A tissue identification framework for medical images and applications to digital radiographic images and MRI

In this thesis, we present a clustering-based medical image segmentation framework for the identification of targeted tissues or structures. The concept behind our framework is the shortcomings of conventional clustering schemes due to the characteristics of medical images. The framework consists of three main stages, namely region construction, region labeling using the voting procedure which is a local region-based clustering method, and label refinement. While the framework has general utility across different applications and modalities, the stages have to be modified according to the a priori information about image modality and targeted tissues or structures.; We present two applications of our framework. First, the problem of segmentation of in digital radiographic images of extremities into bone and non-bone regions for image enhancement purpose is considered. The adaptation of our framework to this problem is thoroughly covered and the segmentation evaluation is performed using manual tracings as the gold standard.; The cartilage tissue extraction in MRI is performed as a second application. This is a three dimensional problem as opposed to the previous application and a crucial process for diagnosis and monitoring of joint diseases. In addition to cartilage extraction, we present the algorithms that we employed for thickness computation of extracted cartilage tissue. The performance of thickness computation algorithm is evaluated using synthetic three dimensional phantoms. We perform the validation of the whole cartilage extraction-thickness computation system using a histology study which has already been performed on animals as the gold standard.