Automated edge detection: new methodologies for portal imaging

A primary justification for EPIDs in the clinic has been the advantages and convenience offered by digital imaging over film, notwithstanding the generally poorer image quality of EPIDs. One advantage is the potential for automated image manipulation, such as image registration, to assist the physician in patient positioning and collimation verification. The application of edge detection algorithms, such as Sobel, Laplacian and Canny, in the clinic can be adversely limited in portal imaging due to the presence of dense nonisotropic noise, patient anatomy and imaging artifacts, and the lack of a fixed coordinate system that does not require user input for recognition. A novel more robust edge detection algorithm has been developed: a supervised Laplacian of Gaussian (SLoG). This involves, firstly, generating an “approximate” edge using a thresholding technique. The “approximate” edge serves as a criterion to remove any false edges, which would be detected by solely using the traditional LoG algorithm. Simulated and portal imaging results are presented demonstrating the improved performance of the SLoG algorithm over conventional algorithms in the detection of the collimation edge. In pretreatment portal imaging, a bb tray is inserted in the block tray, providing a superimposed grid pattern to define the x and y machine axes, independent of the collimator edges or detector geometry. A novel application of the Hough transform is presented, resulting in automated recognition of the coordinate system associated with the bb tray. This involves convolution with a kernel unique to the bb tray pattern, thresholding and a Hough transform