5 Results

A typical run of the segmentation system is illustrated in Figure  4 . The task is to identify the boundary of a human gall bladder in a series of slices through a 3D ultrasound data set. In the first slice the user specifies B-spline control points near the boundary (Figure  4 (a)) and the system searches normal to the spline for local minima of . Since the boundary is not always well characterised by high gradients, the target points give a fairly poor segmentation (Figure  4 (b)). The user manually corrects the target points to give the desired segmentation (Figure  4 (c)). The system now re-fits the B-spline to the target points and trains local boundary models for each spline segment, so that more appropriate potentials P can be used in the next slice. The B-spline is propagated into the next slice and the 1D searches are repeated (Figure  4 (d)): target points are located at local minima of P along the search lines (Figure  4 (e)). This gives an acceptable segmentation without the need for any user intervention. For comparison, Figure  4 (f) shows the segmentation obtained using the intensity gradient alone.

   
Figure 4: Segmenting a human gall bladder . The figure shows a typical run of the adaptive segmentation system, demonstrating clearly the benefits of statistical boundary models with on-the-fly training. See text for details.

The segmentation system is not computationally expensive and runs fast enough for comfortable interaction. Little user intervention is required after slice 1, since the adaptive models are able to track slow changes in the boundary statistics. Even when user intervention is required, the segmentation is usually considerably better than a segmentation based on gradient information alone.