological method for facial analysis used all 68 automatically detected landmarks. This methodology is not tied to any particular set of landmarks, and it can be applied to other sets of facial annotations. We selected these landmarks since they were pertinent in identifying OI’s previously reported craniofacial manifestations.7

Our morphological approach to understanding OI manifestations had two objectives: determining whether any difference occurred between the study's groups and, if so, identifying these differences. Figure 2 illustrates our shape processing and analysis methodology.

image

Morphological analysis method. The study population was subdivided into subgroups (OI types I, III and IV and control). Generalized Procrustes analysis (GPA) was performed for each group and output the groups’ mean shapes. The four mean shapes were compared using Goodall's F-test. The four mean shapes were then aligned using GPA, and landmark distances were computed to compare morphological differences

Generalized Procrustes analysis (GPA) was used to normalize and align each group's shapes, allowing for an accurate comparison.13 Changes in morphology were deemed more important than measurements since they offer more information about OI’s phenotypic manifestations. GPA was performed for both rotation and scaling to ensure optimal superimposition. A second GPA was performed on the groups’ mean shapes, which allowed for comparisons with our landmark distance calculations.

Goodall's F-test was used to determine whether groups’ facial morphology differed. This test identified statistically significant differences between the groups’ mean shapes.14

To explore and interpret differences, Euclidean distances were computed between corresponding landmarks for each mean shape compared to the control group's baseline mean shape. The resulting landmark distances were grouped according to their anatomical areas to improve the results’ interpretability (Figure 3B). We refer to these differences as anatomical shape discrepancy (Equation 1).

image

Anatomic discrepancy plot and landmark labels. A, Osteogenesis imperfecta groups’ mean shapes’ landmark distances compared to the control group's mean shape. Distances were compiled using anatomic regions to enhance our results’ interpretability. These distances are relative, and units were not measured in absolute values since the shapes were superimposed by generalized Procrustes analysis. B, Anatomic labelling of the facial landmarks

Equation1 : Anatomical Shape Discrepancy. urn:x-wiley:16016335:media:ocr12491:ocr12491-math-0001(1)

To validate our novel approach, we compared the results of our proposed morphological analysis with facial ratios analyses commonly used in orthodontics.15 Four facial ratios were measured (Equation 2) biocular width to bitemporal width (Ratio 1), bimandibular width to bitemporal width (Ratio 2), bitemporal width to facial height (Ratio 3) and lower face height (LFH) (Figure 1B). The study's groups were compared using analysis of variance (ANOVA) with Bonferroni's post hoc tests.

Equation2 : Facial Ratios. urn:x-wiley:16016335:media:ocr12491:ocr12491-math-0002(2)

Predictive models were trained to detect OI types. Three binary classifiers were trained. The training and test sets were created from an 80/20 split of the data set and used for all our models.

The first model, named PCALog (Figure 4), comprised 138 input features, 68x and 68y coordinates, and a subject's age (1) and sex (1). For this model, a principal component analysis (PCA) was used to reduce the dimensions to 30 features, which were then passed through a logistic regression layer (Figure 4).

image

Schematic representation of the PCALog model

The second and third models were pre-trained AlexNet16 and VGG1617 convolutional neural networks. Both models used RGB images with a fixed size of 224 pixels in height and width as inputs. Transfer learning was used for both deep learning architectures.

To assess the predictive models’ efficacy, precision-recall and F1 scores were calculated for each class and for the test set's total results (Equation 3).

Equation3 : Prediction Model Accuracy Metrics. urn:x-wiley:16016335:media:ocr12491:ocr12491-math-0003(3) 3 RESULTS

All of our mean shape comparisons between the OI type groups and the control group were statistically significant—except for our comparison between individuals with OI type I and individuals with OI type IV. Figure 3A summarizes the morphological differences between the OI type groups compared to the control group, according to their anatomical regions. OI type III showed considerably greater morphological differences compared to OI types I and IV. OI type I most closely resembled the control group. Overall, subjects’ temple and eye landmarks showed the most pronounced discrepancies compared to the baseline shape.

