基于流向特征熵和测地线距离的粘连血管型肺结节聚类分割

肺癌计算机辅助诊断(lung cancer CAD)是辅助医生定量判别结节良恶性的新技术.倍增率是临床上判断结节良恶性的指标,而精确地分割结节又是计算倍增率的前提.因为结节和血管的CT值相近,所以难以正确分割粘连血管型结节.血管里充满着流向同一方向的血液,使得大部分血管像素梯度的法向量(流向特征)都指向同一方向,流向特征熵值小;而结节上的像素梯度法向量方向杂乱无章,流向特征熵值大.大部分血管像素到结节中心的测地线距离比结节像素到结节中心的距离大.基于上述血管与结节差异,文中提出了一种基于流向特征熵和测地线距离的K均值聚类算法来分割结节.针对132个临床CT影像的肺结节(104个孤立型和28个粘连血管型),12个LIDC集合1的肺结节(4个孤立型和8个粘连血管型)和182个LIDC集合2的肺结节(25个孤立型肺结节和157个粘连血管型肺结节),评估实验结果和影像科医生手工绘制的金标准相比较,孤立型肺结节分割正确率分别为100/104(96.2%),4/4(100%),24/25(96.0%),粘连血管型分割正确率分别为26/28(92.9%),7/8(87.5%)和149/157(94.9%).实验表明,该方法能在短的时间内正确地分割孤立型结节和粘连血管型结节且具有好的鲁棒性.

Juxta-vascular nodule segmentation based on the flowing entropy and geodesic distance feature

Computed aided diagnosis （CAD） of lung CT is a new quantitative analysis imaging technique to distinguish malignant nodules from benign ones. Nodule growth rate is a key indicator for judgment of benign or malignant nodule. Accurate nodule segmentation is the premise condition of calculating nodule growth rate. However, it is difficult to segment Juxta-vascular nodules, due to similar gray levels between nodule and vessel. To distinguish the nodule region from the adjacent vessel region, a flowing direction feature, referred to as the direction of normal vector for a pixel, is introduced. Since all blood has the same flowing direction through a vessel, the normal vectors of pixels in the vessel region typically point to similar orientations while the directions of those in the nodule region can be viewed as disorganized. So the entropy value of flowing direction features in a neighbor region for a vessel pixel is bigger than that for a nodule pixel. And vessel pixels typically have a larger geodesic distance to the nodule center than nodule pixels. Based on k-Means clustering method, the flowing entropy feature, combined with geodesic distance feature, is proposed to solve the segmentation problem of the vessel attachment nodule. The validation of the proposed segmentation algorithm was carried out on Juxta-vascular nodules （104 solid nodule and 28 Juxta-vascular nodule）, identified in the Chinalung-CT screening trail and on the lung image database consortium （LIDC） dataset. Among them, there are 12 nodules in the first LIDC database （4 solid nodules and 8 Juxta-vaseular nodules） and 182 nodules in the second LIDC database （25 solid nodules and 157 Juxta-vascular nodules）. Comparison is done between the gold standard and experimental results. The correct segmentations on solid nodule are 100/104（96.2%）, 4/4（100%） and 24/25（96.0%）, respectively, while the correct segmentations on Juxta-vascular nodule are 26/28（92.9%）,7/8（87.5%） and 149/157（94.9%）, respectively, showing that the proposed method has low time complexity and high accurate rate.