[A pilot study on clinical application of three-dimensional morphological completion of lesioned mandibles assisted by generative adversarial networks].

Objective: To explore the clinical application pathway of the CT generative adversarial networks (CTGANs) algorithm in mandibular reconstruction surgery, aiming to provide a valuable reference for this procedure. Methods: A clinical exploratory study was conducted, 27 patients who visited the Department of Oral and Maxillofacial Surgery, Xiangya Hospital of Central South University between January 2022 and January 2024 and required mandibular reconstruction were selected. The cohort included 16 males and 11 females, with the age of (46.6±11.5) years; among them, 7 cases involved mandibular defects crossing the midline. The CTGANs generator produced 100 images, and the mean squared error (MSE) was calculated for differences between any two generated images. Preoperative cone-beam CT data from 5 patients were used to construct a labeled test database, divided into groups: normal maxilla, normal mandible, diseased mandible, and noise (each group containing 70 cross-sectional images). The CTGANs discriminator was used to evaluate the loss values for each group, and one-way ANOVA and intergroup comparisons were performed. Using the self-developed KuYe multioutcome-option-network generation system (KMG) software, the three-dimensional (3D) completion area of the mandible under cone-beam CT was defined for the 27 patients. The CTGANs algorithm was applied to obtain a reference model for the mandible. Virtual surgery was then performed, utilizing the fibular segment to reconstruct the mandible and design the surgical expectation model. The second-generation combined bone-cutting and prebent reconstruction plate positioning method was used to design and 3D print surgical guides, which were subsequently applied in mandibular reconstruction surgery for the 27 patients. Postoperative cone-beam CT was used to compare the morphology of the reconstructed mandible with the surgical expectation model and the mandibular reference model to assess the three-dimensional deviation. Results: The MSE for the CTGANs generator was 2 411.9±833.6 (95%CI: 2 388.7-2 435.1). No significant difference in loss values was found between the normal mandible and diseased mandible groups (P>0.05), while both groups demonstrated significantly lower loss values than the maxilla and noise groups (P<0.001). All 27 patients successfully obtained mandibular reference models and surgical expectation models. In total, 14 162 negative deviation points and 15 346 positive deviation points were observed when comparing the reconstructed mandible morphology with the surgical expectation model, with mean deviations of -1.32 mm (95%CI:-1.33--1.31 mm) and 1.90 mm (95%CI: 1.04-1.06 mm), respectively. Conclusions: The CTGANs algorithm is capable of generating diverse mandibular reference models that reflect the natural anatomical characteristics of the mandible and closely match individual patient morphology, thereby facilitating the design of surgical expectation models. This method shows promise for application in patients with mandibular defects crossing the midline.