Research On Computer-aided Detection For Pulmonary Nodules Using CT Images

Lung cancer is one of the common malignant tumors. The key to reduce lung cancer's mortality is early detection, early diagnosis and early treatment. Nodule is one of the important signs of pulmonary lesions, whose properties and prognosis can be indicated by nodules, so its detection and diagnosis has great significance to determine lung lesions. In other words, one of the key technologies of lung cancer's early diagnosis is detecting small nodules in CT images, and judging whether they have a deteriorating tendency. The CT scan is an important means of diagnosing lung cancer. But, the images'quantity generated by CT scanning would directly burden doctors' workload and therefore increase the pretermission and inaccuracy rate of diagnosis. Image processing technologies such as segmentation, extraction,3D reconstruction and displaying, make computer-aided reading CT images possible. These technologies which can assist doctors to analysis lesions and other interested regions qualitatively and quantitatively, can improve efficiency of medical diagnosis and reduce the burden of doctors. This paper aims to apply computer-aided diagnosis of pulmonary nodules to clinical diagnosis. It can give some suggestions where the suspected nodules are before doctors diagnose. Therefore, they don't need to browse all CT images repeatedly, which reduce their workload and improve the diagnostic efficiency.The computer-aided detection of lung nodules proposed in this paper includes four steps:lung parenchyma extraction, interested regions extraction, interested regions 3D reconstruction, and pulmonary nodules recognition. Firstly, for getting the lung binary images, the iteration threshold method is employed to segment the images preliminarily. Based on the characteristic of lung CT images, an effective morphological filtering method is employed on the binary images. Then, the contour tracing and seed filling methods are used to segment the whole lung parenchyma. Three-dimension reconstruction is proposed after the whole Region of Interesting (ROI) is extracted by the Fuzzy C-mean clustering algorithm. Owing to a great deal of false positives lie in ROI, the SVM classifier is designed to distinguish nodules from normal areas with good detected result after the effective features extraction.In this paper,1243 ROI are obtained, including 104 nodules and 1139 not nodules, from 60 CT images. Better detection results are obtained after repeated experiments. The sensitivity and specificity of SVM classifier reach 90.42% and 99.77% respectively.