This article describes the clinical validation study setup, statistical analysis and results for a deep learning algorithm which detects dental anomalies in intraoral radiographic images, more specifically caries, apical lesions, root canal treatment defects, marginal defects at crown restorations, periodontal bone loss and calculus. The study compares the detection performance of dentists using the deep learning algorithm to the prior performance of these dentists evaluating the images without algorithmic assistance. Calculating the marginal profit and loss of performance from the annotated paired image data allows for a quantification of the hypothesized change in sensitivity and specificity. The statistical significance of these results is extensively proven using both McNemar's test and the binomial hypothesis test. The average sensitivity increases from $60.7\%$ to $85.9\%$, while the average specificity slightly decreases from $94.5\%$ to $92.7\%$. We prove that the increase of the area under the localization ROC curve (AUC) is significant (from $0.60$ to $0.86$ on average), while the average AUC is bounded by the $95\%$ confidence intervals ${[}0.54, 0.65{]}$ and ${[}0.82, 0.90{]}$. When using the deep learning algorithm for diagnostic guidance, the dentist can be $95\%$ confident that the average true population sensitivity is bounded by the range $79.6\%$ to $91.9\%$. The proposed paired data setup and statistical analysis can be used as a blueprint to thoroughly test the effect of a modality change, like a deep learning based detection and/or segmentation, on radiographic images.

Obstructive sleep apnoea (OSA) is a sleep-related breathing disorder, characterized by repetitive airway obstructions, causing disruptive snoring and daytime sleepiness. Maxillomandibular advancement (MMA), which enlarges the upper airway, is a therapeutic surgical approach. However, no study has performed an upper airway sub-region analysis using validated three-dimensional (3D) anatomical and technical limits on cone beam computed tomography (CBCT). Hence, this prospective, observational trial was performed to evaluate 3D volumetric changes in the upper airway according to validated 3D cephalometric landmarks, before and after MMA, for all patients with a polysomnography diagnosis of OSA (apnoea–hypopnoea index (AHI) ≥5). The secondary objective was to evaluate the impact of MMA on the AHI and in a subjective manner with the Epworth Sleepiness Scale (ESS) and OSA questionnaire. Eleven consecutive OSA patients were included. A significant volume increase in the oropharynx (P = 0.002) and hypopharynx (P = 0.02) was observed, in contrast to a non-significant volume reduction in the nasopharynx (P > 0.05). The median AHI (P = 0.03) and ESS score (P = 0.004) decreased significantly as a result of surgery. In conclusion, MMA significantly enlarges the airway volume of the oropharynx and hypopharynx and is associated with improved quality of life.

AIM:The aim of this study was to evaluate the accuracy of 3D soft tissue predictions generated by a computer-aided maxillofacial planning system in patients undergoing orthognathic surgery.METHODS AND MATERIALS:Twenty patients with dentofacial dysmorphosis were treated with orthognathic surgery after a preoperative orthodontic treatment. Fourteen patients had an Angle Class II malocclusion; three patients had an Angle class III malocclusion, and three patients had an Angle Class I malocclusion. Skeletal asymmetry was observed in six patient. The surgeries were planned using the Maxilim software. Computer assisted surgical planning was transferred to the patient by digitally generated splints. The validation procedures were performed in the following steps: (1) Standardized registration of the pre- and postoperative Cone Beam CT volumes; (2) Automated adjustment of the bone-related planning to the actual operative bony displacement; (3) Simulation of soft tissue changes; (4) Calculation of the soft tissue differences between the predicted and the postoperative results by distance mapping.STATISTICAL ANALYSIS AND RESULTS:Eighty four percent of the mapped distances between the predicted and actual postoperative results measured between -2 mm and +2 mm. The mean absolute linear measurements between the predicted and actual postoperative surface was 1.18. Our study shows the overall prediction was dependent on neither the surgical procedures nor the dentofacial deformity type.CONCLUSION:Despite some shortcomings in the prediction of the final position of the lower lip and cheek area, this software promises a clinically acceptable soft tissue prediction for orthognathic surgical procedures.

