Computer-aided Diagnosis: Automatic Lung Nodule Detection In Chest Radiograph

Lung cancer is one of the most common and lethal kinds of cancer. The incidence of lung cancer ranks second in all kinds of cancers, which makes it the most common death causation of malignancy in the world. The incidence and death rate of lung cancer raises fast especially in recent fifty years. The five-year survival ratio of lung cancer patients can be highly improved from 14% up to 49% if the lesion is detected in the initial phases. The initial phase of lung cancer presents as lung nodule. So the initial detection and diagnosis of lung nodule value much in the cure of lung cancer.Computer-Aided Diagnosis (CAD) originated from implementation of computer technology into diagnostic imaging field. Based on digital image processing technology, CAD has helped radiologists a lot. Because the fast calculation speed and good repeatability, CAD can greatly reduce the work intensity of diagnostic imaging physicians and guarantee the quality at the same time. CAD implements professional computer algorithm to analyze medical image. Lesions could be detected to help radiologists improve detection rate. So CAD is known as the"second pair of eyes"of radiologists.Since the research group in University of Chicago began the study on detection of pulmonary nodules in the last century 90's, CAD system based on chest X-ray images has been developed over 10 years and has some good research results. But clinical application has not been conducted in China. Therefore, lung nodules CAD system with strong applicability is worth of research to satisfy the actual need of doctors. The objective of this study is to develop an efficient and stable CAD system, which could automatically detect lung nodules in chest X-ray images. Four major research areas are discussed: lung region segmentation, extraction of interested region, feature extraction and selection, and feature classification. (1) Since lung nodules may exist at any location in the lung area, the segmentation of lung area is the basis of diagnosis. In the section, the paper proposes a hybrid algorithm which combines adaptive threshold method and revised contour integration and edge tracking algorithms. (2) In the extraction of interested region process, due to the interested region is suspected as pulmonary nodule, these candidates should be selected to do further analysis and judgments. This paper presents an improved multi-scale template recognition algorithm for pulmonary nodules. (3) In the feature selection part, based on the clinical characteristics of lung nodules, we propose to extract features both in the original region of interest and the enhanced region of interest. The features include traditional location information, geometric features, improved gray grads information features, and radiation features of boundary designed by authors. Feature selection was conducted by genetic algorithm. (4) Last, these features are selected by classifier of cost sensitive support vector machine in order to detect nodules.The present system is proved to be an integrated computer aided diagnosis system by the analytical description. Some results have been achieved with the modified segmentation and multi-scaled model pulmonary nodules detection algorithm. And several improved characteristics have optimized the classification in this thesis. Also, the feature selection module and modified characteristic sensitive SVM classifier improved the system performance. The final practiced SVM training model yielded a sensitivity of 0.7043 under a specificity of 0.9677. The false positive number for each image averaged 2.82, and the AUC valued 0.8272.