【摘要】 输尿管内窥镜手术是目前针对肾结石的主流治疗方案，其外形细长、镜体柔软，能够灵活穿越人体自然腔道的内径狭窄的尿道和输尿管，检查视野范围更广，使医生能够更好地观察到病变区域．但一般的输尿管内窥镜仅配备单目摄像头进行配合手术操作，无法借助额外设备获取数据导致了其图像信息的匮乏；同时，相比于胃肠、鼻镜等手术场景，本研究的肾内场景在不具备公开数据集的同时，图像质量参差不齐，表面纹理细节不足，孔洞区域褶皱少，受模糊反光等干扰大，都易使深度估计受到影响针对以上问题，提出了一种基于改进风格迁移模型的深度估计方法．该方法首先根据术前CT图像重建肾脏内部腔道模型并提取中心路径，将虚拟内窥镜的摄像头设置在插值后的路径点上，构建了虚拟内窥镜漫游图像与深度估计图像一一对应的数据集，并基于此数据集训练了一个深度估计模型；随后，使用添加高效通道注意力(ECA)模块的改进风格迁移模型，将真实内窥镜图像域迁移至虚拟内窥镜图像域；最后，再将经由风格迁移产生的虚拟内窥镜图像送入上述训练得来的深度估计模型中，最终实现真实内窥镜图像的深度估计．所提方法的可行性及有效性在输尿管钬激光碎石术的图像中得到验证．更多还原

【Abstract】 Using flexible ureteroscopes is the mainstream treatment method for nephrolithiasis owing to their slender and flexible structure,allowing them to navigate through the narrow natural passages of the human body,such as the urethra and ureters. The elongated shape of these ureteroscopes provides a broad visual inspection range,enabling physicians to effectively observe affected areas. However,conventional ureteroscopes typically feature a singlecamera system for surgical operations,leading to limited image data. This limitation results in insufficient information as additional imaging devices cannot be leveraged. Unlike surgical scenarios such as gastrointestinal or nasal endoscopy,the kidney’s internal environment lacks publicly available datasets. The images captured during surgery exhibit varying quality,with insufficient surface texture details,fewer folds in cavity areas,and susceptibility to interference such as blurriness and reflections. These challenges can significantly impact depth estimation. Herein,a depth estimation method is proposed,which involves leveraging an improved style-transfer model to address the aforementioned issues. The method begins by reconstructing the internal cavity model of the kidney based on preoperative computed tomography images and extracting the central path. The involved virtual endoscope’s camera is then positioned at interpolated path points,creating a dataset that correlates virtual endoscope roaming images with depth estimation images. A depth estimation model is trained using this dataset. Subsequently,an improved style-transfer model incorporating an efficient channel attention module is employed to transfer the real endoscopic image domain to the virtual endoscopic image domain. Finally,the virtual endoscopic images generated through style-transfer are input into the previously trained depth estimation model,achieving depth estimation of real endoscopic images. The feasibility and effectiveness of the proposed method are validated using images obtained from ureteroscopic holmium laser lithotripsy procedures.更多还原