|
In this paper, we present a review of the research conducted by our group to design an automatic endoscope navigation and advisory system. The whole system can be viewed as a two-layer system. The first layer is at the signal level, which consists of the processing that will be performed on a series of images to extract all the identifiable features. The information is purely dependent on what can be extracted from the 'raw' images. At the signal level, the first task is performed by detecting a single dominant feature, lumen. Few methods of identifying the lumen are proposed. The first method used contour extraction. Contours are extracted by edge detection, thresholding and linking. This method required images to be divided into overlapping squares (8 by 8 or 4 by 4) where line segments are extracted by using a Hough transform. Perceptual criteria such as proximity, connectivity, similarity in orientation, contrast and edge pixel intensity, are used to group edges both strong and weak. This approach is called perceptual grouping. The second method is based on a region extraction using split and merge approach using spatial domain data. An n-level (for a 2' by 2' image) quadtree based pyramid structure is constructed to find the most homogenous large dark region, which in most cases corresponds to the lumen. The algorithm constructs the quadtree from the bottom (pixel) level upward, recursively and computes the mean and variance of image regions corresponding to quadtree nodes. On reaching the root, the largest uniform seed region, whose mean corresponds to a lumen is selected that is grown by merging with its neighboring regions. In addition to the use of two- dimensional information in the form of regions and contours, three-dimensional shape can provide additional information that will enhance the system capabilities. Shape or depth information from an image is estimated by various methods. A particular technique suitable for endoscopy is the shape from shading, which is developed to obtain the relative depth of the colon surface in the image by assuming a point light source very close to the camera. If we assume the colon has a shape similar to a tube, then a reasonable approximation of the position of the center of the colon (lumen) will be a function of the direction in which the majority of the normal vectors of shape are pointing. The second layer is the control layer and at this level, a decision model must be built for endoscope navigation and advisory system. The system that we built is the models of probabilistic networks that create a basic, artificial intelligence system for navigation in the colon. We have constructed the probabilistic networks from correlated objective data using the maximum weighted spanning tree algorithm. In the construction of a probabilistic network, it is always assumed that the variables starting from the same parent are conditionally independent. However, this may not hold and will give rise to incorrect inferences. In these cases, we proposed the creation of a hidden node to modify the network topology, which in effect models the dependency of correlated variables, to solve the problem. The conditional probability matrices linking the hidden node to its neighbors are determined using a gradient descent method which minimizing the objective cost function. The error gradients can be treated as updating messages and ca be propagated in any direction throughout any singly connected network to adjust the network parameters. With the above two- level approach, we have been able to build an automated endoscope navigation and advisory system successfully.
|
