Since three dimensional models of anatomical structures are now routinely used by surgeons, there is a clear need to validate the process by which such models are generated. One method of validating the result of relaxing Dual SurfaceNets is by visual inspection. The 3D model can be superimposed onto the original grayscale image, as shown in Fig. 2d. The borders of the model can be confirmed by examining each slice of the image.
In SurfaceNets, each node of the model is guaranteed to lie within one voxel of the original binary segmentation. Dual SurfaceNets can uphold the same constraints, but in practice these constraints need to be relaxed slightly to effectively combine the information in both nets. The distance that a node strays from its initialization point (the center of its cell) can be constrained during the relaxation. Furthermore, upon convergence, the distribution of displacements can be analyzed to determine the goodness of the fit. For both nets of the femur data set, over 97% of the points lie within one voxel of their starting position. Therefore, the final model is not only very smooth, but also faithful to the input segmentation (see [3] for more details).
The validation process is often hindered by the difficulty in obtaining ground truth. While we do not have explicit ground truth, we generated the low resolution femur model and then compared the result with the high resolution femur segmentation. Ideally, each point of the low resolution model should fall near the high resolution surface. Even though the voxel extents of the axial and sagittal scans used in generation of the nets are 4.28mm and 5.78mm respectively, the majority of model points fall within one millimeter of the high resolution model. Furthermore, 98% lie within one sub-sampled voxel of the original data [3]. The model produced by Dual SurfaceNets on the low resolution scans is a good estimate of the high resolution model, while true to the input images.