肺部轮廓畸变辅助诊断算法的设计与实现

发布时间：2018-03-11 12:20

本文选题：医学图像处理　切入点：机器学习　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：自动化医学图像的病变定位是计算机医疗辅助的一个关键技术。目前的方法中,需要医生通过人工的方式对医学图形进行诊断,毫无疑问,这需要巨大的人工成本。在现有的医学数据库中,存在着海量的数据。而计算机的辅助可以帮助医生减少工作量,提高诊断的效率,这个课题一直以来都被作为一个重要的课题进行探讨。许多图像相关的分割算法以及匹配模型被提出来,比如基于图像灰度的算法,基于图像梯度的算法,基于模板比配模型等等。但由于医学图像的复杂性,事实上,整个流程仍然需要医生以交互式的方式对图像进行分割和定位,不能够实现大规模的医学图像处理。这篇论文针对肺部医学CT图像序列的轮廓畸变诊断建立了一套全自动的诊断算法。这套算法主要包含轮廓掩膜提取算法和轮廓畸变判断算法。实际产生病变的肺结构中通常会存在许多病症,轮廓畸变只是其中的一部分病症,但无论要对何种病症进行诊断,都需要首先获得肺轮廓的掩膜。这就是说轮廓掩膜提取算法为所有的诊断算法提供了前提性的输入。轮廓畸变判断算法基于提取算法的输入,对轮廓建立子轮廓的特征向量。在训练数据中,通过交互完成对轮廓异常区域的信息标注。通过机器学习获得最后的诊断模型,对判定为异常的轮廓的异常区域进行局部的高亮标注。
[Abstract]:The localization of pathological changes in automatic medical images is a key technology of computer assisted medical treatment. In the present method, doctors are required to diagnose medical graphics manually, and there is no doubt that, There is a huge amount of data in existing medical databases, and computer support can help doctors reduce their workload and improve diagnostic efficiency. Many image segmentation algorithms and matching models have been proposed, such as image grayscale based algorithms, image gradient-based algorithms, image gradient-based algorithms. But because of the complexity of medical images, in fact, the whole process still needs doctors to segment and locate the images in an interactive way. Large scale medical image processing can not be realized. This paper establishes a set of automatic diagnosis algorithms for contour distortion diagnosis of lung medical CT image sequence. This set of algorithms mainly includes contour mask extraction algorithm and wheel. An algorithm for judging the profile distortion. There are usually many diseases in the lung structure that actually causes the disease. Contour distortion is just part of it, but whatever the diagnosis, It is necessary to obtain the mask of the lung contour first. This means that the contour mask extraction algorithm provides a leading input for all diagnostic algorithms. The contour distortion judgment algorithm is based on the input of the extraction algorithm. In the training data, the information of the contour anomaly region is annotated by interaction. The final diagnosis model is obtained by machine learning. Local highlighting of abnormal areas identified as abnormal contours is carried out.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R563;TP391.41

【参考文献】