Bui Nhat Linh, Ong Sim Heng, Foong Kelvin Weng Chiong
Department of Electrical and Computer Engineering, National University of Singapore, Singapore, Singapore,
Int J Comput Assist Radiol Surg. 2015 Aug;10(8):1269-77. doi: 10.1007/s11548-014-1134-5. Epub 2014 Dec 12.
A patient-specific upper airway model is important for clinical, education, and research applications. Cone-beam computed tomography (CBCT) is used for imaging the upper airway but automatic segmentation is limited by noise and the complex anatomy. A multi-step level set segmentation scheme was developed for CBCT volumetric head scans to create a 3D model of the nasal cavity and paranasal sinuses.
Gaussian mixture model thresholding and morphological operators are first employed to automatically locate the region of interest and to initialize the active contour. Second, the active contour driven by the Kullback-Leibler (K-L) divergence energy in a level set framework to segment the upper airway. The K-L divergence asymmetry is used to directly minimize the K-L divergence energy on the probability density function of the image intensity. Finally, to refine the segmentation result, an anisotropic localized active contour is employed which defines the local area based on shape prior information. The method was tested on ten CBCT data sets. The results were evaluated by the Dice coefficient, the volumetric overlap error (VOE), precision, recall, and F-score and compared with expert manual segmentation and existing methods.
The nasal cavity and paranasal sinuses were segmented in CBCT images with a median accuracy of 95.72 % [93.82-96.72 interquartile range] by Dice, 8.73 % [6.79-12.20] by VOE, 94.69 % [93.80-94.97] by precision, 97.73 % [92.70-98.79] by recall, and 95.72 % [93.82-96.69] by F-score.
Automated CBCT segmentation of the airway and paranasal sinuses was highly accurate in a test sample of clinical scans. The method may be useful in a variety of clinical, education, and research applications.
特定患者的上气道模型对于临床、教育和研究应用至关重要。锥形束计算机断层扫描（CBCT）用于对上气道进行成像，但自动分割受到噪声和复杂解剖结构的限制。针对CBCT头部容积扫描开发了一种多步水平集分割方案，以创建鼻腔和鼻窦的三维模型。
首先采用高斯混合模型阈值化和形态学算子自动定位感兴趣区域并初始化活动轮廓。其次，在水平集框架中由库尔贝克-莱布勒（K-L）散度能量驱动活动轮廓来分割上气道。K-L散度不对称性用于在图像强度的概率密度函数上直接最小化K-L散度能量。最后，为了细化分割结果，采用基于形状先验信息定义局部区域的各向异性局部活动轮廓。该方法在十个CBCT数据集上进行了测试。通过骰子系数、体积重叠误差（VOE）、精度、召回率和F分数对结果进行评估，并与专家手动分割和现有方法进行比较。
在CBCT图像中，鼻腔和鼻窦的分割中位数准确率为：骰子系数95.72%[四分位间距93.82 - 96.72]，VOE为8.73%[6.79 - 12.20]，精度为94.69%[93.80 - 94.97]，召回率为97.73%[92.70 - 98.79]，F分数为95.72%[93.82 - 96.69]。
在临床扫描测试样本中，气道和鼻窦的CBCT自动分割具有很高的准确性。该方法可能在各种临床、教育和研究应用中有用。
