Postdoc: Integrating Soft Segmentation With Intensity-Based Matching for 2D/3D Image Data Registration

shahidi， ramin斯坦福大学cise实验计算机科学博士后助理：将软分割与基于强度的匹配相结合用于2D/3D图像数据配准“软”分割是图像处理领域的一项最新创新，它试图在对各种图像元素进行分类的同时保留图像中尽可能多的信息。软分割为图像中的单个像素产生灵活的（例如，模糊或概率）标签，而不是强制决定每个像素。软分割是一种有效的方法，用于噪声图像，如术中透视x线图像，其中信息的保存是至关重要的。近年来，各种有前途的体素属性或基于强度的匹配算法被开发用于三维（3D）医学图像配准，但这些算法在存在明显噪声的情况下是不够的。脊柱图像包含可变形结构（脊柱）中的刚性元素（椎骨）。配准是在单个椎体上进行的，因此，脊柱图像包含结构化和非结构化噪声。提出了将软标签应用于脊柱二维图像的分割，用于二维/三维图像配准的研究。博士后助理将协助1)调整现有的基于基准的临床脊柱导航系统，使用基于图像的预分割图像精细配准，2)开发一个半自动化的透视图像分割，在感兴趣的区域周围生长一个边界框，3)开发模糊和概率“软”标签来分割透视图像。4)将软标签卷积成梯度匹配算法和基于互信息的强度匹配算法。

0104114Shahidi, RaminStanford UniversityCISE Postdoctoral Associates in Experimental Computer Science: Integrating Soft Segmentation with Intensity-Based Matching for 2D/3D Image Data Registration'Soft' segmentation is a recent innovation in image processing which attempts to preserve the maximum amount of information possible in an image while classifying various image elements. Soft segmentation produces flexible (e.g., fuzzy or probabilistic) labels for individual pixels in an image as opposed to forcing a decision about each pixel. Soft segmentation is an effective method for noisy images, such as intra-operative fluoroscopic x-ray images, where preservation of information is critical. In recent years, a variety of promising voxel-property or intensity-based matching algorithms have been developed for three-dimensional (3D) medical image registration, but these algorithms are inadequate in the presence of significant noise. Spine images contain rigid elements (vertebrae) within a deformable structure (spine). Registration is performed on a single vertebra, therefore, spine images contain both structured and unstructured noise. Research is proposed to apply soft labels to the segmentation of two-dimensional (2D) images of the spine for 2D/3D image registration. The postdoctoral associate will assist in 1) adapting an existing fiducial-based clinical spinal navigation system to use image-based fine registration using pre-segmented images, 2) developing a semi-automated segmentation of fluoroscopic images growing a bounding-box around the region of interest, 3) developing both fuzzy and probabilistic 'soft' labels for segmenting the fluoroscopic images, and 4) the convolution of soft labels into gradient and mutual information-based intensity-based matching algorithms.