Tahmasebi Nazanin, Boulanger Pierre, Noga Michelle, Punithakumar Kumaradevan
Annu Int Conf IEEE Eng Med Biol Soc. 2018 Jul;2018:5906-5909. doi: 10.1109/EMBC.2018.8513607.
Delineation of lung tumor from adjacent tissue from a series of magnetic resonance images (MRI) poses many difficulties due to the image similarities of the region of interest and surrounding area as well as the influence of respiration. However, accurate segmentation of the tumor region is essential in planning a radiation therapy to prevent healthy tissues from receiving excessive radiation. The manual delineation of the entire MRI sequence is tedious, time-consuming and costly. This study investigates how one can perform automatic tracking of tumor boundaries during radiation therapy using convolutional neural networks. We proposed to use a convolutional neural network architecture with modified Dice metric as the cost function. The proposed approach was evaluated over 600 images in comparison to expert manual contours. The proposed method yielded an average Dice score of $0.91 \pm 0.03$ and Hausdorff distance of $2.88 \pm 0.86$ mm. The proposed approach outperformed recent state-of-the-art methods in terms of accuracy in the delineation of the mobile tumors.
从一系列磁共振成像（MRI）中区分肺部肿瘤与相邻组织存在许多困难，这是由于感兴趣区域与周围区域的图像相似性以及呼吸的影响。然而，肿瘤区域的准确分割对于规划放射治疗以防止健康组织受到过度辐射至关重要。手动勾勒整个MRI序列既繁琐、耗时又昂贵。本研究探讨了如何使用卷积神经网络在放射治疗期间自动跟踪肿瘤边界。我们建议使用一种以修改后的骰子系数作为代价函数的卷积神经网络架构。与专家手动轮廓相比，在600张图像上对所提出的方法进行了评估。所提出的方法平均骰子评分为$0.91 \pm 0.03$，豪斯多夫距离为$2.88 \pm 0.86$毫米。在所勾勒的移动肿瘤的准确性方面，所提出的方法优于最近的先进方法。
