一种针对脑部图像分割强度不均匀性的改进方法

An Improved Method for Intensity Inhomogeneity of Brain Image Segmentation

强度不均匀性是医学图像中常见的问题,对图像的精确分割提出了许多挑战,图像分割是计算机视觉和计算的基础步骤,提出了一种基于模糊C均值(FCM)的能量最小化方法,将全局聚类和局部聚类相结合,用于磁共振(MR)脑图像的偏场估计和分割。该方法将MR图像分解为两个分量作为全局聚类项的优点,充分利用了表征组织物理性质的真实图像和解释强度不均匀性的偏置场及其各自的空间特性。MR图像的分解描述了整个图像中偏移场的变化,其中组织边界的某些深层变化细节可能会丢失。该方法利用了图像局部区域的不同偏移场的局部聚类项,较好地处理了不同组织间强度的深刻变化。由于局部聚类方法对偏移场的分布缺乏全局控制,此文利用了全局聚类和局部聚类的优点,考虑了两者的结合。在该方法中,通过能量最小化过程同时实现了偏移场估计和组织分割。用FCM迭代优化能量最小化问题,通过真实图像和合成图像与相关模型的对比实验,证明了该模型在偏差校正和分割精度方面的优越性。

Intensity inhomogeneity is a common problem in medical images and poses many challenges for accurate segmentation of images.Image segmentation is a foundation step for computer vision and computing.A fuzzy C-means(FCM)-based energy minimization method that combines global clustering and local clustering was proposed for bias field estimation and segmentation in magnetic resonance(MR)brain images.The advantage of this method is that the MR image is decomposed into two components as global clustering items,taking full advantage of the real image that characterizes the physical properties of the tissue,bias fields that explain the intensity inhomogeneity,and their respective spatial properties.The decomposition of MR images describes the changes in bias field of the whole image,where the details of some deep changes of tissue boundaries may be lost.This method uses the local clustering terms of different bias fields in the local area of the image to better deal with the profound changes of strength between different tissues.As there is a lack of global control for distribution of bias field in local clustering method,we took advantage of global clustering and local clustering,and considered the combination of them.In the proposed method,bias field estimation and tissue segmentation were simultaneously achieved by an energy minimization process.The problem of energy minimization was iteratively optimized by FCM.Comparison experiments on some real and synthetic images with relevant models demonstrate the advantages of the proposed model in terms of bias correction and segmentation accuracy.