基于集成分类器和随机森林算法的肺部CT图像处理应用研究
发布时间：2019-01-15 08:54
【摘要】：目的:近年来伴随着雾霾天气的不断加重,间质性肺病的患病率和死亡率持续上升,早期临床症状不明显,只有30%患者肺活检诊断可以发现间质性肺部疾病的症状。间质性肺病虽发展至晚期较易诊断,但早已失去早期诊断的意义。肺部CT(Computed Tomography,CT)对肺组织和间质更能细致显示其形态结构变化,尤其在判定间质性肺病这类以肺部周边病变为主的肺部疾病方面具有独特的诊断价值。在临床辅助诊断中,由于肺实质部分与其它气管、支气管等组织存在粘连,致使在确定病灶区域时存在一定的模糊信息。结合图像预处理、特征提取、分类器等相关算法实现对肺部CT图像的疾病分类以及肺实质分割,为疾病的诊断提供更多更加有效的病灶信息,有利于间质性肺病后续的治疗。方法:本研究针对不同类型肺部疾病在CT图像上相似病理特征,通过PHOG(Pyramid Histogram of Oriented Gradients,PHOG)算法提取肺部CT图像的方向梯度信息,采取“投票”的思想训练基分类器成为集成分类器,构建一个鲁棒的分类模型,进而对冗杂的肺部CT图像实现患病和健康两种不同表征类型的有效区分。其次,课题研究针对肺部CT图像肺实质与非肺实质部位存在不清晰边界的问题,利用肺部CT图像纹理变化大,灰度对比明显的特征,运用灰度共生矩阵算法获取纹理特征同时融合灰度特征构成特征矩阵,选取随机森林作为分类器,提出一种超像素与随机森林复合的分割算法,实现肺实质的准确分割。结果:为测试算法模型的性能选取日内瓦大学一个公开的间质性肺病的数据库ILDs(Interstitial Lung Diseases)。实验结果表明,基于集成分类器算法的肺部CT图像疾病分类模型的准确率达到94.55%,敏感度为86.44%,取得比较理想的分类结果;基于随机森林的肺部CT图像分割算法在健康肺部CT图像的准确率高达99.09%,肺部纤维化、毛玻璃、肺气肿、肺结节患病图像分割准确率均在90%以上。结论:在基于肺部CT图像疾病分类方面,本文提出的基于集成分类器的分类模型,能够高精度的实现肺部CT图像健康和患病两种类型的分类,且鲁棒性较好。肺部CT图像分类算法模型敏感度虽还有待进一步提高,但是对于肺病的临床诊断治疗方案的确定具有一定的现实意义。在肺实质分割方面,提出的基于随机森林分类器的方法,能够准确高效的实现不同种类病理表征肺部CT图像的肺实质部分的分割。在患病严重的肺部CT图像的分割和算法的运算效率还需进一步研究,其对于开展肺部CT图像的检测、量化等进一步的工作仍然具有很好的应用前景。
[Abstract]:Objective: in recent years, with the worsening of haze weather, the morbidity and mortality of interstitial pulmonary disease are rising, and the early clinical symptoms are not obvious. Only 30% of the patients can find the symptoms of interstitial lung disease by lung biopsy diagnosis. Although it is easy to diagnose interstitial pulmonary disease at late stage, it has long lost the significance of early diagnosis. Pulmonary CT (Computed Tomography,CT) can show the morphologic and structural changes of lung tissue and interstitial more carefully, especially in the diagnosis of interstitial pulmonary disease, which is dominated by pulmonary peripheral lesions. In clinical assistant diagnosis, there is some fuzzy information in the determination of lesion area due to the adhesion of lung parenchyma with other trachea and bronchi. Combined with image preprocessing, feature extraction, classifier and other related algorithms to achieve lung CT image disease classification and lung parenchyma segmentation, for the diagnosis of disease to provide more effective focus information, conducive to the follow-up treatment of interstitial pulmonary disease. Methods: based on the similar pathological features of different types of lung diseases on CT images, the directional gradient information of lung CT images was extracted by PHOG (Pyramid Histogram of Oriented Gradients,PHOG algorithm. The idea of "vote" is adopted to train the base classifier as an integrated classifier, and a robust classification model is constructed, which can effectively distinguish ill and healthy lung CT images. Secondly, aiming at the problem of unclear boundary between lung parenchyma and non-pulmonary parenchyma in lung CT images, the feature of large texture change and obvious grayscale contrast of lung CT image is used. Using gray level co-occurrence matrix algorithm to obtain texture features and fuse gray features to form feature matrix, select random forest as classifier, propose a segmentation algorithm which is composed of hyperpixel and random forest, and realize accurate segmentation of lung parenchyma. Results: to test the performance of the algorithm model, select ILDs (Interstitial Lung Diseases)., an open database of interstitial pulmonary diseases at the University of Geneva. The experimental results show that the disease classification model of lung CT image based on the integrated classifier algorithm has the accuracy of 94.555.The sensitivity is 86.44. The accuracy rate of lung CT image segmentation based on random forest is 99.09% in healthy lung CT image. The segmentation accuracy of pulmonary fibrosis, glass, emphysema and pulmonary nodule disease image is over 90%. Conclusion: in the aspect of disease classification based on lung CT image, the classification model based on integrated classifier can achieve the classification of lung CT image health and disease with high accuracy and good robustness. Although the sensitivity of lung CT image classification algorithm model needs to be further improved, it has a certain practical significance for the clinical diagnosis and treatment of lung disease. In the aspect of lung parenchyma segmentation, the proposed method based on stochastic forest classifier can accurately and efficiently segment the lung parenchyma of different kinds of pathological CT images. The segmentation of lung CT images and the computational efficiency of the algorithm need to be further studied, which still has a good application prospect for the detection and quantification of lung CT images.
