BME MS Defense: Sindhuja Tirumalai Govindarajan
Automated Segmentation of Blood Vessels in the Presence of Fibrosis in Volumetric Lung CT Images
Supervised by Professor Walter O'Dell
Radiation sensitivity of lungs raises concern for radiation toxicity in normal tissue surrounding the lesions treated with radiation therapy. A non-invasive assay has been developed for quantification of dose-response relationship of normal lung tissue in patients receiving high-dose stereotactic body radiation therapy to the lung. Follow-up volumetric CT scans obtained from patients show focal regions of fibrosis corresponding to the high-dose region and little observable long-term damage at distant locations in the lung. Apparent change in image intensity is measured using pixel-by-pixel comparison of fibrotic regions in the follow-up scans with pre-treatment scans. Blood vessels with image intensity similar in Hounsfeld units to lung tissue contribute to these pixel-by-pixel calculations, leading to erroneous results. Hence, an accurate segmentation of blood vessels in regions of fibrosis or cancer lesion is required.
The main aim of this work is to automatically segment blood vessels and extract desired fibrosis tissue structure in thoracic CT images. Different approaches to vessel tree segmentation exist including morphological methods, multi-scale approaches, matching filters approaches, centerline extraction procedures, ridge-based methods, and region-based approaches. Due to relatively large background noise in lung CT images and the unpredictable nature of fibrous tissue shape and size, we chose region-based approaches. In this project, I propose a novel hybrid algorithm that performs nearly automatic segmentation on pulmonary blood vessels based on 3D region growing method combined with morphological operators and Euclidean distance measures. 3D seeded region growing algorithm was developed on phantom lung CT scans and tested on a larger phantom data set. The algorithm was further tested on patient lung scans with minor modification of parameters. To measure the performance of the algorithm, the results were compared to a ground truth database obtained by manual segmentation done by an independent reader. The algorithm is simple and efficient, and can be adapted to other imaging modalities or clinical applications. The programming languages and software tools used in this work were ImageJ, ITK-SNAP and MATLAB.