Registration of chest X-ray images based on the contour of the lung

OData support
Dr. Pataki Béla József
Department of Measurement and Information Systems

Cancer is one of the leading death causes nowadays in the developed countries including Hungary. Lung cancer is the kind of disease which causes the most death. The early detection of lung cancer can make the disease treatable and in many cases completely curable.

The most common imaging method which makes early detection of lung cancer possible is the chest X-ray. In these images the early stage cancer is hard to detect even for experienced doctors. Therefore it is important to develop an algorithm which can help doctors in examining in chest X-rays.

In my thesis I created a method which registers a patient former chest X-ray image onto his recent one. This method provides an opportunity for the doctor where he can swap two images or use a difference image created from these pictures. With this help he can recognize the small differences between the two images easier. This algorithm is a fully automatic method based on pairs of characteristic points.

The structure of the algorithm is the following. After the preprocessing of the two picture of the patient I searched the possible points of the contour of the lung on both images. Using these points I created the contour after local corrections. The results were measured by two types of errors and these error values were comparable with the error values of the hand drawn contours. These lines could be great help for other methods using chest X-rays.

Using the contours and the specific features of the lung I created the pairs of points which are needed for the registration. I used these points and the pairs of the inner points of the lungs which were formerly computed to implement various transformations. I examined several transformations by two kind of error measure. By both of these measures the best working transformation was the second order polynomial transformation by using both the points of the contour and the points of the inner part of the lungs.

The average root mean squared error of this transformation is 6.88 millimeter. This error value and doctor's subjective opinion of the visual appearance of the method provides an opportunity for the algorithm to be part of the running medical system.


Please sign in to download the files of this thesis.