摘要目的:提出一种半监督学习的器官几何轮廓动态生成方法，并基于膀胱容积变化及与子宫的相对位置关系，实现宫颈癌放疗中子宫几何轮廓的精准生成。方法:回顾性收集南方医科大学南方医院放疗科2023年1月至12月收治的60例宫颈癌患者的120组盆腔定位CT图像（膀胱充盈与排空像）。提出一种基于压缩-激励通道注意力机制的条件生成对抗网络（CGAN），通过突出膀胱与子宫间的关键特征联系，学习盆腔器官的相对解剖位置及子宫-膀胱运动相关性，实现基于膀胱状态变化的子宫几何轮廓精确生成。模型生成性能使用Dice相似系数（DSC）、交并比（IoU）和95%豪斯多夫距离（HD95）进行量化评估，并与GAN模型、CGAN模型、Pix2Pix模型3种模型进行比较。通过配对样本 t检验进行两两比较。 结果:本研究提出的SE-CGAN模型在测试集上取得最优性能，DSC为0.83±0.09，IoU为0.71±0.05，HD95为（6.74±1.23）mm，较GAN、CGAN、Pix2Pix模型DSC分别提升7.5%、4.9%、3.6%（均 P<0.001），HD95均值降低32.9%~45.3%。SE-CGAN模型与其他3种模型的生成性能指标差异均有统计学意义（均 P<0.001），而CGAN与Pix2Pix模型的生成性能指标差异没有统计学意义。可视化结果进一步显示，GAN模型生成的子宫轮廓与真实形态偏差较大，边缘模糊；CGAN和Pix2Pix模型生成的轮廓虽具有一定重叠度，但边界还原精度不足；而SE-CGAN模型生成的子宫轮廓与真实轮廓高度吻合，边界细节清晰，生成质量最佳。 结论:本研究提出的生成对抗网络方法，通过建立膀胱状态对子宫几何轮廓的动态调节作用的生成机制，以任意一次定位CT的膀胱几何轮廓来精确生成子宫几何轮廓，有效解决了盆腔器官运动带来的放疗靶区不确定性问题。

abstractsObjective:To propose a semi-supervised learning method for dynamic generation of organ geometric contours, leveraging bladder volume variations and its relative position to the uterus to accurately generate uterine contours in cervical cancer radiotherapy.Methods:A total of 120 sets of pelvic planning CT images (including both full and empty bladder scans) from 60 patients with cervical cancer treated at the Department of Radiation Oncology, Nanfang Hospital of Southern Medical University between January and December 2023 were retrospectively collected. A conditional generative adversarial network (CGAN) based on a squeeze-and-excitation channel attention mechanism was proposed to accurately generate uterine geometric contours under varying bladder filling states. By emphasizing the critical spatial relationships between the bladder and uterus, the model learned the relative anatomical positions of pelvic organs and their motion correlations. The generative performance was quantitatively evaluated using the average Dice similarity coefficient (DSC), intersection over union (IoU), and the 95 th percentile Hausdorff distance (HD95), and was compared with GAN model, CGAN model, and Pix2Pix model. Pairwise comparisons were perfomed by paired-sample t-test. Results:The proposed SE-CGAN model achieved the best performance on the test set, with DSC of 0.83±0.09, IoU of 0.71±0.05, HD95 of (6.74±1.23) mm, improving DSC by 7.5%, 4.9%, and 3.6% compared to the GAN, CGAN, and Pix2Pix models, respectively (all P<0.001), and reducing the mean HD95 by 32.9%-45.3%. Statistical analysis revealed significant differences between SE-CGAN model and the other 3 baseline models, whereas no significant difference was observed between CGAN model and Pix2Pix model. The visualization results further demonstrated that the GAN model produced uterine contours deviated greatly from the real shape, and the edge was fuzzy; CGAN and Pix2Pix model achieved better overlap but lacked of precision in boundary reconstruction. In contrast, the contours generated by SE-CGAN model closely matched the ground truth with clearly defined edges, indicating superior reconstruction accuracy. Conclusions:In this study, we propose a generative adversarial network method that establishes a dynamic modulation mechanism by which the bladder state influences the uterine geometric contour, enabling accurate generation of the uterine contours from the bladder contours of any given localization CT scan. This approach effectively addresses the uncertainty in radiotherapy target delineation caused by pelvic organ motion.