Analysis of chest radiographs using texture based image registration methods

OData support
Supervisor:
Orbán Gergely Gyula
Department of Measurement and Information Systems

In the developed world lung cancer is becoming a more and more common cause of death, increasingly burdening the hospitals in terms of analyzing chest radiographs. There are however some decision support systems at our disposal, which can speed up the evaluation process without the need for a radiologist's supervision. For years, the BME Department of Measurement and Information Systems performed research in digital processing of X-ray

images, including the methods introduced in this thesis.

The thesis is about two areas of decision support, symmetry analysis and temporal follow-ups. Symmetry-analysis provides a possibility to emphasize crucial information regarding the diagnosis, while temporal follow-ups provide information about the progression

of the illness. Both problems are to be solved by alignment, i. e. registration of images or image regions. Symmetry analysis involves the registration of the two lung areas on one image, and the subtraction of both, temporal follow-ups require the registration of two images recorded at different times. The thesis introduces two applications solving the two presented problems, after reviewing the theoretical background and the relevant literature of the field.

The symmetry analyzer concludes the registration using a texture based optimization method with different algorithms. One of these is a novel expansion of existing algorithms, which proves to be a valid alternative to other known registration methods. Based on verification results on synthetic and real X-ray images, it is statable that the method successfully performs the desired tasks.

The method realizing a temporal follow-up is based on finding and matching keypoints. After matching the points of the images with SIFT the algorithm removes the outliers from the transformation model using the RANSAC method. Based on the remaining set of pairs a neural network calculates the transformation between the images. Results show an excellent registration performance, clinical applicability would however require further considerations.

Downloads

Please sign in to download the files of this thesis.