Automatic preoperative 3d model registration in laparoscopic liver resection
Automatic preoperative 3d model registration in laparoscopic liver resection
复制标题
DOI：
10.1007/s11548-022-02641-z
复制
发表时间：
2022-05-23
影响因子：
3
通讯作者：
Bartoli, A.
中科院分区：
文献类型：
--
作者：
Labrunie, M.;Ribeiro, M.;Bartoli, A.
目的。增强现实（AR）在腹腔镜肝切除术中需要在腹腔镜图像上找到解剖标志和轮廓。它们用于配准术前CT分割得到的三维模型。现有的AR系统依赖于外科医生1)标注地标和轮廓，2)提供初始注册。这些重要的任务需要外科医生的注意，这可能会干扰手术。我们提出了自动解决这两个任务的方法，因此可以自动注册。方法。标志是下脊和镰状韧带。我们通过从68个过程中提取的1415个标记图像的新数据集中训练U-Net来解决1)。我们采用一种新颖的自动粗到精姿态估计方法来求解2)，该方法在迭代鲁棒过程中包含可见性推理。此外，我们建议将脊划分为六个解剖子部分，使其在配准中的标注和使用更加准确。结果。我们的方法检测轮廓的误差相当于一个有经验的外科医生。由于检测不足，其检测脊和韧带的误差较大。尽管如此，我们的方法成功地初始化了3个临床程序的肿瘤靶标注册误差分别为22.4、14.8和7.2 mm。相比之下，手动初始化的误差分别为30.5、15.1和16.3 mm。结论。我们的结果是有希望的，表明我们已经找到了一种合适的方法。
Purpose. Augmented Reality (AR) in Laparoscopic Liver Resection requires anatomical landmarks and the silhouette to be found on the laparoscopic image. They are used to register the preoperative 3D model obtained from CT segmentation. The existing AR systems rely on the surgeon to 1) annotate the landmarks and silhouette and 2) provide an initial registration. These non-trivial tasks require surgeon attention which may perturb the procedure. We propose methods to solve both tasks, hence registration, automatically. Methods. The landmarks are the lower ridge and the falciform ligament. We solve 1) by training a U-Net from a new dataset of 1415 labelled images extracted from 68 procedures. We solve 2) by a novel automatic coarse-to-fine pose estimation method, including visibility-reasoning within an iterative robust process. In addition, we propose to divide the ridge into six anatomical sub-parts, making its annotation and use in registration more accurate. Results. Our method detects the silhouette with an error equivalent to an experienced surgeon. It detects the ridge and ligament with higher errors owing to under-detection. Nonetheless, our method successfully initialises the registration with tumour target registration errors of 22.4, 14.8 and 7.2 mm for 3 clinical procedures. In comparison, the errors from manual initialisation are 30.5, 15.1 and 16.3 mm. Conclusion. Our results are promising, suggesting that we have found an appropriate methodological approach.