Table 1 summarizes the study group's ratios. Between-group comparisons were statistically significant for all facial metrics except lower face height (LFH). The biocular width to bitemporal width (Ratio 1) differed between the control group and the OI type I group. Bimandibular width to bitemporal width (Ratio 2) differed between the control group and the OI types I and III groups. Bitemporal width to facial height (Ratio 3) was statistically different for the control group and all the OI type groups, as well as between OI types I and III groups and the OI types III and IV groups.

Table 2 summarizes the different models’ accuracy metrics. Precision indicates the positive predictive value and recall indicates the sensitivity. The F1-score is a metric that tests a predictive model's accuracy, considering both precision and recall. The model based on facial morphology classifiers performed perfectly in our test set. The AlexNet and VGG models performed less accurately.

TABLE 2. Accuracy metrics from the classifiers test set Model Class Precision Recall F1-Score PCALog Control 1.00 1.00 1.00 OI Subject 1.00 1.00 1.00 AlexNet Control 0.88 0.78 0.82 OI Subject 0.84 0.91 0.88 VGG16 Control 0.86 0.89 0.87 OI Subject 0.91 0.89 0.90 Abbreviation: OI, Osteogenesis imperfecta. 4 DISCUSSION

In this study, we used automated facial morphology analysis based on computer vision to objectively assess OI-related facial characteristics. This novel approach allowed us to easily and efficiently conduct face detection, labelling and shape analysis without human annotation. We found the morphological differences between individuals with OI and the control group to be more pronounced at the level of subjects’ eyes and temples. Morphological differences across all face labels were consistently more severe for the OI type III group. Using GPA and other shape analysis methods, we reliably compared our study groups’ facial morphology. This approach simplified our interpretation of our findings using an anatomic discrepancy plot. Moreover, we were able to train a model that can classify OI subjects and control subjects based on their morphological features.

The OI and control samples were not matched according to subjects’ ethnicity, given the amount of variation among the OI subjects. The sample's heterogeneous ethnicity presents a limitation to our study vis-à-vis our mean shape and predictive models. Our facial shape analyses were conducted on groups that included male and female participants; additionally, we did not further subdivide the groups according to subjects’ age. Further studies could investigate such morphological differences according to subjects’ genders and changes with age.

We used 68 landmarks to compare the groups’ mean shapes and evaluate differences in facial features. Mean shapes were computed using all landmarks at once. We compared facial features by grouping landmarks according to their anatomic labels (Figure 3B). A mean shape comparison allowed us to determine whether patients had similar facial morphology. Anatomic regions’ landmark distances detailed the groups’ differences. Figure 3A illustrates the differences in anatomic regions between each group. A statistically different facial morphology, associated with deviations in anatomic regions compared to the control group, could signify underlying craniofacial malformations. Therefore, this difference should be considered during orthodontics’ diagnostic phase.

Individuals with OI types I, III and IV presented with different facial morphologies compared to our control group. All three ratios that incorporated bitemporal width differed significantly between groups. The LFH measurement, which only considered the face's vertical aspect, was the only metric that showed no intergroup differences. Thus, differences in facial width are a key characteristic of OI facial morphology. These findings align with the triangular facial morphology described in previous articles.7, 9, 18

Our results also associate OI with a different morphology at the level of the temples. Differences in mean shapes can be easily visualized and interpreted using an anatomic discrepancy plot. In a separate study using cone-beam–computed tomography, we also observed prominent temporal bones among severe OI patients.19 Our present analysis suggests that these manifestations vary by OI type and that OI type III patients are the most severely and consistently affected among OI groups.

Our morphological approach is a more reliable method of comparing facial morphology than any single measurement or metric, such as facial ratios. A more intuitive interpretation of our results, encompassing all the facial landmarks, allowed us to determine where differences occurred. Mean shape testing seemed the most efficient way to detect morphological differences between groups since it incorporates the complete set of landmarks at once. Moreover, anatomic discrepancy analysis provided us with more precise information about such differences and was less cumbersome to interpret. This type of analysis could extend in future studies correlating morphological changes with genetic mutations. This technique could also be adapted to automate the analysis of other types of medical imaging, such as computed tomography (CT) scans, radiographs and 3D models. Such an adaptation could achieve a deeper understanding of various genetic mutations’ manifestations. In orthodontics, these techniques offer potential as a reliable method to assess and track the effects of orthodontic and dentofacial orthopaedic treatments, orthognathic surgery or cleft palate management on soft-tissue facial morphology.

