Research On Computer-aided Feature Extraction And Classification Of Lung Cancer

The lung cancer is one of the severest malignancies dangerous for human's health, whose incidence and mortality are both highly increasing in all countries, especially in the latest half century. To improve lung cancer patients' viability, the earliest diagnosis and therapy of lung cancer is essential. The CT scanning for lung cancer has obviously higher identification rate than that of ordinary X-ray photo, therefore, it has become the most effective imaging method for lung cancer detection at present. However, the images' quantity generated by CT scanning is in multitude, which would directly burden doctors' workload and therefore increase the pretermission or inaccuracy rate of diagnosis. The goal of this paper is to develop a computer-aided diagnosis system for lung cancer based on CT images, which can identify the suspect pulmonary nodules after the automatic analysis of CT images, and greatly helps doctors improve the diagnostic quality and efficiency.The paper separates the whole image analysis work into four steps: (1) the extraction of pulmonary parenchyma by several image processing operations, such as thresholding, morphologic algorithms, boundary tracking and etc; (2) the segmentation of pulmonary parenchyma images by three algorithms, which are the K-means clustering (KM), the fuzzy C-means clustering (FCM), and the one based on Gibbs random field and fuzzy C-means clustering (GFCM), in order to attain regions of interest (ROIs, including pulmonary nodules, pulmonary blood vessels, pulmonary bronchi and etc.); (3) the selection of five effecitive features from ten orginal alternate ones of ROIs by the theory of feature space optimization design and the standardization of these five features to acquire ten-dimensional input eigenvector by fuzzy theory; (4) the design of a BP-neural network classification tool to realize the final classification of pulmonary nodules.In system tests, we used pulmonary nodule images from 57 patients by four brands of spiral CT scanning, including totally 213 pieces of images, 262 pulmonary nodules. The area below the ROC curve in the test result was 0.9848; when the cut-off point was 0.5, the sensitivity was 96.18%, the specificity was 96.25%, the average false positive of every image was 0.493, and the accuracy rate was 96.24%.On the whole, the system realized the automatic computer detection of pulmonary nodules...