| Craniomaxillofacial defects and deformities are a kind of diseases that seriously endanger human physical and mental health.They not only cause damage to the patient’s facial structure and function,but also seriously affect the patient’s normal production,life and social life,and bring significant psychological harm to the patient.Despite the increasing development of science and technology,it is still challenging to improve the surgical treatment effect of such diseases and meet the expected appearance requirements of patients.At present,computer-aided surgical design plays an important role in the treatment of craniomaxillofacial defects and deformities.Compared with the past,it can make surgical planning more accurately and visually.However,in the process of surgical planning,it is still dominated by the experience of doctors,mainly aiming at the design of bones;Patients are concerned about the change of face shape and appearance,and more about the soft tissue results.The two have different concerns,resulting in contradictions,resulting in poor communication between doctors and patients,resulting in unsatisfactory treatment results for some patients.Although some software have the function of predicting the postoperative type in the process of surgical planning,it only provides a face reference for surgical scheme planning and cannot meet the expected appearance requirements of patients.In order to unify the concerns of doctors and patients to the soft tissue appearance in the process of surgical planning,improve the role of patients’ appearance needs in surgical planning,and achieve the satisfactory treatment effect of patients,this paper proposes a goal-oriented reverse surgical planning method for craniomaxillofacial malformations,That is to use the patient’s expected appearance as a guide to plan the reverse bone surgery plan.Different from the current surgical scheme design directly for bones,this method first edits the expected appearance of the patient in three dimensions on the software before designing the operation,and then uses the neural network model constructed in this paper to predict the position and shape of the underlying bone corresponding to the expected appearance,which is used to guide the planning of the surgical scheme.This reverse surgical planning method avoids the excessive dependence on doctors’ subjective experience in previous surgical design,improves the role of patients’ appearance needs in surgical design,and can better improve patients’ treatment satisfaction.The premise of realizing the goal-oriented reverse surgery planning method is to establish a neural network model to predict the underlying bone with the appearance of soft tissue.Neural network model is an algorithm specially used in the field of deep learning to analyze data,learn and mine its internal laws and relationships from a large number of data,and then judge and predict new data or events in the real world.In this paper,the normal population facial CT soft and hard tissue model database is established to prepare data for the construction of neural network model.Through the development of point cloud sampling program EFPs software,the facial soft and hard tissue point cloud sampling in the database is completed.The sampled point cloud data is used for training and the neural network model FPNet is constructed.Then,the prediction accuracy of the network model is evaluated through experiments,and a preliminary attempt is made to be used in clinical reverse surgery scheme planning.At the same time,aiming at the complex problem of surgical planning of large-scale craniomaxillofacial bone defects,using the craniomaxillofacial bone point cloud data in the database,this paper constructs a GRNet network model for automatic point cloud completion and reconstruction of craniomaxillofacial bone defects,evaluates its completion accuracy and applies it to the planning of clinical surgery.The full text will be expanded from the following four parts:The first part is to establish the craniomaxillofacial soft and hard tissue model database and develop the point cloud sampling software EFPs.1027 cases of craniomaxillofacial CT data of normal young people were collected.According to the inclusion and exclusion criteria,613 cases(271 males and 342 females,with an average age of 28.9 years)were finally obtained for database construction.In mimics21 Establish the three-dimensional model of the soft tissue,including the three-dimensional coordinate system of the soft tissue and the three-dimensional reconstruction of the soft tissue,and unify the three-dimensional coordinate system of the soft tissue,including the three-dimensional reconstruction of the soft tissue.The three-dimensional soft and hard tissue models of each structured CT data are stored independently,and the database is constructed by naming the file with digital serial number.In order to complete the point cloud sampling of facial soft and hard tissue three-dimensional model data,this paper assumes that a series of small balls with different radii intersect with the soft and hard tissue surface,with a fixed angle on the intersection line θ Based on the idea of point cloud sampling,an orderly point cloud sampling software for craniomaxillofacial soft and hard tissue matching-EFPs software is developed.The software can sample the paired ordered point cloud of soft and hard tissues in different