In Partial Fulfillment of the Requirements for the Degree of
Doctor of Philosphy
Will defend his dissertation
The presence of palpable thyroid nodules is a common clinical problem and epidemiologic studies show that up to 7% of the US adult population has single or multiple nodules within the thyroid gland. Fine needle aspiration (FNA) is widely accepted as the most direct and accurate diagnostic procedure in the management of nodular thyroid disease. It is also the most cost-effective tool for the initial screening and triage of thyroid nodule cases. However, suspicious results related to cytopathologic difficulties in differentiating between benign and malignant follicular tumors lead to unnecessary surgical procedure for many patients. As such, there is a concomitant increase in the demand for tools for assessment of FNA cytology (FNAC) leading to improved diagnoses. In this thesis, we develop a systematic approach to image analysis, and develop methods and tools to measure quantitative parameters of normal vs. abnormal cytology. Specifically, our approach to quantifying FNAC specimens is to use multispectral digital microscopy and image analysis algorithms to provide a comprehensive description of specimen morphology and characterization of the cytology and to evaluate the benefit of the use of computer-aided screening in identifying patients with an increased risk of malignancy. We investigate methods of constructing probabilistic models with high dimensional data sets. We show that the spectral information, and the corresponding extension provide valuable discriminative power for differentiating thyroid FNA biopsies. The specific objective of our work is to develop a novel two-stage cell segmentation framework based on a discriminative probabilistic approach that models contextual interactions to differentiate different components of FNAC images by imposing both spatial and spectral regularizations and recover implicit shape information from overlapping objects. The potential benefit of obtaining accurate quantitative parameters through the developed segmentation framework for tumor characterization and classification is also presented.