Multiple Volume Results

We have applied our segmentation method to several multiscan nonuniform datasets to produce high-resolution level set models. The parameters used for these segmentations are listed in Table 8.2. a and i are weights that the user adjusts to balance attraction to edges with curvature-based smoothing during the level set deformation process.

8.4.2.1 Griffin Dataset

The griffin dataset was created with a volumetric laser scan reconstruction algorithm [49]. This algorithm creates a high-resolution volumetric representation of an object by merging multiple depth maps produced via a laser scan. The original griffin dataset has a resolution of 312 x 294 x 144. We have extracted two nonuniform datasets from this high-resolution representation by copying every sixth plane of data in the X direction and every tenth plane in the Y direction. The two derived nonuniform griffin datasets have the following resolution: 52 x 294 x 144 and 312 x 30 x 144. Isosurfaces have been extracted from these datasets, appropriately scaled in the low-resolution direction, and are presented in the first two images in Fig. 8.9 (top). Each low-resolution scan inadequately captures some important geometric feature of the griffin. We have performed a reconstruction from the undersampled nonuniform scans to produce the result in Fig. 8.9 (top). The method produces a high-resolution (312 x 294 x 144) level set model that contains all of the significant features of the original scan.

8.4.2.2 Mouse Embryo Dataset

The first three scans in Fig. 8.9 (bottom) are derived from a high-resolution MR scan of a mouse embryo. They are subsampled versions of a 256 x 128 x 128

Figure 8.9: Nonuniform datasets merged to produce high-resolution level set models, (top) laser scan of a figurine and (bottom) MR scan of a mouse embryo.

volume dataset, and have the following resolutions: 26 x 128 x 128, 256 x 16 x 128, and 256 x 128 x 13. The last image in Fig. 8.9 presents the result produced by our multiscan segmentation method. The information in the first three scans has been successfully used to create a level set model of the embryo with a resolution of 256 x 128 x 130. The finer features of the mouse embryo, namely its hands and feet, have been reconstructed.

8.4.2.3 Zucchini Dataset

The final dataset consists of three individual MRI scans of an actual zucchini. The separate scans have been registered manually and are presented on the left side of Fig. 8.10, each with a different color. The resolutions of the individual scans are 28 x 218 x 188, 244 x 25 x 188, and 244 x 218 x 21. This image highlights the rough alignment of the scans. The right side of Fig. 8.10 presents the result of our level set segmentation. It demonstrates that our approach is able to extract a reasonable model from multiple datasets that are imperfectly aligned.

Figure 8.10: Three low-resolution MR scans of a zucchini that have been individually colored and overlaid to demonstrate their imperfect alignment. The level set model on the right is derived from the three low-resolution scans.

Was this article helpful?

0 0

Post a comment