基于LevelSet的超声乳腺肿瘤图像的轮廓提取

摘要

Breast cancer is the most prevalent cancer in women. Precisely extracting the structural and functional information contained in breast images is current challenge. This research focused on automatic contouring for breast tumors in sonography with LevelSet segmentation. The basic idea was: ① the regions of interest were roughly located. ② Used a sophisticated preprocessing filter [modified curvature diffusion equation (MCDE)] to reduce noise, but preserved the shape and contrast of the breast tumor. ③ Produced the potential edge according to the gradient of image. ④ Set the initial contour. ⑤ Adjusted the speed function of the LevelSet method to seek a better method of automatically extracting the contours of breast tumors from ultrasound (US) images. In the process of exploration, two methods of improved LevelSet were involved: geodesic active contour method and threshold Levelset. Due to weak boundary information of the breast tumor in sonography, the contouring results obtained with computer simulation revealed that the threshold LevelSet identified better contours than the other two methods, and similar to those obtained with manual sketching%　　乳腺癌的发病率在女性癌症中居首位，在计算机辅助下精确地提取出乳腺图像中包含的结构信息和功能信息是当前面临的难题。本研究主要探讨利用水平集分割方法对超声乳腺图像进行肿瘤的自动分割。基本思路：①粗提取肿瘤所在的区域；②采取改进的各项异性扩散方程(MCDE)，在保证乳腺肿瘤边界信息不被削弱的前提下削减超声图像的噪声；③利用图像梯度获取边界的潜在图像；④设置初始轮廓；⑤调整水平集中的速度函数，寻求适用于乳腺超声图像中肿瘤轮廓的提取方法，探索过程涉及两种改进的 LevelSet 方法：测地活动轮廓(geodesic active contour)和阈值水平集(threshold LevelSet)。结果表明，针对乳腺超声肿瘤的弱边界信息，利用阈值水平集分割可以自动获取与手动勾画较为相似的轮廓。