Multilevel cluster ensembling for histopathological image segmentation
Please cite this item using this persistent URLhttp://hdl.handle.net/11693/15250
In cancer diagnosis and grading, histopathological examination of tissues by pathologists is accepted as the gold standard. However, this procedure has observer variability and leads to subjectivity in diagnosis. In order to overcome such problems, computational methods which use quantitative measures are proposed. These methods extract mathematical features from tissue images assuming they are composed of homogeneous regions and classify images. This assumption is not always true and segmentation of images before classification is necessary. There are methods to segment images but most of them are proposed for generic images and work on the pixel-level. Recently few algorithms incorporated medical background knowledge into segmentation. Their high level feature definitions are very promising. However, in the segmentation step, they use region growing approaches which are not very stable and may lead to local optima. In this thesis, we present an efficient and stable method for the segmentation of histopathological images which produces high quality results. We use existing high level feature definitions to segment tissue images. Our segmentation method significantly improves the segmentation accuracy and stability, compared to existing methods which use the same feature definition. We tackle image segmentation problem as a clustering problem. To improve the quality and the stability of the clustering results, we combine different clustering solutions. This approach is also known as cluster ensembles. We formulate the clustering problem as a graph partitioning problem. In order to obtain diverse and high quality clustering results quickly, we made modifications and improvements on the well-known multilevel graph partitioning scheme. Our method clusters medically meaningful components in tissue images into regions and obtains the final segmentation. Experiments showed that our multilevel cluster ensembling approach performed significantly better than existing segmentation algorithms used for generic and tissue images. Although most of the images used in experiments, contain noise and artifacts, the proposed algorithm produced high quality results.