Introduction: To correct dentofacial deformities, a combination of orthodontic treatment and orthognathic surgery is needed. Prediction software packages are beneficial in treatment planning and achieving improved outcomes, but before using any software, its reliability and reproducibility must be assessed. The aim of this study was to evaluate the accuracy of 2-dimensional Dolphin (version 10; Dolphin Imaging & Management Solutions, Chatsworth, Calif) and 3-dimensional Maxilim (Medicim, Sint-Niklaas, Belgium) softwares in predicting the soft-tissue profiles of patients who had Le Fort I osteotomies. Methods: The presurgical and postsurgical cone-beam computed tomography synthesized lateral cephalograms of 13 patients were collected. Using the Dolphin and Maxilim softwares, the postsurgical profiles were predicted. The positions of the soft-tissue landmarks in profile views were compared with landmarks in the postsurgical photographs. The data were analyzed with the coefficient of reliability and paired-sample t tests. Results: The alpha values of the interclass correlations for each landmark in the x and y planes were between 0.96 and 0.99, except for stomion superior in Maxilim (0.83). The 95% confidence interval and the absolute mean of the error showed that errors in the Dolphin software were greater than those in the Maxilim software, but the differences were not significant (P>0.05), except for soft-tissue A-point. The greatest errors were seen in the chin region. The prediction errors of the nasolabial and mentolabial angles were greater; the prediction error in the Dolphin software was 9 degrees, which has clinical significance. Conclusions: The Dolphin and Maxilim softwares are both appropriate for clinical use. Their inaccuracies in the prediction of the chin region should be considered in complicated surgical planning.

Background and objectives: The purpose is to present and validate a new and innovative ‘surface to Cone-beam CT (CBCT)’ registration method to obtain a 3D virtual augmented model (AUM) of the patient's face appropriate for orthognathic surgery planning.

Purpose State of the art computer aided implant planning procedures typically use a surgical template to transfer the digital 3D planning to the operating room. This surgical template can be generated based on an acrylic copy of the patient's removable prosthesis-the so-called radiographic guide-which is digitized using a CBCT or CT scanner. Since the same accurate fit between the surgical template and the patient as with the radiographic guide and the patient should be ensured, a procedure to accurately digitize this guide is needed.Methods A procedure is created to accurately digitize radiographic guides based on a calibrated segmentation. Therefore, two steps have to be executed. First, during a calibration step a calibration object is CBCT or CT scanned and a calibration algorithm which results in an optimal threshold value is executed. Next the guide is CBCT or CT scanned and a 3D model is created using the obtained optimal threshold. To validate our method, we compared a high accuracy laser scanned copy of the guide with the generated 3D model by creating a distance map between both models.Results The procedure was performed for different CBCT and CT scanners, and the digitization error for each scanner was defined. The 90th percentile error measured on average 0.15 mm, which was always less than the applied voxel size for all CBCT and CT test scans.Conclusions The calibration procedure evaluated in this study solves the known problem of digitizing a radiographic guide based on non-standardized gray value CBCT images. The procedure can easily be executed by a clinician and allows an accurate digitization of a radiographic guide using a CBCT or CT scanner. Starting from this digitization, an accurate surgical template can be made which has a good fit on the patient's remaining teeth and surrounding soft tissues.

In this paper a method is presented for the automated identification of cephalometric anatomical landmarks in craniofacial cone-beam CT images. This method makes use of statistical models, incorporating both local appearance and shape knowledge obtained from training data. Firstly, the local appearance model captures the local intensity pattern around each anatomical landmark in the image. Secondly, the shape model contains a local and a global component. The former improves the flexibility, whereas the latter improves the robustness of the algorithm. Using a leave-one-out approach to the training data, we assess the overall accuracy of the method. The mean and median error values for all landmarks are equal to 2.55 mm and 1.72 mm, respectively.

