甲状腺结节超声图像分割算法研究

Ultrasonic image segmentation algorithm for thyroid nodules

超声甲状腺结节分割是发现与识别甲状腺良恶性肿瘤的关键技术之一。针对模糊聚类法无法准确分割超声图像甲状腺结节边缘,而局部拟合(RSF)模型法对手动初始化轮廓敏感的问题,提出一种融合空间约束模糊C均值聚类和局部拟合RSF模型的分割结节方法。用空间约束模糊C均值聚类法(SKFCM)对图像进行聚类并二值化聚类结果作为RSF模型法初始轮廓,克服了RSF模型法对初始轮廓敏感问题,水平集演化参数也将通过聚类结果自动给出,不再需要人为设定。同时改进了RSF模型法拟合项,并利用高斯正则化规则RSF模型水平集,提高了RSF模型演化效率,缩短了收敛时间。仿真实验结果表明,提出的甲状腺结节超声图像分割方法能够快速准确地分割出结节区域。

Ultrasonic thyroid nodule segmentation is one of the key techniques to discover and identify the benign and malignant thyroid tumors. Fuzzy clustering nlethod can not accurately segment the edge of thyroid nodules in ultrasound images and Regin- Scalable Fitting（RSF） model is sensitive to manual initialization. This paper presents a hybrid segmentation method that utilizes both the space constraint of the Gaussian kernel induced by fuzzy C-means clustering and region scalable fitting （RSF） model. The clustering algorithm is used to pre-segment the image and hinarise the results of segmentation, which solves the initial contour of RSF. The controlling parameters of the level set evolution are estimated by the results of pre-segment and not by manual setting pa- rameters. The enerb7 item of Region scalable fitting model has been improved and Gaussian filtering is utilized to regularize RSF level set function, which improves segmentation efficiency and reduces the convergence time. Simulation experiment results show that the proposed model ensures an improvement in segmentation accuracy.