基于多尺度小波边缘检测改进GVF snake的颈部淋巴结超声图像轮廓提取

EXTRACTING CONTOUR OF CERVICAL LYMPH NODE ULTRASOUND IMAGE WITH GVF SNAKE IMPROVED BASED ON MULTI-SCALE WAVELET EDGE DETECTION

提取超声图像的轮廓对医学诊断有着积极意义。然而,由于超声图像具有目标与背景间对比度低、信噪比低等特点,以往的边缘检测算法在解决图像噪声、精确定位边缘以及获得连续光滑的边缘线之间的矛盾均未得到理想的效果。GVF snake能较好地解决以上矛盾,且具有更大的捕获范围和更强的凹陷域收敛性。但GVF snake初始轮廓线需手工勾勒,不仅比较繁琐,而且目标提取的结果在很大程度上受人工初始化的影响。为此,提出一种多尺度小波变换模极大值与GVF snake算法结合的方法来提取颈部淋巴结超声图像轮廓。该方法首先运用小波变换模极大值多尺度边缘检测算法得到目标图像的边缘图,再在边缘图上分别选取上、下、左、右四个不同方向若干个特征点,即可自动获得较为客观的初始轮廓线,最后利用GVF snake模型提取图像的精确轮廓。实验表明该方法能得到目标图像连续光滑的轮廓线,同时比GVF snake提取的轮廓更加精确,更能反映轮廓的局部细节。此外,由于初始轮廓更加接近给定图像的真实边缘,从而减少了梯度矢量流力场迭代(GVF)次数,提高了轮廓的收敛速度。

It is of positive significance to extract the contour of ultrasound images for medical diagnosis.However,since the ultrasound images have low resolution and SNR,all previous edge detection algorithms cannot achieve ideal effects on overcoming the contradiction between removing image noise,precisely positioning the edge and obtaining continuous and smooth contour lines.GVF snake model can well solve the contradiction above,and has a larger capture range and stronger sag exposed area convergence.However the initial contour of GVF snake must be hand-sketched,it is cumbersome and the results of object extraction are heavily influenced by artificial initialisation. Therefore,we propose a method which combines the multi-scale wavelet transform modulus maxima with GVF snake algorithm to extract the contour of cervical lymph node ultrasound image.First,the method uses wavelet transform modulus maxima multi-scale edge detection algorithm to obtain the edge chart of object image,and then it selects a number of characteristic points from four different directions of upper, lower,left and right,thus can automatically get a rather objective initial contour line.Finally,it uses the GVF snake model to extract the precise contour of the image.Experiments show that the method makes the contour line of the object image smooth and continuous;the contour it extracted is more accurate compared with the GVF snake model,and better reflects local details.In addition,since the initial contour is more approximate to real edge of the given image,thus it reduces the iteration numbers of gradient vector flow （GVF）field and improves the convergence speed of the contour.