Three-dimensional cephalometric analysis offers advantages over the traditional two-dimensional analysis, however, also increases the number of cephalometric landmarks(1). Therefore we present an automated method, which can significantly reduce the time needed by the physician for the identification of the landmarks. This method consists of two phases. In the first phase prior knowledge is captured from a set of training data containing images with manually annotated cephalometric landmarks, whereas in the second phase this knowledge is used for the automated identification of cephalometric landmarks. The method is validated using a leave-one-out approach on the training data, resulting in a mean and median error value of 2.31mm and 1.52mm. The mean distance for all landmarks split in the horizontal, vertical and transversal direction was equal to 0.85mm, 0.88mm and 0.99 mm respectively, whereas the median distance was equal to 0.53mm, 0.47mm and 0.55mm, respectively. Therefore, the accuracy of the method presented in this abstract is comparable to the inter-observer variability as reported by Swennen et al(1).

The segmentation of teeth is of great importance for the computer aided planning of dental implants or orthodontic treatment, however, is hampered by the presence of metallic streak artifacts present in the Cone-Beam Computed Tomography (CBCT) images and the lack of contrast between the teeth and the bone. We propose a new method for the segmentation of teeth from CBCT images, using prior knowledge on the shape of each tooth and the image grey values. The prior knowledge on the shape is obtained from a set of training data consisting of CBCT images from 22 patients with corresponding segmented teeth. As an initial validation, the method is applied to the segmentation of 6 CBCT images and compared to the manual segmentation. The average and maximal distance between the manual segmentation and segmentation obtained by the method is equal to 0.6 mm and 4.7 mm, respectively.

Accurate preoperative planning is mandatory for orthognathic surgery. One of the most important aims of this planning process is obtaining good postoperative dental occlusion. Recently, 3D image-based planning systems have been introduced that enable a surgeon to define different osteotomy planes preoperatively and to assess the result of moving different bone fragments in a 3D virtual environment, even for soft tissue simulation of the face. Although the use of these systems is becoming more accepted in orthognathic surgery, few solutions have been proposed for determining optimal occlusion in the 3D planning process. In this study, a 3D virtual occlusion tool is presented that calculates a realistic interaction between upper and lower dentitions. It enables the surgeon to obtain an optimal and physically possible occlusion easily. A validation study, including 11 patient data sets, demonstrates that the differences between manually and virtually defined occlusions are small, therefore the presented system can be used in clinical practice.

A generic model-based segmentation algorithm is presented. Based on a set of training data, consisting of images with corresponding object segmentations, a local appearance and local shape model is build. The object is described by a set of landmarks. For each landmark a local appearance model is build. This model describes the local intensity values in the image around each landmark. The local shape model is constructed by considering the landmarks to be vertices in an undirected graph. The edges represent the relations between neighboring landmarks. By implying the markovianity property on the graph, every landmark is only directly dependent upon its neighboring landmarks, leading to a local shape model. The objective function to be minimized is obtained from a maximum a-posteriori approach. To minimize this objective function, the problem is discretized by considering a finite set of possible candidates for each landmark. In this way the segmentation problem is turned into a labeling problem. Mean field annealing is used to optimize this labeling problem. The, algorithm is validated for the segmentation of teeth from cone beam computed tomography images and for automated cephalometric analysis.

This paper presents an algorithm for non-rigid registration of breast MRI follow-up images that compensates for differences in patient positioning while maintaining real anatomical and pathological changes. The proposed method uses a biomechanical model to constrain the deformation of the internal breast tissue according to elastic continuum mechanics, which is driven by suitable boundary conditions that align the breast surfaces in the images to be registered. Typically, such boundary conditions impose one-to-one surface point correspondences that are established a priori. We investigate alternative, more flexible boundary conditions that do not depend on fixed point correspondences and do not assume completely accurate breast surface segmentation in both images. More specifically, we allow for sliding motion of one surface over the other during deformation as well as for restricted motion perpendicular to the initially segmented boundary surface, based on the internal elastic forces and local intensity information. We evaluate the impact of different boundary conditions on registration quality from the subtraction images obtained for repeated scans of healthy volunteers with intermediate repositioning, using rigid body and free form whole volume intensity based registration for comparison, and also present initial results for actual patient data. Our results demonstrate a drastic reduction in subtraction artifacts using our model, without compromising the biomechanical validity of the deformation field such as unrealistically large local volume changes as with traditional voxel intensity based registration.

