脑肿瘤三维可视化模型自动重建技术的开发及临床应用

Development and clinical application of automatic reconstruction technology for a three-dimensional visualization model of brain tumors

摘要:

目的 研究和开发一种基于头颅多模态核磁共振成像（magnetic resonance imaging，MRI）的影像学数据，自动重建常见脑肿瘤及其周围重要结构的三维可视化模型，并验证其效能及临床适用性。方法 收集常见脑肿瘤头颅多模态核磁共振影像数据，并将其分为训练集、验证集和临床测试集。在训练及验证中，通过3D深度卷积神经网络的算法训练系统自动分割并重建出脑肿瘤及周围结构的能力；在临床测试集中分别用系统及人工手动的方法完成重建，比较本系统自动重建与手动重建之间的重建效率及图像质量。结果 在完成1例肿瘤及周围结构一体化模型重建的时间花费上面，系统用时由人工用时的（5 442±623） s减少至（657±78） s，差异有统计学意义（t=27.530，P=0.000）。且系统重建出的模型与原始影像学图像具有高度一致性（Dice系数为0.92），系统重建出的图像与人工重建的图像在质量方面并无明显差异。结论 基于多模态影像学数据，运用深度学习等算法对脑肿瘤及周围结构进行自动分割及全自动三维可视化重建，具有准确、高效、可靠的优点，对于脑肿瘤的诊断和手术计划的制定具有重要意义。

Abstract:

Objective To study and develop a three-dimensional visualization model that automatically reconstructs common brain tumors and important structures around them based on multimodal magnetic resonance imaging（MRI） data of the head and to validate its performance and clinical applicability.Methods Multimodal MRI data of the head with common brain tumors were collected and divided into training set，validation set，and clinical testing set. In the training and verification sets，the system's ability to automatically segment and reconstruct brain tumors and surrounding structures was trained through the algorithm of 3D depth convolutional neural network. In the clinical testing set，the reconstruction was completed by the system and a human，respectively，and the reconstruction efficiency and image quality were compared between the two methods.Results The time spent on completing the integrated model reconstruction of a tumor and its surrounding structure was significantly reduced from 5442±623 seconds （by a human） to （657±78） seconds（by the system）（t=27.530，P=0.000）. Meanwhile，the model reconstructed by the system had high consistency with the original image（Dice coefficient=0.92），and there was no significant difference in the image quality between the system and human reconstruction.Conclusion The automatic segmentation and fully automatic 3D visualization reconstruction of brain tumors and their surrounding structures using algorithms such as deep learning based on multimodal imaging data are accurate，efficient，and reliable，which is of great significance in the diagnosis of brain tumors and the formulation of surgical plans.