regions of the face,and transform the disordered three-dimensional point cloud of facial soft and hard tissues into the paired point cloud data of ordered soft and hard tissues,which makes data preparation for the later network model training.In the second part,a convolutional neural network model FPNet for bone prediction in different regions of the face is constructed and its prediction accuracy is evaluated.Referring to the "V" structure of UNET network model,this paper constructs a "W" convolutional neural network model FPNet with two "V" structures in series.The point cloud sampling software EFPs constructed in the first part is used to sample the point cloud of different regions(chin region,mandibular angle region and zygomatic region)of the facial soft and hard tissue model in the database.Among the 613 paired point cloud data of soft and hard tissues sampled,500 point cloud data are randomly selected as the training set to train the FPNet model,and 50 data are used as the verification set to verify the model.The prediction accuracy of FPNet model is evaluated with 63 point cloud data in the test set and 10 sample data outside the new database.The evaluation results show that the prediction accuracy of soft tissue point cloud in the chin,mandibular angle and zygomatic region of FPNet model constructed in this paper is 81.5%,90.5% and 93.8%respectively.In the third part,a reverse surgical planning method guided by the expected appearance goal is proposed to verify the accuracy of surgical planning,and a preliminary clinical application is carried out.The reverse operation planning method proposed in this paper is to edit the soft tissue model reconstructed from the patient’s CT data on Geomagic wrap software before operation,so that the edited face effect can meet the patient’s expected requirements and be regarded as the expected face.Then,the soft tissue point clouds of chin,mandibular angle and zygomatic area with expected appearance are predicted in FPNet model to obtain the bone point cloud of its corresponding area.The bone point cloud is reconstructed as the target bone to guide the surgical planning.In order to verify the accuracy of reverse surgery planning,10 patients who have completed facial contour plastic surgery were selected,and their postoperative type was taken as the expected appearance.The corresponding bone morphology was predicted by FPNet model,and the predicted bone was taken as the target bone for surgery planning on the preoperative CT bone data of patients.The accuracy of the planning scheme was evaluated by comparing the planning scheme with the actual CT bone of the patient.After evaluation,the accuracy of this surgical planning method in Chin,mandibular angle and zygomatic area was 78.3%,87.6% and 90.2% respectively.The preliminary clinical application of this method was carried out,and 5 cases of facial contour plastic surgery were planned,and satisfactory surgical results were observed.In the fourth part,a neural network model GRNet model for automatic point cloud completion of craniomaxillofacial bone defects is constructed,the prediction accuracy of the model is evaluated,and the preliminary clinical application is carried out.In this part,the GRNet network model and Poin Tr network model are first built,and then each craniomaxillofacial bone sample in the database is randomly cut 10 times to generate 10 corresponding bone defect samples.The data set is enhanced 10 times,forming a total of6130 pairs of defect complete paired point cloud data for network model training.The test set evaluation of the two trained models and the preclinical evaluation of 10 different types of defect data outside the database were carried out respectively.The results show that the accuracy of 2mm error threshold of GRNet model evaluated by the test set is 83%,which is better than 63.9% of Poin Tr model.GRNet model was initially used in clinical planning of 3 cases of craniomaxillofacial large-scale cross midline defect deformity,and the operation effect was satisfactory.The point cloud completion of GRNet model was used in orthognathic surgery planning,and the relocation of abnormal jaw in the process of orthognathic surgery planning was regarded as the point cloud automatic completion of maxillary and mandibular defects.Three cases of surgery planning were carried out with GRNet model,and the operation effect was satisfactory.To sum up,by establishing the craniomaxillofacial CT database of normal young people and developing the point cloud sampling software EFPs,this paper constructs the FPNet neural network model based on point cloud deep learning for the first time,and realizes the purpose of predicting the underlying bone with the facial soft tissue appearance;A reverse surgical planning method for craniomaxillofacial malformations based on FPNet model and guided by the goal of expected appearance is proposed for the first time.After evaluation,its accuracy is within the clinically acceptable range,and it is preliminarily used in the surgical planning of craniomaxillofacial malformations;The GRNet network model for automatic point cloud reconstruction of large-scale craniomaxillofacial bone defects is constructed for the first time.After accuracy evaluation,it is preliminarily used in clinical craniomaxillofacial bone defects and orthognathic surgery planning,and the clinical effect is satisfactory. |
PDF Full Text Request