1. Introduction
Craniomaxillofacial reconstructive surgery is a complex and difficult multidisciplinary technique due to the intricate anatomy of the skull. The aim of craniomaxillofacial reconstruction surgery is to preserve the patient’s appearance, strengthen facial functions, and regain the bilateral symmetry of the craniofacial skeleton. However, craniomaxillofacial reconstructive surgery may lead to complications such as bone disorders, congenital deformities, trauma, pathologies, genetic abnormalities, and cancers. A precise recognition of the bilateral symmetry facial midplane is an imperative step for pre-surgical planning and implant design techniques. For facial restoration, this midplane plays a major role when one side of the image is replicated and used as a guide to recreate the deformed or injured side. The authors of [
1] have established a widely agreed approach for defining the midline of the craniofacial skeleton. To date, however, the most popular two-dimensional image application method, or the midsagittal plane (MSP) for a three-dimensional object, is the method introduced by [
2].
There are a few approaches that aim to simplify the task of locating the midline plane of the facial skeleton. One technique proposed by [
3] describes a semiautomatic system that, in conjunction with surface models reconstructed from computed tomography images (CT), uses principal component analysis (PCA) and the iterative closest point (ICP) alignment method. The first step is to determine the direction of the mirror plane correctly. This was achieved by using PCA to match the replicated mesh and the initial mesh roughly. Then, the ICP algorithm was described by a refined registration. The downside of this approach was the dependency on the central point of the image for the approximation of the symmetrical plane (obtained using the average of the vertices of the facial mesh). If the central point was in the wrong position due to any external factors (such as imperfect symmetry), this approach would lead to a symmetrical plane in the wrong direction and position. In addition, this algorithm is not able to adjust and learn from previous images to improve its performance, limiting its capabilities.
Alternatively, [
4] determines the midline symmetry plane by using boney landmarks to create a midline representing facial symmetry. For a stack of horizontal lines crossing bilaterally through the facial skeleton containing boney landmarks, this approach essentially measures the midline symmetry plane as a perpendicular midpoint. This approach involves the manual collection of a variety of cephalometric boney landmarks in the dataset by either specifically locating the landmarks on the plane (which requires great attention by an expert user) or by using the midline as a reference and locating the landmarks at equal distances from the midline. However, manual skeletal landmark selection is ineffective, time-consuming, and reliant on an expert operator, resulting in errors in the measurement of the symmetry plane [
5] outlines an ICP-based process for automated symmetry plane detection of 3D asymmetrically scanned human faces that uses particle swarm optimization. This approach starts with a discrete 3D model. The symmetry plane is tested by a tentative first attempt using a PCA algorithm. The model is then refined iteratively by a Levenberg–Marquardt algorithm before its final prediction is obtained. This revised version enhances the shortcomings of [
6], but the current implementation also struggles to integrate self-learning to maximize the result of the model and misses the ability to learn from previous versions.
By minimizing the error-index of the symmetry plane, an automated method based on an iterative process was recently proposed by [
7]. To automatically correct the initial symmetry plane, with a significant contribution to the use of the rotation matrix derived from the registration process, this method performs analytical data analysis in 3D point sets derived from CT images. First, the plane was divided into two groups by the initial symmetry plane estimated by the PCA and the collection of skull points. Then, to match two point-sets, the ICP registration method was used.
Most recently, ref. [
8] introduced a novel automatic concept for determining the bilateral symmetry midline of the facial skeleton based on invariant moments. This technique creates a dataset from images aligned using cephalometric landmarks. The images are then rotated from 14° to 15° with a resolution of 0.5°. Then, after comparing different feature extractors, pseudo-zernike moments were selected for having the best accuracy using the k-nearest neighbors classifier. Finally, after detecting the rotation degree of the image, the midpoint is calculated using geometric moments. However, this model still has some limitations. For instance, this method uses 2D images with an image resolution of 128 × 128 which becomes difficult in real applications on different image modalities such as computed tomography and magnetic resonance imaging (MRI). Additionally, this technique was not tested on non-symmetrical skull images which may affect its results.
Thus, this study aims to present a 3D technique for automatically calculating the craniofacial symmetry midplane from CT scans using convolutional neural network (CNN) and geometric moments.
Figure 1 shows the overview of the proposed method. First, using 3D U-net, the skull is removed from CT images to create a dataset. Then, based on the cephalometric landmarks, the CT image is aligned in the coronal and transverse planes. The image is then duplicated and two datasets of 441 images, per image, of 0.5° resolution is created from −5° to 5°. These sets of images are presented to a 3D rotate invariant CNN. After CNN determines the rotation degree of these images in the coronal and transverse planes, the skull midpoints are calculated using 3D geometric moments. Finally, by joining the midpoints and grades described by the CNN, the midplanes can be constructed.
4. Conclusions
This study aimed to introduce a 3D automatic technique for determining the craniofacial symmetry midplane from CT scans using the convolutional neural network and gemetric moments. A total of 195 symmetric CT images were used to evaluate this method using the CQ500 database while eight asymmetric CT images from TCIA database were used to evaluate the performance in asymmetric images. For symmetric images, this method obtained results close to 99%. However, for asymmetric images, the method needs further development to improve its results. CNNs offer an effective alternative to the pseudo-zernike moments method and conventional landmark-based symmetry scores that depend on the general positions of cephalometric landmarks. CNNs are also an alternative to PCA-ICP techniques, which depend on the manual selection of the central point which cannot be improved. With the proposed technique, the central point could be found as the centroid of an image, and then the symmetrical midplane can be constructed for symmetric images. In this study, we have shown the proposed technique to be reliable and to provide the midplane symmetry plane with great accuracy in symmetric images. This method can be used as a tool to aid surgeons in reconstructive craniofacial surgeries.