|
Applied
Math Seminar |
|
Radiation Oncology treats cancer by delivering relatively small doses of radiation over about six weeks in order to eliminate cancer without destroying or chronically damaging healthy tissue in and around the growth. The technology for planning and delivering these treatments has developed rapidly over the last decade. CT and MRI scans are used as three-dimensional anatomical models to ensure that the treatments deliver a dose distribution that conforms geometrically to the tumor target. The balance between dose to normal surrounding structures and the dose to the tumor is then optimized by modulating the intensity of the treatment fields. The process depends critically upon segmentation of the target volume and normal structures. Segmentation is carried out using manual tools in the current state of the art. This part of the process consumes about one-half of the 2-8 hours per patient required to plan the treatments. An accurate and reliable automated segmentation algorithm would have a large impact on radiation oncology. However, the problem is quite difficult to solve. Anatomical structures can be recognized by trained personnel in CT and MRI scans not only by the shape of the gray scale patterns but also from the context of the patterns. Not surprisingly, gray scale classification and boundary detection are inadequate for anatomical segmentation. It is therefore usually assumed that a segmentation algorithm must use templates developed from training (similar to human training) and contextual pattern recognition in order to segment the wide range of shapes, sizes, and locations in which normal anatomical structures present themselves in the human body. |