The goal of my thesis is to describe a fully-automatic and robust algorithm, which is capable of segmenting the shadow of the heart on digital X-ray radiographs. An accurate and precise segmentation algorithm is prerequisite to many complex image processing methods such as the compensation of the segmented shadow, thus automatic detection algorithm could detect pathological changes behind the heart. Segmenting anatomical formations is a difﬁcult task, because they often have large variation, and anatomical boundaries might not coincide with detected edges.
The segmentation algorithm use the (approximate) contour of the lungs for initialization. A simple model of the heart is also generated with supervised learning technique. The ﬁrst approximation of the location of the heart is performed using a special variant of the Hough transformation, called Hough Transform for Natural Shapes (HNS). Later, I use various ﬁltering techniques in order to produce a one-pixel-wide closed contour from the noisy and incomplete edge image.
My algorithm, though doesn’t solve the problem perfectly, show improvement over the other algorithms tested by the MIT department in the past. At the end of my thesis, I show multiple methods to evaluate the performance of the algorithm, using the radiographs from the JPRS database.