Research On The Segmentation Algorithms Of Ground Glass Opacity

As the most deadly tumor, the lung cancer deprived 1.1 billion of men’s and 0.49 billion of women’s lives just in 2012. The best way to increase the patients’ five year survival rates and to release the burdens of patients’ families is to diagnose the cancers in their early stages. The ground glass opacity(GGO) is one of the most impartant appearances of the early-stage lung cnacer, but not a specific performance. At present the malignancy of GGO is detected by the analyses of nodule’s morphological features, such as the size, the shape, the margin, the interface, the air bronchograms, the adjacent structures, the solid components et al, and by the analyses of the change of these nodule’s features. All to all, the segmentation of the GGO is an essential step.Recently, severial algorithms have been proposed for the GGOs’ segmentation. But none of them is performed perfectly. What’s more, all the algorithms put forward are performed to get the main body of GGO, neglecting the margin structures such as pine-like process, cusp angles et al, which will cause a over-segment of GGO. The over-segment of GGO will lead some bad results like the wrong measurment of the GGO’s size, the wrong analysis of GGO’s structures and the wrong jugement of GGO’s malignancy. In this paper, a new model, a region based adaptive weight Markov Random field model, is used to get a precise GGO. However, some parameters in this model badly rely on the training set, and the process of the segmentation is complex. In order to solve this problem, two more models was proposed. This paper contains four parts, the extraction of lung parenchyma, the GGO segmentation based on region based adaptive weight MRF, the GGO segmentation based on MRF-2D-OSTU model, the GGO segmentation based on finite similar-possion mixture model.The extraction of lung parenchyma is a main step to eliminate the bad eff ects of the useless parts in CT when we segment the GGO and anslyse GGO’s structures. And it is a good way to reduce computation cost and increase the segmentation accuracy. In this section, two steps are used to get the precise GGO. Fistly, the OSTU model is used to get a thresholding, and the method of connected domian labeling used to fill the holes. Secondly, the rolling-ball algorithm and convex hull based method are used to repair the parenchymas’ margin.The region based adaptive weight Markov Random Field(MRF) model is a MRF model whose transition probability is based on local regions’ membership. This parameter has a big value if a local region seriously belongs to GGO or normal parenchyma, otherwise it owns a small value. Based on this model, a precise segmentation result is got.The MRF based 2D-OSTU is a model that a smoothing energy is introduced to the between-class variance calculating function. This energy is well used to adjust the number and size of the potential areas when the optimum threshold is got.A finite class Possion Mixture Model is a model which has a good use of the relationship between pixels and their neighbourhoods and the facts that the distributions of GGOs’ gray value and normal parenchymas’ gray value are similar to Possion distribution.In this paper, the Experts Grading Method is used to evaluate the segmentation results. And the probabily of error and the relative ultimite measurement accuracy of the size and form is calculated to compare different algorithms. All the results demonstrate that all the three methods get good GGO segmentation results.