Research On GGO Nodule Detection In Lung CT Based On Super-pixels

In clinical medical field, CT images have been widely applied into the disease diagnosis with the growth of CT imaging technology and a large number of CT images have been produced in hospital. CT image is the way to know the condition of patients and doctors need to scan a lot CT images during the diagnosis. Computer aided diagnosis is an important way to release the burden of doctors. In this essay, we investigate to develop an algorithm for automatic detection of ground glass opacity nodules for computer-aided diagnosis of lung diseases, which can reduce the manual effort and improve the accuracy in lung disease diagnosis. The main research work in this essay is introducing the technology of super-pixel into the detection of Ground Glass Opacity(GGO) nodules in two-dimensional lung CT images. We extract super-pixels as Regions of Interest(ROI) of GGO and take super-pixels as basic processing units to further detect GGO. The main contributions in this essay can be described as below:Detection of ROIs based on SLIC algorithm which can generate super-pixels. In this paper, first, changes of window level and window width, lung extraction, simple thresholding and morphological operations are used in order to get a CT image which may contain GGO and the result CT image is then converted to gray image. By applying SLIC(Simple Linear Iterative Clustering) method, super-pixels segmented in the gray image can be regarded as initial GGO candidate regions.Extraction of GGO based on SVM(Supported Vector Machine). After being converted into gray image, the result image still contains other structures and can’t provide as much information as color images do, which leads to inaccurate initial candidate regions. Thus, to detect GGO completely, we need to process ROIs further. We use morphological operations to process each super-pixel region and apply traditional rule filters and a learning-based SVM(Supported Vector Machine) classifier to reduce false positive regions. Finally, we use another SVM classifier to identify GGOs, distinguishing GGO nodules from others. After all the operations, the last GGO nodules can be derived.The database used in our experiment is from LISS database with all real cases. We did the research on the database and carried out three-fold cross validation experiments on GGO data. The sensitivity of the experiment result can reach 88.89%, the specificity can reach 77.27% and the false positive on each scan can be 5.2. The experiment result has proved that our method of detecting GGO nodules by generating super-pixels turns out to be effective.