S16-01 SESSION 16: PLANNING/IMAGING – PART II QUANTIFICATION OF HEAD SHAPE FROM 3D PHOTOGRAPHY FOR AUTOMATIC DETECTION OF CRANIOSYNOSTOSIS

Abstract

Introduction: Although suture fusion in patients with craniosynostosis can be identified from CT images, these are typically acquired after detecting symptoms to avoid radiation. 3D photography has gained popularity in craniofacial imaging as a cheap and non-invasive modality. We present a method to automatically detect and quantify head malformations caused by craniosynostosis from 3D photographs. Methods: We collected CT images of 201 subjects without cranial pathology (age 1.93 ± 1.69 years) and 217 with various types of craniosynostosis (age 0.69 ± 1.16 years), and 3D photographs of 89 patients with craniosynostosis (age 1.17 ± 1.85 years) acquired with the 3DMDhead scan. We used image processing to segment the head from CT and 3D photography, and to label the cranial bones from CT automatically. The region of each cranial bone in 3D photography was determined using a reference template. We created a normative statistical head shape model from the subjects without pathology using principal component analysis. We calculated local head malformations for each patient by comparing their head shape with its projection on the statistical model using distances and curvature differences, and compared these values obtained from CT and from 3D photography using a paired Wilcoxon rank-sum test on a dataset of patients with both pre-operative CT image and 3D photograph (N=34). We used these metrics calculated our entire dataset to train a support vector machine classifier to distinguish between subjects with and without craniosynostosis using their head shapes, and to determine the type of craniosynostosis in patients with single suture fusion. We evaluated the accuracy of our methods using cross-validation. Results: We obtained similar quantifications of distances and curvature differences from normality from CT images and from 3D photographs using our paired dataset (p-values 0.39 and 0.44, respectively). We found significant differences in these metrics with respect to the normative model at all cranial bones in patients with sagittal fusion, and at the frontal area in patients with metopic or coronal fusion (p<0.001). We obtained 95.5% sensitivity and 95.5% specificity identifying patients with craniosynostosis with our classifier. In addition, we obtained 99.6% sensitivity and 99.3% specificity identifying the fused suture in patients with single fusion. Conclusion: Quantitative image analysis can identify and classify cranial malformations caused by craniosynostosis from 3D photography. Our approach has the potential to detect subtle malformations without radiation and before patients become symptomatic.