Research On The Key Technologies Of Computer Aided Diagnosis For Lung Nodule Diseases In Computed Tomography Images

Lung cancer is one of the most fatal cancers worldwide, and computed Tomography (CT) scan is believed to be the best imaging method for lung cancer detection. Clinical research shows that early diagnosis and treatments play a vital role in improving the cure rate and prognosis of lung cancer patients, and the computer-aided diagnosis of lung cancer, having advantages of reducing the radiologist’s workload and the oversight, as well as improving lung cancer’s diagnostic accuracy in CT images, has therefore become a worldwide research hotspot. The key technologies of lung cancer computer-aided diagnosis in CT images has been studied, including imaging database constructing method, lung parenchyma segmentation methods, and nodule detection algorithms, with emphasis on the detection methods for juxta-pleural nodules, solid nodules, and GGO(Ground-Glass Opacity) nodules.Aiming at solving the shortcomings of the LIDC (Lung Imaging database Consortium) database, a lung imaging database constructing method for computer-aided diagnosis is studied. A lung cancer imaging database framework is proposed and implemented. It includes data storage model,"gold standard" generation based on multi-expert’s annotation, as well as imaging and annotation visualization. This framework offers not only a database for lung cancer computer-aided diagnosis research, but also a good reference for constructing other diseases’imaging database.The lung segmentation and the juxta-pleural nodule detection have been studied, and a novel method based on multi-scale relatively convex hull is proposed for lung contour repair and juxta-pleural nodule detection. The method is composed of the following steps:1) the initial segmentation lung boundary is analyzed;2) incorrect concavity candidates are identified by using multi-scale relatively convex hull;3) each incorrect-concavity candidate is characterized based on a defined set of features, such as the span, depth, and area of incorrect-concavity candidates. Then rule-based filtering is used to distinguish between the incorrect concavities and correct concavities caused by lung structure;4) the incorrect concavities are repaired by replacing each incorrect concavity with suitable segmentation boundary points;5) the difference between the borders of the original lung and the smoothed one is considered as juxta-pleural nodule candidates at some a hull value of the scale parameter;6) each nodule candidate is characterized based on23features; and7) a support vector machine (SVM) classifier is used to reduce false positive. Experimental result shows that this method is able to improve the precision of lung segmentation effectively, and achieved94.5%sensitivity and7.75false positive per scan for juxta-pleural nodule detection.A method for solid nodules detection is studied in this paper. It is composed of several steps:1) a multi-scale dot filter is developed to initially detect the candidate solid nodules;2) then23features are extracted from each candidate, such as the fraction of the candidate nodule surface that is attached to other solid structure, the variance of the pixels that locate in different slice;3) the feature set is optimized using GA, and a SVM classifier is used to identify the true and false positive nodules. Experiments result shows that the detection method achieved98%sensitivity and20false positive per scan, or92%sensitivity and4.7false positive per scan.The detection method of GGO nodules has been studied, and a new detection method of GGO is proposed based on shape index and adaptive nonlinear filter. The method consists of several steps as follows:1) the volumetric shape index map, which is based on local Gaussian and mean curvatures, are calculated for each voxel in the lung region;2) adaptive threshold method is applied to extract GGO candidate nodules from the shape index map;3) blood vessels and extracted candidate nodules are removed from lung CT image, and adaptive nonlinear filter is used to stretch the contrast between GGO nodule and lung parenchyma, then threshold-based approach is used to extract GGO candidate nodules which are missed in the first stage.4) a SVM classifier is applied to reduce the number of false positive. The result shows that the detection method achieved94.4%sensitivity with12.5false positive per scan.Finally, the method for evaluating nodule detection algorithms has been studied. A unified evaluation platform which is consist of lung imaging database, the "gold standard", and the evaluation algorithm is put forward as well as realized for nodule detection algorithms’evaluation and comparison.