目的 探讨基于U型卷积神经网络(U-shaped convolutional neural network, U-net)建立的前列腺磁共振图像自动化分割和重建3D模型对腹腔镜前列腺癌根治术进行术中认知导航的效果。方法 应用含有人工注释的共5 000张前列腺癌磁共振影像训练集,训练U-net,构建了一套以临床需求为导向,稳定高效的全卷积神经网络算法模型,对前列腺磁共振图像进行区域化、多结构和精细自动化分割,并将分割数据使用医学影像处理交互平台(Medical Image Interaction Tool Kit,MITK)自动重建,以STL格式输出建模信息,应用平板电脑在术中展示前列腺模型,进行认知导航。结果 基于201例前列腺癌患者的磁共振图像训练样本,在经典U-net基础上通过适应性改良,建立了一套结构简单、性能优秀的U-net,可以实现对前列腺、肿瘤、精囊腺、直肠等重要结构的单独分割,并进行三维可视化,直观地显示手术关键部位的结构关系和肿瘤侵犯程度。术中通过平板电脑同步展示3D模型,成功进行认知导航。结论 通过改良的U-net可以自动化完成前列腺磁共振图像的结构化分割,通过重建局部解剖部位的3D模型用于术中认知融合导航,可以达到肿瘤可视化、降低手术切缘阳性率、提高手术效果的作用。

Objective: To investigate the efficacy of intraoperative cognitive navigation on laparoscopic radical prostatectomy using 3D prostatic models created by U-shaped convolutional neural network (U-net) and reconstructed through Medical Image Interaction Tool Kit (MITK) platform. Methods: A total of 5 000 pieces of prostate cancer magnetic resonance (MR) imaging discovery sets with manual annotations were used to train a modified U-net, and a set of clinically demand-oriented, stable and efficient full convolutional neural network algorithm was constructed. The MR images were cropped and segmented automatically by using modified U-net, and the segmentation data were automatically reconstructed using MITK platform according to our own protocols. The modeling data were output as STL format, and the prostate models were simultaneously displayed on an android tablet during the operation to help achieving cognitive navigation. Results: Based on original U-net architecture, we established a modified U-net from a 201-case MR imaging training set. The network performance was tested and compared with human segmentations and other segmentation networks by using one certain testing data set. Auto segmentation of multi-structures (such as prostate, prostate tumors, seminal vesicles, rectus, neurovascular bundles and dorsal venous complex) were successfully achieved. Secondary automatic 3D reconstruction had been carried out through MITK platform. During the surgery, 3D models of prostatic area were simultaneously displayed on an android tablet, and the cognitive navigation was successfully achieved. Intra-operation organ visualization demonstrated the structural relationships among the key structures in great detail and the degree of tumor invasion was visualized directly. Conclusion: The modified U-net was able to achieve automatic segmentations of important structures of prostate area. Secondary 3D model reconstruction and demonstration could provide intraoperative visualization of vital structures of prostate area, which could help achieve cognitive fusion navigation for surgeons. The application of these techniques could finally reduce positive surgical margin rates, and may improve the efficacy and oncological outcomes of laparoscopic prostatectomy.