Given the small data set of subjects with OI, we had insufficient data to train and evaluate a multi-class classifiers model. Including too few types III and IV patients would have created a skewed training set prone to overfitting. Moreover, this small sample makes assessing a model's accuracy difficult. For these reasons, we aimed to train a simple binary classifier in order to illustrate the use of machine learning models with minimal data.

Our trained models’ results are especially promising for the detection and classification of OI. We observed the efficiency of logistic regression using landmark data, compared to pre-trained deep neural networks. The PCALog model outperformed common deep learning architectures—probably due to our small sample size. The ability to train successful classifiers using landmarks and general subject data as input features offers potential for new diagnostics aids. Since OI is a rare disease, the main limitation in our models’ sophistication and accuracy is the size of our study's data set. Simple models can still perform effectively for preprocessed data, such as landmarks. Our PCALog model was trained only on control and OI patients. At its current developmental stage, this model cannot be considered a diagnostic model or screening tool. OI patients may share some morphological features with other syndromes, which could lead to misdiagnosis. The efficacy of our PCALog model, trained on small amounts of data, shows potential for clinical applications of machine learning methods to assess and classify rare diseases and syndromes’ facial features. These models and methods may be applied to other conditions associated with changes in facial morphology, such as Marfan, Crouzon, Ehler-Danlos and Treacher Collins syndromes.

5 CONCLUSIONS

Facial morphology among OI patients presents specific features, mostly at the level of the temples and the eyes. Using a small data set, we successfully trained our PCALog model to distinguish individuals with OI from controls, based on morphological landmarks.

ACKNOWLEDGEMENTS

The Brittle Bone Disease Consortium (1U54AR068069-0) is part of the National Center for Advancing Translational Sciences (NCATS) at the Rare Diseases Clinical Research Network (RDCRN), and it is funded through a collaboration between the Office of Rare Diseases Research (ORDR), NCATS, the National Institute of Arthritis and Musculoskeletal and Skin Diseases (NIAMS), the National Institute of Dental and Craniofacial Research (NIDCR), and the Eunice Kennedy Shriver National Institutes of Child Health and Development (NICHD). This article's authors are solely responsible for its content, and this article does not necessarily represent the National Institutes of Health's official views. The Brittle Bone Disease Consortium is also supported by the Osteogenesis Imperfecta Foundation. The authors also acknowledge the contributions of the following BBDC site coordinators: M. Abrahamson (OHSU), S. Alon (UCLA), M. Azamian and A. Turner (BCM), C. Brown (Nemours Alfred I. duPont Hospital), E. Carter and E. Yonko (HSS), A. Caudill (Chicago Shriners), K. Dobose (KKI), M. Durigova (Montreal Shriners Hospital for Children), A. Giles and E. Rajah (CNMC), M. Gross-King (Tampa Shriners), and E. Strudthoff (UNMC).

CONFLICT OF INTEREST

Frank Rauch: Dr Rauch reports grants from National Institutes of Health (NIH), during the conduct of the study. Personal fees from Novartis, personal fees from Mereo, grants from Mesentech, outside the submitted work.

AUTHOR CONTRIBUTIONS

Maxime Rousseau: Participated in conceptualization, data curation, formal analysis, investigation, methodology, software and writing. Javier Vargas: Participated in methodology, software and writing. Frank Rauch: Participated in supervision, validation and writing. Juliana Marulanda: Participated in project administration and writing. Jean-Marc Retrouvey: Participated in conceptualization, supervision, validation and writing.

APPENDIX

Brendan Lee, V. Reid Sutton, Sandesh CS Nagamani, Francis Glorieux, Janice Lee, Paul Esposito, Maegen Wallace, Michael Bober, David Eyre, Danielle Gomez, Gerald Harris, Tracy Hart, Mahim Jain, Deborah Krakow, Jeffrey Krischer, Eric Orwoll, Lindsey Nicol, Cathleen Raggio, Peter Smith, and Laura Tosi.

