【作者】 王光洁；

【导师】 张晔；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2007， 硕士

【摘要】 尽管在过去的20年中有众多国内外专家对微钙化点目标的计算机辅助检测算法做了大量的研究,但迄今为止,在微钙化点目标的自动识别算法上面仍然存在目标难以检出以及检测结果假阳性高的问题,通过对这些问题进行探讨,以及对微钙化点目标特征提取问题的研究,本文结合数学形态学和机器学习分类器等知识解决此类困难。本文所研究的对微钙化点目标的检测技术主要分为两部分。首先,针对微钙化点目标检测困难的问题,结合微钙化点多呈小尺寸且近似圆形颗粒状的这一形态信息,提出了一种基于数学形态学的多结构元素多路加权的形态梯度合成的目标增强算法,该算法能够增强图像中具有微钙化点形状特征区域的灰度,非常有利于后面对真实的目标检测。在此基础之上我们使用迭代顺序滤波法粗检测出图像中的微钙化点目标。实验证明这种基于形态增强的目标粗检测算法对真实微钙化点目标的检测率较高,但是此时的检测结果中存在一定数量的假阳性目标。为了降低粗检测结果的假阳性,我们分析验证并且提取了粗检测结果中真实和假阳的微钙化点目标在空域和小波域上面的一些有效的特征,用以训练SVM学习机,最后,用构造出来的SVM分类器对未经过训练的粗检测的疑似目标做去伪判决。实验证明该方法可以去除粗检测结果中大量的假阳性目标,显著降低最终检测结果的假阳性,提高了微钙化点目标的检测性能。实验结果表明,本文所提出的算法对微钙化点目标具有良好的检测性能,即使在致密的乳腺图像中也能够正确检测出真实目标,并且检测结果具有较低的假阳性率。对大量乳腺图像所进行的微钙化点目标检出率达到81.7%,假阳性率为12.2%,比较好的解决了目前微钙化点检测技术中存在的目标检测率低和假阳性率高的问题。更多还原

【Abstract】 Though lots of investigations on the microcalcification detecting technique were done by many experts home and abroad in the past 20 years, there still exists the major problem that the targets detecting results expressed a low sensitivity with a high false-positive rate. We’v done research on this problem and investigated to extract microcalcification’s features, to solve this problem with the knowledge of Mathematical Morphology and Machine Learning Classifier.The microcalcification detecting technique investigated by our dissertation includes two parts as follows: Firstly, depending on the microcalcifications’morphology information that they’re tiny granule and approximate round in shape, we contrive a multi-structure-access morph-grads enhancement algorithm based on Mathematical Morphology, and regions with the similar morph-features to the microcalcifications will get a higher brightness, then we detect the mircocalcification targets with the iterative ordinal filter, and the experiment results prove that the target’s morph-enhancement algorithm can heighten the detecting sensitivity, but the there are still a number of false positives in our results.In order to reduce the amount of false positives in our coarse detecting results, we do analysis and validations to extract those efficient features both in spatial and wavelet domain from moiety of our coarse detecting results which are chosen randomly, as the training samples to SVM classifier, those unchosen detecting results can be the test samples of the trained SVM classifier, and a large quantity of false positives in those test samples will be wiped off from the classifier’s adjudication. The experiment results prove that lots of false positives are wiped off by SVM classifier, so false-positive rate can be remarkable decreased, and the targets detectability is improved.Experiments express that the approach investigated by our dissertation has a good performance on the microcalcification targets detection, we can even detect targets correctly in very dense mammograms, also with low false-positive rate. Our experiment expresses that we have 87.1% as sensitive rate and 12.2% false-positive rate. Thus, we solve those so far existing major problems in the microcalcification targets detecting field.更多还原