Edge Detection and Correction Algorithm for 3d Ct Images

The normal direction of the bone contour in computed tomography (CT) images provides mportant anatomical information and can guide segmentation algorithms. Since various bones in CT images have
different sizes, and the intensity values of bone pixels are generally nonuniform and noisy, estimation of the normal direction using a single scale is not reliable. We propose a multiscale approach to estimate the normal direction of bone edges. The reliability of the estimation is calculated from theestimated results and, after re-scaling, the reliability is used to further correct the normal direction. The proposed algorithm starts with initial seed search in the input image. Initial seeds can be obtained by thresholding or manual selection. Starting from an initial seed, to find an edge point near the
seed or to correct the location of the seed we calculate the normal direction at the seed and construct a 1-D signal that consists of the intensity of the pixels along the normal direction centered at the seed. We then use an edge filter to find the edge point in the 1-D signal. The deviation of the edge point in the
1-D signal from the center is used to obtain the location of the edge point in the 2-D image. From the edge point, we predict the next edge point along the edge. The predicted point is taken as a seed and the process is repeated until no new edge point can be predicted. We then go to the next initial seed, and the entire process is repeated until all of the initial seeds are used up. Finally,the edge map is postprocessed to produce a better edge map of the bone. Normal-direction estimation is based on the observation that the intensity of pixels near an edge changes more rapidly along the normal direction than along
the tangent direction.
Initial seeds are obtained from some methods such as thresholding or manual selection. Based on the intensity of pixels normal direction is found.By using kalman filter 1D signal is generated.The deviation of the edge point in the 1-D signal from the center is used to obtain the location of the edge point in the 2-D image. From the edge point, we predict the next edge point along the edge. The predicted point is taken as a seed and the process is repeated until no new edge point can be predicted.We then go to the next initial seed, and the entire process is repeated until all of the initial seeds are used up.
Ideally, bone contours are closed. However, due to noise, the contour resulting from the above described process may not be closed and some new information or assumption is needed to enclose the contour. One common approach is Canny’s hysteresis thresholding which uses two separate threshold values.
Those pixels with an edge strength greater than the larger threshold value are deemed to be edge pixels, while those with an edge strength less than the lower threshold value are not on an edge.In this algorithm Instead of canny filter they are using the smoothed 1-D signal along the tangent direction of the edge so that the neighbors’ edge strength could influence the edge strength of the predicted point. The difference in these two methods is that the latter uses more than one neighbor’s information, whereas the former uses the information of the nearest neighbor only. The cost of the latter is that the tangent direction or the normal direction needs to be estimated more accurately than in Canny’s method.
How to find seed points?
Simple way of finding seeds is to utilize the hounsfield values of pixels. For which the value is 0 for water and –1000 or 1023 for air. Muscle and bone are the principal tissues with Hounsfield values greater than that of water in CT slices. To find seeds near the edge of bone we first “windowed” the intensity, namely, we set those pixels with values less than zero to be zero, and set those pixels with values greater than some value.. Next, we calculated edge strength.
Why normal direction?
Normal-direction estimation is based on the observation that the intensity of pixels near an edge changes more rapidly along the normal direction than along the tangent direction.
How to Generate 1D signal?
Once we have estimated and corrected the normal direction,and have obtained the optimal at pixel , we can then construct a 1-D signal composed by the intensity of pixels along the normal direction and centered at pixel . To decrease the perturbations caused by noise and quantization effects we use a smoothed 1-D signal along the tangent direction of the edge. As stated above, we assume that the normal direction is estimated well so that the intensity of changes most rapidly at the edge point. If the normal direction is not estimated well, the response of to an edge filter is weaker, and the detected edge point is not as reliable. Therefore, similar to the estimated normal direction, the reconstructed may not be completely reliable. Because is determined from the normal direction, the reliability
How to delete the bone edges?
In order to aid in segmentation, which is the identification of distinct bones or bone fragments in a given CT slice, we post process the edge map to eliminate edges that join contours of distinct bones, which we call inter-osseal edges. when two bones are either nearby and or have an indistinct region separating them.