【学位授予单位】：山东中医药大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R816.4;TP391.41
本文编号：2409055
[Abstract]:Objective: in recent years, with the worsening of haze weather, the morbidity and mortality of interstitial pulmonary disease are rising, and the early clinical symptoms are not obvious. Only 30% of the patients can find the symptoms of interstitial lung disease by lung biopsy diagnosis. Although it is easy to diagnose interstitial pulmonary disease at late stage, it has long lost the significance of early diagnosis. Pulmonary CT (Computed Tomography,CT) can show the morphologic and structural changes of lung tissue and interstitial more carefully, especially in the diagnosis of interstitial pulmonary disease, which is dominated by pulmonary peripheral lesions. In clinical assistant diagnosis, there is some fuzzy information in the determination of lesion area due to the adhesion of lung parenchyma with other trachea and bronchi. Combined with image preprocessing, feature extraction, classifier and other related algorithms to achieve lung CT image disease classification and lung parenchyma segmentation, for the diagnosis of disease to provide more effective focus information, conducive to the follow-up treatment of interstitial pulmonary disease. Methods: based on the similar pathological features of different types of lung diseases on CT images, the directional gradient information of lung CT images was extracted by PHOG (Pyramid Histogram of Oriented Gradients,PHOG algorithm. The idea of "vote" is adopted to train the base classifier as an integrated classifier, and a robust classification model is constructed, which can effectively distinguish ill and healthy lung CT images. Secondly, aiming at the problem of unclear boundary between lung parenchyma and non-pulmonary parenchyma in lung CT images, the feature of large texture change and obvious grayscale contrast of lung CT image is used. Using gray level co-occurrence matrix algorithm to obtain texture features and fuse gray features to form feature matrix, select random forest as classifier, propose a segmentation algorithm which is composed of hyperpixel and random forest, and realize accurate segmentation of lung parenchyma. Results: to test the performance of the algorithm model, select ILDs (Interstitial Lung Diseases)., an open database of interstitial pulmonary diseases at the University of Geneva. The experimental results show that the disease classification model of lung CT image based on the integrated classifier algorithm has the accuracy of 94.555.The sensitivity is 86.44. The accuracy rate of lung CT image segmentation based on random forest is 99.09% in healthy lung CT image. The segmentation accuracy of pulmonary fibrosis, glass, emphysema and pulmonary nodule disease image is over 90%. Conclusion: in the aspect of disease classification based on lung CT image, the classification model based on integrated classifier can achieve the classification of lung CT image health and disease with high accuracy and good robustness. Although the sensitivity of lung CT image classification algorithm model needs to be further improved, it has a certain practical significance for the clinical diagnosis and treatment of lung disease. In the aspect of lung parenchyma segmentation, the proposed method based on stochastic forest classifier can accurately and efficiently segment the lung parenchyma of different kinds of pathological CT images. The segmentation of lung CT images and the computational efficiency of the algorithm need to be further studied, which still has a good application prospect for the detection and quantification of lung CT images.
【学位授予单位】：山东中医药大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R816.4;TP391.41
【参考文献】
相关期刊论文 前2条
1 刘静;郭建;贺遵亮;;基于Gist和PHOG特征的场景分类[J];计算机工程;2015年04期
2 龚敬;王丽嘉;王远军;孙希文;聂生东;;基于灰度积分投影与模糊C均值聚类的肺实质分割[J];中国生物医学工程学报;2015年01期
，本文编号：2409055
本文链接：https://www.wllwen.com/yixuelunwen/yundongyixue/2409055.html
最近更新
教材专著