Maxillary distraction osteogenesis is indicated in severe angle class III malocclusions, and severe maxillary hypoplasia among some cleft patients and other craniofacial deformities. Twenty patients, aged 8–48 years (mean 17.8 ± 10.5 SD) with maxillary and midfacial hypoplasia were treated. The follow-up period was 13–65 months (mean 35 ± 16.3 SD). A trans-sinusal maxillary distractor was placed intraorally at each side of the maxilla. The distraction vector was predicted using specialist software, and was transferred to the patients using stereolithographic models and individual templates. A (high) Le Fort I type osteotomy was performed. The amount of activation varied from 8 to 17.5 mm (mean 13.1 ± 2.9 SD). Soft and hard tissue formation resulted in complete healing across the distraction gaps. The distractors are almost completely submerged, and can be left in place as long as necessary to avoid relapse. Wit's appraisal was used to measure the stability of the long-term distraction results. Results up to 5 years after distraction showed considerable maxillary advancement with long-term stability. Ongoing growth of the facial skeleton must be considered when distraction osteogenesis is chosen in growing patients.

The state-of-the-art diagnostic tools in oral and maxillofacial surgery and preoperative orthodontic treatment are mainly two-dimensional, and consequently reveal limitations in describing the three-dimensional (3D) structures of a patient's face. New 3D imaging techniques, such as 3D stereophotogrammetry (3D photograph) and cone-beam computed tomography (CBCT), have been introduced. Image fusion, i.e. registration of a 3D photograph upon a CBCT, results in an accurate and photorealistic digital 3D data set of a patient's face. The purpose of this study was to determine the accuracy of three different matching procedures. For 15 individuals the textured skin surface (3D photograph) and untextured skin surface (CBCT) were matched by two observers using three different methods to determine the accuracy of registration. The registration error was computed as the difference (mm) between all points of both surfaces. The registration errors were relatively large at the lateral neck, mouth and around the eyes. After exclusion of artefact regions from the matching process, 90% of the error was within+/-1.5 mm. The remaining error was probably caused by differences in head positioning, different facial expressions and artefacts during image acquisition. In conclusion, the 3D data set provides an accurate and photorealistic digital 3D representation of a patient's face.

In the field of maxillofacial surgery, there is a huge demand from surgeons to be able to pre-operatively predict the new facial outlook after surgery. Besides the big interest for the surgeon during the planning, it is also an essential tool to improve the communication between the surgeon and his patient. In this work, we compare the usage of four different computational strategies to predict this new facial outlook. These four strategies are: a linear Finite Element Model (FEM), a non-linear Finite Element Model (NFEM), a Mass Spring Model (MSM) and a novel Mass Tensor Model (MTM). For true validation of these four models we acquired a data set of 10 patients who underwent maxillofacial surgery, including pre-operative and post-operative CT data. For all patient data we compared in a quantitative validation the predicted facial outlook, obtained with one of the four computational models, with post-operative image data. During this quantitative validation distance measurements between corresponding points of the predicted and the actual post-operative facial skin surface, are quantified and visualised in 3D. Our results show that the MTM and linear FEM predictions achieve the highest accuracy. For these models the average median distance measures only 0.60 mm and even the average 90% percentile stays below 1.5 mm. Furthermore, the MTM turned out to be the fastest model, with an average simulation time of only 10 s. Besides this quantitative validation, a qualitative validation study was carried out by eight maxillofacial surgeons, who scored the visualised predicted facial appearance by means of pre-defined statements. This study confirmed the positive results of the quantitative study, so we can conclude that fast and accurate predictions of the post-operative facial outcome are possible. Therefore, the usage of a maxillofacial soft tissue prediction system is relevant and suitable for daily clinical practice.

Procedure for preoperatively a prediction of postoperative image of at least part of a body, comprising the steps of: - determining a preoperative 3D model of at least part of a body, - acquiring preoperative photograph 2D of said at least part of said body from one or more display positions, - matching said preoperative 3D model with said preoperative photograph 2D, - determining a deformation field from the deformation of said preoperative 3D model in a postoperative 3D model predicted from said at least part of said body, and - deriving a predicted that postoperative 3D model postoperative image, said postoperative predicted image a picture in 2D obtained by deformation of said preoperative photograph 2D using said deformation field.