胸部X線影像中肺部結(jié)節(jié)的自動(dòng)檢測(cè)技術(shù)——不平衡數(shù)據(jù)集的分類.pdf

1、上海交通大學(xué)碩士學(xué)位論文胸部X線影像中肺部結(jié)節(jié)的自動(dòng)檢測(cè)技術(shù)——不平衡數(shù)據(jù)集的分類姓名：王娟娟申請(qǐng)學(xué)位級(jí)別：碩士專業(yè)：生物醫(yī)學(xué)工程指導(dǎo)教師：任秋實(shí);張繼武20070101II COMPUTED-AIDED DETECTION IN CHEST X-RAY —— CLASSIFICATION ON IMBALANCED DATA ABSTRACT The lung cancer is the No. 1 Cancer Killer wo

2、rldwide, and the cure of tumors is always the aim that people pursued. The key of treatments is early diagnosis. Being regarded as one of the general clinical examination, chest radiography is ubiquitous in clinical prac

3、tice, it make early diagnosis of primary lung cancer significant. But at the same time it encountered many problems, for example radiographer often missed tumors and diagnosis falsely when they faced so many images. The

4、techniques of image processing and pattern recognition make it possible to computer-aided read images, it helps doctors to increase the workflow efficiency and reduce their eye fatigue, thus may improve cancer diagnostic

5、 accuracy. The detection of lung nodule includes pre-processing of images, the extraction of pulmonary parenchyma, the segmentation of region of interest (ROI), features extraction of ROI and classification. The performa

6、nce is bad when we use traditional classification algorithms, because the number of normal samples is far more than the number of cancer samples. Therefore, it is very important to research on classification on imbalance

7、d data. We do research from there aspects: Firstly, we propose a novel over-sampling approach named LLE-Based SMOTE. It can reduce the imbalance of data through incorporating over-sampling method with nonlinear dimension

8、al reduction. The locally linear embedding algorithm (LLE) is first applied to map the high-dimensional data into a low-dimensional space, where the input data is more separable, and thus can be over-sampled by synthetic

9、 minority over-sampling technique (SMOTE).Then the synthetic data points generated by SMOTE are mapped back to the original input space as well through LLE. Secondly, we introduce support vector machines with different c