Research On Segmentation Algorithms Of Regions Of Interest For Pulmonary Lesions

Lung cancer is one of the severest malignancies dangerous for human's health. According to the report of World Health Organization (WHO), death rate caused by lung cancer has already jumped to the highest among all cancers in the world. With the development of computer, computer-aided diagnosis (CAD) has become one of the major research subjects in medical imaging, diagnostic radiology and computer science. Many new technologies of CAD is appearing and making rapid development. The search indicated that CAD has a positive effect in improving diagnostic accuracy rating and reducing ignored diagnosis. Radiologists can evaluate, measure and timely diagnose the medical images by CAD system. However, the segmentation of pulmonary lesions region of interest (ROI) is an important component of CAD system.In this thesis, the basic methods of the image segmentation are analyzed. For the segmentation of pulmonary ROI, three algorithms are studied:(1) Segmentation algorithm of pulmonary ROI by two-dimensional maximum fuzzy entropy threshold method based on weighted multi-scale CI. This algorithm first calculated the weighted multi-scale CI of the image, which will constitute a two-dimensional feature space with gray-scale, and then get threshold to segment ROI by two-dimensional maximum fuzzy entropy criteria; (2) Segmentation algorithm of pulmonary ROI based on mean shift clustering with Hessian matrix, this algorithm carries the dot filtering based on Hessian matrix on pulmonary image to locate ROI, combining the Hessian matrix filtering feature with the intensity range and the spatial position of each pixel to form.a feature vector set, and consequently, groups all feature vector sets of ROI into different clusters with mean shift clustering algorithm; (3) Segmentation algorithm of pulmonary ROI based on improved Markov Random Field (MRF). The segmentation issue is firstly transformed to the Maximum A Posteriori (MAP) by Bayes theorem, and further to the minimum energy function. Then, the algorithm involves the intensity differences, the distances and correlation between two classifications based on the traditional MRF potential function. Finally, the Iterative Conditional Model (ICM) is employed to find out the solution of problem.As an important part of pulmonary nodules detection, the suspected regions of pulmonary lesions can be segmented effectively by ROI segmentation algorithms proposed in the thesis. Pulmonary CT images are used to test each algorithm, and the results show that three ROI segmentation algorithms can effectively segment a nodule region, meanwhile removing the non-nodular regions and reducing the false positive rate.