OrthoForecast

To predict the soft tissue profile after orthognathic surgery, OrthoForecast uses data provided by patients who have previously undergone orthognathic surgery. The OrthoForecast database contains information obtained from 400 patients, including 100 with facial asymmetry, 100 with a skeletal class II jaw relationship, and 200 with a skeletal class III jaw relationship. These patients underwent sagittal split ramus osteotomy (SSRO; 386 cases) or intraoral vertical ramus osteotomy (IVRO; 14 cases), with (352 cases) or without (48 cases) Le Fort I osteotomy. Table 1 shows a summary of the entire subject sample. Frontal and lateral cephalograms and facial photographs were taken routinely before and more than 1 year after surgery and stored in JPEG format. It was assumed that any postsurgical oedema in the soft tissue had resolved by the stated postoperative time point (>1 year after surgery). Lateral and frontal cephalograms and photographs were taken with the teeth in intercuspal position and the head in a position such that the Frankfort horizontal plane was parallel to the floor.

The data were standardized as follows. The orbitale–porion distance of all subjects was defined based on lateral cephalograms. Photographs were superimposed on each cephalogram by changing their size. The size of the frontal cephalogram was adjusted by the size of the lateral cephalogram. Then, according to the adjusted frontal cephalogram, the size of the frontal photograph was modified referring to the distance between the ear rod and soft tissue menton in lateral view. Finally, the modified photograph was superimposed on the frontal cephalogram.

A single examiner located 24 feature points on each photograph, including 11 points on the frontal photograph and 13 points on the lateral photograph, as shown in Figure 1 and described in detail in Table 2 . Feature points were chosen on the basis of a facial recognition system, which has a reliability in identifying the arrangement of facial feature points of 97.6%. Before taking the measurements, the accuracy of point plotting was investigated by three examiners. Three arbitrary points were marked on the computer display and plotted twice to obtain the x and y coordinates. Differences in these data were examined for the two plots by means of two-way analysis of variance (ANOVA). The x and y coordinates of all the points showed no significant differences between the first and second procedures for each of the three examiners and between the three examiners, confirming the accuracy of the point plotting. Displacements of the feature points between photographs taken before and after surgery were registered.

Illustrations of the front (A) and lateral face (B), showing 24 feature points. (A) 1L, left eyespot; 1R, right eyespot; 2, apex of the nose; 3L, left alar curvature; 3R, right alar curvature; 4, centre of the lip; 5L, left oral angle; 5R, right oral angle; 6, soft tissue menton; 7L, left soft tissue gonion; 7R, right soft tissue gonion. (B) 1, eyespot; 2, apex of the nose; 3, upper cheek; 4, midpoint of the cheek; 5, inferior cheek; 6, subnasale; 7, upper lip; 8, lower lip; 9, oral angle; 10, soft tissue B-point; 11, soft tissue pogonion; 12, soft tissue menton; 13, soft tissue gonion.

To predict a patient’s postoperative profile, OrthoForecast applies a matching algorithm to the arrangement of the facial feature points on the patient’s presurgical image. From the findings, five candidates are chosen from the stored data of 400 previous patients. The displacement of the feature points between images obtained before and after surgery is reflected in the patient’s facial photograph. The presurgical feature points of the matched candidate are relocated onto the patient’s facial photograph, and the patient’s profile is morphed according to the displacement of the postsurgical feature points.

Assessment of accuracy of the OrthoForecast prediction

The study was approved by the university hospital ethics committee. Fifteen patients with facial asymmetry (asymmetry group), 15 patients with a skeletal class II jaw relationship (class II group), and 15 patients with a skeletal class III jaw relationship (class III group) were enrolled. Frontal and lateral cephalograms and facial photographs were taken before orthognathic surgery and used in the OrthoForecast program to predict the outcomes of orthognathic surgery. For each patient, the software identified five candidates. The candidate with the greatest facial resemblance to the actual image was selected by the first author and used for comparison measurements.

Twenty-four feature points on the predicted and actual images (11 points on the frontal image and 13 points on the lateral image) were digitized by semiautomatic image recognition. Differences between the feature points in the predicted and actual facial images were assessed. For all measurements, the mean and standard deviation (SD) were calculated. Some angles were also compared, as illustrated in Figure 2 . Line 1 was defined as the line connecting the bilateral mandibular angles, line 3 as the line connecting the bilateral eyespots, and line 2 as the line connecting the bilateral oral angles. In the asymmetry group, we measured angle 1 (between lines 1 and 3) and angle 2 (between lines 2 and 3) on the frontal facial photograph to evaluate the facial asymmetry. In the class II and class III groups, we measured angle 3 – the angle formed by lines connecting the eyespot, upper lip, and soft tissue B-point on the lateral facial photograph. Angle 3 was used to assess facial convexity (originally defined as the soft tissue convexity).

Construction of feature points and lines used in the analysis, and angular measurements. (A) 1, Angle between line 1 (connecting the bilateral mandibular angles) and line 3 (connecting the bilateral eyespots); 2, angle between line 2 (connecting the bilateral oral angles) and line 3; (B) 3, angle formed with lines connecting the eyespot point, upper lip point, and soft tissue B-point.

Objective evaluation of the similarity between the OrthoForecast prediction and the actual image

To analyze the accuracy of postsurgical prediction, we used the method described by Giangreco et al. For each case, the images simulated by OrthoForecast and the actual postoperative photographs were compared on a slide. Examiners (10 fellow dental staff members and 27 orthodontists) graded the similarity of the images as ‘very similar’, ‘similar’, ‘alike’, ‘hardly alike’, or ‘different’.

One of the 15 patients in each group was used as a control for calibrating the evaluators. For the evaluation of the control case, the predicted image was replaced by the patient’s actual postsurgical image. In other words, two identical postsurgical images were shown to the evaluators: one as the predicted image and one as the actual postsurgical image. The control was expected to be classified as ‘very similar’ or ‘similar’. If an evaluator graded a control as ‘alike’, ‘hardly alike’, or ‘different’, then that evaluator was disqualified.

Statistical analysis

The Tukey–Kramer test was used to compare the preoperative, predicted, and postoperative reference angles. A P -value of <0.05 was considered statistically significant. The analysis was carried out using SPSS version 17.0 statistical software (SPSS Inc., Chicago, IL, USA).