基于交叉熵和GVF-Snake的子宫肌瘤高强度聚焦超声图像自动分割算法

Automatic segmentation method of uterine tumor high intensity focused ultrasound image based on cross entropy and GVF-Snake

目的:高强度聚焦超声(HIFU)广泛应用于肿瘤无创治疗,目前手术中多用超声成像技术进行导航,但由于HIFU图像对比度低,信噪比低以及目标边界模糊等缺点,HIFU图像的目标识别与分割是重点也是难点,所以需要提出能够自动快速获得HIFU图像肿瘤轮廓的分割方法。方法:GVF-Snake模型算法能够有效地利用超声图像的局部与整体信息实现边界的准确定位,非常适用于HIFU图像分割,但是作为参数活动轮廓模型,GVF-Snake对初始轮廓的依赖性较强,而且通常采用手画初始轮廓,增加了人为因素对试验结果的干预。针对GVF-Snake的相关特性,本文提出用二维最小交叉熵阈值分割法来提取初始轮廓。交叉熵是度量两个统计概率分布之间信息量差异的物理量,分别表征分割前后图像中像素特征向量的概率分布,当原始图像和分割图像之间的信息量差异最小时,便得到最优阈值。使用二维最小交叉熵算法求得初始轮廓后,进而使用GVF-Snake模型收敛,得到最终结果。结果:该算法对HIFU图像中子宫肌瘤的识别与分割具有较为理想的效果,统计结果显示灵敏度平均值达到87.56%,标准化的Hausdorff距离指数平均值达到4.95%,整体算法的运行时间平均值达到2.16 s。结论:该分割算法通过GVF-Snake自动生成初始轮廓,避免了人为干预,整体分割算法快速精准,取得了较好的实验结果,为其在HIFU设备的应用奠定了基础。

Objective High intensity focused ultrasound （HIFU） is widely applied in noninvasive tumor treatment field, and ultrasonic imaging technology is always applied for navigation in the current operation. The target identification and segmentation of HIFU image is important and diffcult for the low contrast, low signal-to-noise ratio, and fuzzy target boundaries in HIFU image. A segmentation method which can automatically and rapidly obtain the tumor contour in HIFU image is necessary. Methods GVF-Snake model algorithm, which could effectively use the partial and the whole information of ultrasonic image to realize the accurate positioning of boundaries, was appropriately used in the HIFU image segmentation. As a parameter active contour model, GVF-Snake had deep dependence on the initial contour. The initial contour was delineated manually, which increased the interference of human factors on the test results. For the relative features of GVF- Snake, two- dimensional （2D） minimum cross entropy thresholding method was proposed to extract the initial contour. Cross entropy was used to measure the information difference between two probability distributions, separately representing the characteristic vector probability distribution of pixels in images before and after segmentation. When the information differences between the images before and after segmentation were minimal, the optimal threshold value was obtained. After that the initial contour was obtairled by using 2D minimum cross entropy thresholding method, the GVF-Snake model was used to converge the contour, obtaining the final result. Results The performance of the proposed method was satisfactory in recognizing and segmenting HIFU images of uterine tumor. The statistical results showed that the average sensitivity index reached 87.56, and that the average value of normal Hausdorff distance reached 4.95%, and that the average running time of the proposed algorism was 2.16 s. Conclusion The proposed method automatically generates the initial contour, avoiding human intervention. And the proposed segmentation algorithm runs accurately and quickly, achieves good experimental results, which lays foundation for the application in HIFU devices.