Abstract:
Cervical cancer is the second most common cancer in Thai female. One of the standard treatments is brachytherapy. In brachytherapy, radioactive seeds are inserted into the patients vagina. In Thailand, the performing radiology oncologist manually defines the location of cancer, urinary bladder and large intestine. The manual process takes time and tolls. In this thesis, we propose the segmentation method to locate the patients urinary bladder for the brachytherapy. The bladder lumen is first located. In contrast to the one in diagnostic imaging, the lumen in brachytherapy is inhomogeneous and may partially contain an ambiguous boundary. With conventional segmentation methods, different image characteristic leads to inaccurate segmentation. The local intensity mean is used to suppress the effect of the intensity inhomogeneity and the ambiguous boundary. Since the lumen is far brighter than the wall, the proposed method, namely directional local mean difference level set method, has the zero-level contour converged to the region whose local intensity mean inside is higher than the mean outside. After the lumen has been located, the boundary of the bladder is detected. Since the boundary is mostly ambiguous, and the wall thickness is not even, conventional methods, whose result is based on the edge detection and the assumption of the even wall thickness, perform poorly. In this thesis, the procedure of the oncologist is imitated. First, the distinct edge is detected, and if it is not available, the shape of the lumen and the wall thickness are used. Since the wall between bladder and vagina is thinner, two thickness thresholds are utilized. The bladders boundary is then smoothed by first-order Savitzky-Golay filter. The proposed method was compared with coupled directional level set and the method proposed by Ma et.al. The experiment on 100 images demonstrated that this proposed method provided the most similar result to the one by the expert oncologist. It was also converged in every case, while the others failed in some cases. Furthermore, contrary to the other two, it did not require a fine parameter tuning.