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

REFERENCES

1Tauer JT, Robinson M-E, Rauch F. Osteogenesis imperfecta: new perspectives from clinical and translational research. JBMR Plus. 2019; 3(8):e10174. https://doi.org/10.1002/jbm4.10174 2Graff K, Syczewska M. Developmental charts for children with osteogenesis imperfecta, type I (body height, body weight and BMI). Eur J Pediatr. 2017; 176(3): 311- 316. https://doi.org/10.1007/s00431-016-2839-y 3Sillence DO, Rimoin DL. Classification of osteogenesis imperfect. Lancet. 1978; 1(8072): 1041- 1042. 4Rossi V, Lee B, Marom R. Osteogenesis imperfecta: advancements in genetics and treatment. Curr Opin Pediatr. 2019; 31(6): 708- 715. https://doi.org/10.1097/MOP.0000000000000813 5Shapiro JR, Byers PH, Glorieux FH, Sponseller PD Osteogenesis imperfecta: a translational approach to brittle bone disease. Academic Press; 2013. http://site.ebrary.com/id/10756382 Accessed April 21, 2021. 6Rousseau M, Retrouvey JM. Osteogenesis imperfecta: potential therapeutic approaches. PeerJ. 2018; 6: e5464-e5464. 7Chang PC, Lin SY, Hsu KH. The craniofacial characteristics of osteogenesis imperfecta patients. Eur J Orthod. 2007; 29(3): 232- 237. 8Rizkallah J, Schwartz S, Rauch F, et al. Evaluation of the severity of malocclusions in children affected by osteogenesis imperfecta with the peer assessment rating and discrepancy indexes. Am J Orthod Dentofacial Orthop. 2013; 143(3): 336- 341. 9Retrouvey J-M, Taqi D, Tamimi F, et al. Oro-dental and cranio-facial characteristics of osteogenesis imperfecta type V. Eur J Med Genet. 2019; 62(12): 103606. https://doi.org/10.1016/j.ejmg.2018.12.011 10Rousseau M, Retrouvey JM. pfla: a python package for dental facial analysis using computer vision and statistical shape analysis. J Open Source Softw. 2018; 3(32): 855-855. 11Zhang K, Li Z, Qiao Y, Zhang Z. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Process Lett. 2016; 23(10): 1499- 1503. https://doi.org/10.1109/LSP.2016.2603342 12Sagonas C, Tzimiropoulos G, Zafeiriou S, Pantic M A semi-automatic methodology for facial landmark annotation. In: 2013 IEEE Conference on Computer Vision and Pattern Recognition Workshops. 2013:896-903. https://doi.org/10.1109/CVPRW.2013.132 13Airey DC, Wu F, Guan M, Collins CE. Geometric morphometrics defines shape differences in the cortical area map of C57BL/6J and DBA/2J inbred mice. BMC Neurosci. 2006; 7: 63. https://doi.org/10.1186/1471-2202-7-63 14Goodall C. Procrustes methods in the statistical analysis of shape. J R Stat Soc Ser B Methodol. 1991; 53(2): 285- 321. https://doi.org/10.1111/j.2517-6161.1991.tb01825.x 15Negi G, Ponnada S, Aravind NKS, Chitra P. Photogrammetric correlation of face with frontal radiographs and direct measurements. J Clin Diagn Res JCDR. 2017; 11(5): ZC79- ZC83. https://doi.org/10.7860/JCDR/2017/28249.9924 16Krizhevsky A, Sutskever I, Hinton GE Imagenet classification with deep convolutional neural networks. Adv Neural Inf Process Syst. 2012; 25: 1097- 1105. 17Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. ArXiv Prepr ArXiv14091556. Published online. 2014. https://arxiv.org/abs/1409.1556 Accessed January 25, 2021. 18Foster BL, Ramnitz MS, Gafni RI, et al. Rare bone diseases and their dental, oral, and craniofacial manifestations. J Dent Res. 2014; 93(7 Suppl): 7S- 19S. 19Reznikov N, Dagdeviren D, Tamimi F, Glorieux F, Rauch F, Retrouvey J-M. Cone-beam computed tomography of osteogenesis imperfecta types III and IV: three-dimensional evaluation of craniofacial features and upper airways. JBMR Plus. 2019; 3(6):e10124. https://doi.org/10.1002/jbm4.10124

留言 (0)

沒有登入
gif