[Virtual reconstruction and clinical verification of maxillary defect based on deep learning].

Objective: To construct a virtual reconstruction method including midspan maxillary defects and provide clinical reference by training a generative adversarial network (GAN) model. Methods: The CT data of middle-aged Han patients with oral diseases who visited the Department of Radiology, West China Hospital of Stomatology, Sichuan University from June 2015 to June 2022 were collected, where the CT data of 100 healthy maxilla and 15 maxillary defects (5 simple unilateral defects, 5 unilateral defects involving zygomatic bone, 5 midspan defects) were selected. Mimics was used to create spherical phantom and simulate bone defects around the healthy maxillas, including simple unilateral defects, unilateral defects involving zygomatic bone and midspan defects. The original image was set as the correct reference for the reconstruction: artificial defects paired with the correct reference were divided into training set (n=70), validation set (n=20) and test set (n=10), where the first two were used to train the GAN model, and the test set was used to evaluate the GAN performance. Data from 15 clinical defects were imported into the trained GAN model for reconstruction, with mirroring and GAN-based virtual reconstruction for unilateral clinical defects, and only the latter method was adopted for midspan defects. The reconstruction results were divided into mirror reconstruction group (n=10), unilateral defect GAN reconstruction group (n=10) and midspan defect GAN reconstruction group (n=5). The test set, mirror reconstruction group, and unilateral defect GAN reconstruction group were quantitatively evaluated, whose quantitative indicators were Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95), and the group results were subjected to one-way ANOVA and Tukey test. The test set, mirror reconstruction group, unilateral defect GAN reconstruction group and midspan defect GAN reconstruction group were qualitatively scored, and Kruskal-Wallis test and Bonferroni correction were used for the total score of each group. Results: The total differences in the test set, mirror reconstruction group, unilateral defect GAN reconstruction group DCS (0.891±0.049, 0.721±0.047, 0.778±0.057, respectively) and HD95 [(3.58±1.51), (5.19±1.38), (4.51±1.10) mm, respectively] were statistically significant (F=28.08, P<0.001; F=3.62, P=0.041); among them, the test set DSC was significantly larger than the mirror reconstruction group (P<0.05), and the test set HD95 was significantly less than the mirror reconstruction group (P<0.05). Overall difference in qualitative total scores [8 (1), 6 (2), 6 (2), and 4 (2) points, respectively] in the test set, mirror reconstruction group, unilateral defect GAN reconstruction group, and midspan defect GAN reconstruction group were statistical significance (H=18.13, P<0.001); pairwise comparison showed that the total score of the test set was significantly higher than that of the mirror reconstruction group (P<0.05). Conclusions: The virtual reconstruction method based on GAN proposed in this study has better virtual reconstruction effect of unilateral defect than mirror technique, and can also realize virtual reconstruction of maxillary midspan defect.目的： 通过训练生成对抗网络（generative adversarial network，GAN）模型，构建一种包括跨中线上颌骨缺损的虚拟重建方法，以期为临床提供参考。 方法： 收集2015年6月至2022年6月于四川大学华西口腔医院影像科就诊的汉族成年口腔疾病患者CT资料，选择100例健康上颌骨及15例上颌骨缺损（5例单纯单侧缺损、5例单侧缺损并累及颧骨、5例跨中线缺损）CT数据。应用Mimics软件在健康上颌骨数据及其附近区域创建球型模体并模拟上颌骨缺损，分别为单纯单侧缺损、单侧缺损并累及颧骨、跨中线缺损，以原始图像为虚拟重建的正确参照；人工缺损与正确参照配对后分为训练集（70例）、验证集（20例）以及测试集（10例），前两者用于训练GAN模型，测试集用于评估GAN性能。15例上颌骨缺损CT数据导入训练完成后的GAN模型中进行虚拟重建，对单侧缺损分别采取镜像和基于GAN的方式进行虚拟重建，而对跨中线缺损仅采取基于GAN的方式进行虚拟重建，重建结果分为镜像重建组（10例）、单侧缺损GAN重建组（10例）和跨中线缺损GAN重建组（5例）。对测试集、镜像重建组、单侧缺损GAN重建组进行定量评价，定量指标为Dice相似性系数（Dice similarity coefficient，DSC）和95%豪斯道夫距离（95% Hausdorff distance，HD95），对各组结果进行单因素方差分析和Tukey检验。对测试集、镜像重建组、单侧缺损GAN重建组和跨中线缺损GAN重建组进行定性评分，对各组总分进行Kruskal-Wallis检验和事后检验（Bonferroni校正法）。 结果： 测试集、镜像重建组、单侧缺损GAN重建组DCS（分别为0.891±0.049、0.721±0.047、0.778±0.057）和HD95［分别为（3.58±1.51）、（5.19±1.38）、（4.51±1.10）mm］的总体差异均有统计学意义（F=28.08，P<0.001；F=3.62，P=0.041）；其中，测试集DSC显著大于镜像重建组（P<0.05），测试集HD95显著小于镜像重建组（P<0.05）。测试集、镜像重建组、单侧缺损GAN重建组、跨中线缺损GAN重建组定性总分［分别为8（1）、6（2）、6（2）和4（2）分］的总体差异有统计学意义（H=18.13，P<0.001）；两两比较显示，测试集总分显著高于镜像重建组（P<0.05）。 结论： 本项研究提出的基于GAN的虚拟重建方法，其单侧缺损虚拟重建效果优于镜像技术，亦可实现跨中线上颌骨缺损的虚拟重建。.