牙根吸收是正畸治疗严重的并发症，建立准确的牙根吸收测量方法在正畸治疗中极为重要。基于GCA模型研究单侧阻生尖牙患者的上颌骨、尖牙和侧切牙体积特征进行测量，评估上颌阻生尖牙的风险因素，为正畸早期干预提供预警形态学指标。

收集45例单侧上颌尖牙阻生患者的治疗前锥形束CT（CBCT）数据，分为训练集32例（70%）、验证集9例（20%）与测试集4例（10%）。经随机旋转、仿射剪切、水平/垂直翻转和图像平移等增强策略处理后，构建4 500张二维图片优化Mask RCNN-GCA模型。模型性能以Dice相似系数（DSC）、灵敏度、特异性与精确度进行评估模型效果。t检验比较阻生侧与非阻生侧的上颌骨及相邻牙体的体积差异。

优化后的模型在上颌骨分割中取得DSC 93.60%、准确度96.50%；尖牙DSC 92.50%、准确度97.50%。模型自动测量结果显示：阻生侧的上颌骨体积（8 992 ± 1 685）mm³、尖牙体积（8 767 mm³）和侧切牙体积（3 028 mm³）均显著小于非阻生侧（t_{上颌骨体积} = 2.90，P_{上颌骨体积} = 0.01；t_尖牙体积 = 2.42，P_尖牙体积 = 0.005；t_{侧切牙体积} = 3.11，P_{侧切牙体积} = 0.006），性别对上颌骨体积参数无显著影响（t_阻生侧 = 0.77，P_阻生侧 = 0.40；t_非阻生侧 = 0.23，P_非阻生侧 = 0.91）；第一前磨牙体积差异无统计学意义（t = 0.77，P = 0.402）。

优化改进后的Mask RCNN-GCA模型可高效、精准、可靠地实现上颌阻生尖牙相关结构的自动分割与体积量化，阻生侧的上颌骨、尖牙及侧切牙体积更小的特征可作为临床正畸早期干预的重要形态学依据，为降低邻牙牙根吸收等并发症风险提供参考。

Root resorption represents a significant clinical challenge in orthodontic practice, necessitating the development of precise diagnostic techniques. This study employed an advanced Mask RCNN-Gate Context Aggregation (GCA) deep learning algorithm to analyze volumetric characteristics of maxillary structures, specifically focusing on canines and lateral incisors in patients with unilateral impacted canines. The study aimed to identify key morphological predictors for maxillary canine impaction and establish quantitative indicators for early clinical intervention.

CBCT data were obtained from 45 patients diagnosed with unilateral maxillary canine impaction. The dataset was partitioned into a training set (32 cases, 70%) , a validation set (9 cases, 20%) , and a test set (4 cases, 10%) . Following preprocessing, which included data augmentation techniques such as random rotation, affine shear, horizontal and vertical flipping, and image translation, a total of 4 500 two-dimensional images were generated to enhance the training of the Mask RCNN-GCA model. Model performance was assessed using the Dice similarity coefficient (DSC) , sensitivity, and accuracy. Additionally, volumetric differences in maxillary bone and dental structures were compared between the impacted and non-impacted sides with t-test analysis.

The refined model demonstrated outstanding segmentation performance, attaining a DSC of 93.60% and an accuracy of 96.50% in maxillary structures; for canine identification, the values were a DSC of 92.50% and an accuracy of 97.50%. Quantitative analysis revealed statistically significant volumetric reductions (t_{maxillary volume} = 2.90, P_{maxillary volume} = 0.01; t_{canine volume} = 2.42, P_{canine volume} = 0.005; t_{lateral incisor volume} = 3.11, P_{lateral incisor volume} = 0.006) on the impacted side for maxillary bone [mean difference: (8 992 ± 1 685) mm³], canine (8 767 mm³) , and lateral incisor (3 028 mm³) compared to contralateral measurements. Gender-based analysis showed no statistically relevant differences in any measured parameters (t_{impacted sides} = 0.77, P_{impacted sides} = 0.40; t_{non-impacted sides} = 0.23, P_{non-impacted sides} = 0.91) . No significant volumetric variation was observed in first premolars (t = 0.77, P = 0.402) .

The enhanced Mask RCNN-GCA framework provided an effective solution for automated, high-precision volumetric assessment of maxillary impaction cases. The identified volumetric disparities in maxillary bone and associated dentition offered clinically relevant morphological markers for early diagnosis and intervention. These findings contributed valuable quantitative references for minimizing orthodontic complications, particularly adjacent root resorption, through timely therapeutic strategies.